The first comprehensive map of mouse brain activity has been unveiled by a large international collaboration of neuroscientists. 

Researchers from the International Brain Laboratory (IBL), including MIT neuroscientist Ila Fiete, published their open-access findings today in two papers in Nature, revealing insights into how decision-making unfolds across the entire brain in mice at single-cell resolution. This brain-wide activity map challenges the traditional hierarchical view of information processing in the brain and shows that decision-making is distributed across many regions in a highly coordinated way.

“This is the first time anyone has produced a full, brain-wide map of the activity of single neurons during decision-making,” explains co-founder of IBL Alexandre Pouget. “The scale is unprecedented as we recorded from over half-a-million neurons across mice in 12 labs, covering 279 brain areas, which together represent 95 percent of the mouse brain volume. The decision-making activity, and particularly reward, lit up the brain like a Christmas tree,” adds Pouget, who is also a group leader at the University of Geneva in Switzerland.

Modeling decision-making

The brain map was made possible by a major international collaboration of neuroscientists from multiple universities, including MIT. Researchers across 12 labs used state-of-the-art silicon electrodes, called neuropixels probes, for simultaneous neural recordings to measure brain activity while mice were carrying out a decision-making task.

“Participating in the International Brain Laboratory has added new ways for our group to contribute to science,” says Fiete, who is also a professor of brain and cognitive sciences, an associate investigator at the McGovern Institute for Brain Research, and director of the K. Lisa Yang ICoN Center at MIT. “Our lab has helped standardize methods to analyze and generate robust conclusions from data. As computational neuroscientists interested in building models of how the brain works, access to brain-wide recordings is incredible: the traditional approach of recording from one or a few brain areas limited our ability to build and test theories, resulting in fragmented models. Now, we have the delightful but formidable task to make sense of how all parts of the brain coordinate to perform a behavior. Surprisingly, having a full view of the brain leads to simplifications in the models of decision-making,” says Fiete.

The labs collected data from mice performing a decision-making task with sensory, motor, and cognitive components. In the task, a mouse sits in front of a screen and a light appears on the left or right side. If the mouse then responds by moving a small wheel in the correct direction, it receives a reward.

In some trials, the light is so faint that the animal must guess which way to turn the wheel, for which it can use prior knowledge: the light tends to appear more frequently on one side for a number of trials, before the high-frequency side switches. Well-trained mice learn to use this information to help them make correct guesses. These challenging trials therefore allowed the researchers to study how prior expectations influence perception and decision-making.

Brain-wide results

The first paper, “A brain-wide map of neural activity during complex behaviour,” showed that decision-making signals are surprisingly distributed across the brain, not localized to specific regions. This adds brain-wide evidence to a growing number of studies that challenge the traditional hierarchical model of brain function, and emphasizes that there is constant communication across brain areas during decision-making, movement onset, and even reward. This means that neuroscientists will need to take a more holistic, brain-wide approach when studying complex behaviors in the future.

“The unprecedented breadth of our recordings pulls back the curtain on how the entire brain performs the whole arc of sensory processing, cognitive decision-making, and movement generation,” says Fiete. “Structuring a collaboration that collects a large standardized dataset which single labs could not assemble is a revolutionary new direction for systems neuroscience, initiating the field into the hyper-collaborative mode that has contributed to leaps forward in particle physics and human genetics. Beyond our own conclusions, the dataset and associated technologies, which were released much earlier as part of the IBL mission, have already become a massively used resource for the entire neuroscience community.”

The second paper, “Brain-wide representations of prior information,” showed that prior expectations — our beliefs about what is likely to happen based on our recent experience — are encoded throughout the brain. Surprisingly, these expectations are not only found in cognitive areas, but also brain areas that process sensory information and control actions. For example, expectations are even encoded in early sensory areas such as the thalamus, the brain’s first relay for visual input from the eye. This supports the view that the brain acts as a prediction machine, but with expectations encoded across multiple brain structures playing a central role in guiding behavior responses. These findings could have implications for understanding conditions such as schizophrenia and autism, which are thought to be caused by differences in the way expectations are updated in the brain.

“Much remains to be unpacked: If it is possible to find a signal in a brain area, does it mean that this area is generating the signal, or simply reflecting a signal generated somewhere else? How strongly is our perception of the world shaped by our expectations? Now we can generate some quantitative answers and begin the next phase experiments to learn about the origins of the expectation signals by intervening to modulate their activity,” says Fiete.

Looking ahead, the team at IBL plan to expand beyond their initial focus on decision-making to explore a broader range of neuroscience questions. With renewed funding in hand, IBL aims to expand its research scope and continue to support large-scale, standardized experiments.

New model of collaborative neuroscience

Officially launched in 2017, IBL introduced a new model of collaboration in neuroscience that uses a standardized set of tools and data processing pipelines shared across multiple labs, enabling the collection of massive datasets while ensuring data alignment and reproducibility. This approach to democratize and accelerate science draws inspiration from large-scale collaborations in physics and biology, such as CERN and the Human Genome Project.

All data from these studies, along with detailed specifications of the tools and protocols used for data collection, are openly accessible to the global scientific community for further analysis and research. Summaries of these resources can be viewed and downloaded on the IBL website under the sections: Data, Tools, Protocols.

This research was supported by grants from Wellcome, the Simons Foundation, the National Institutes of Health, the National Science Foundation, the Gatsby Charitable Foundation, and by the Max Planck Society and the Humboldt Foundation.

3D printing has come a long way since its invention in 1983 by Chuck Hull, who pioneered stereolithography, a technique that solidifies liquid resin into solid objects using ultraviolet lasers. Over the decades, 3D printers have evolved from experimental curiosities into tools capable of producing everything from custom prosthetics to complex food designs, architectural models, and even functioning human organs. 

But as the technology matures, its environmental footprint has become increasingly difficult to set aside. The vast majority of consumer and industrial 3D printing still relies on petroleum-based plastic filament. And while “greener” alternatives made from biodegradable or recycled materials exist, they come with a serious trade-off: they’re often not as strong. These eco-friendly filaments tend to become brittle under stress, making them ill-suited for structural applications or load-bearing parts — exactly where strength matters most.

This trade-off between sustainability and mechanical performance prompted researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Hasso Plattner Institute to ask: Is it possible to build objects that are mostly eco-friendly, but still strong where it counts?

Their answer is SustainaPrint, a new software and hardware toolkit designed to help users strategically combine strong and weak filaments to get the best of both worlds. Instead of printing an entire object with high-performance plastic, the system analyzes a model through finite element analysis simulations, predicts where the object is most likely to experience stress, and then reinforces just those zones with stronger material. The rest of the part can be printed using greener, weaker filament, reducing plastic use while preserving structural integrity.

“Our hope is that SustainaPrint can be used in industrial and distributed manufacturing settings one day, where local material stocks may vary in quality and composition,” says MIT PhD student and CSAIL researcher Maxine Perroni-Scharf, who is a lead author on a paper presenting the project. “In these contexts, the testing toolkit could help ensure the reliability of available filaments, while the software’s reinforcement strategy could reduce overall material consumption without sacrificing function.” 

For their experiments, the team used Polymaker’s PolyTerra PLA as the eco-friendly filament, and standard or Tough PLA from Ultimaker for reinforcement. They used a 20 percent reinforcement threshold to show that even a small amount of strong plastic goes a long way. Using this ratio, SustainaPrint was able to recover up to 70 percent of the strength of an object printed entirely with high-performance plastic.

They printed dozens of objects, from simple mechanical shapes like rings and beams to more functional household items such as headphone stands, wall hooks, and plant pots. Each object was printed three ways: once using only eco-friendly filament, once using only strong PLA, and once with the hybrid SustainaPrint configuration. The printed parts were then mechanically tested by pulling, bending, or otherwise breaking them to measure how much force each configuration could withstand. 

In many cases, the hybrid prints held up nearly as well as the full-strength versions. For example, in one test involving a dome-like shape, the hybrid version outperformed the version printed entirely in Tough PLA. The team believes this may be due to the reinforced version’s ability to distribute stress more evenly, avoiding the brittle failure sometimes caused by excessive stiffness.

“This indicates that in certain geometries and loading conditions, mixing materials strategically may actually outperform a single homogenous material,” says Perroni-Scharf. “It’s a reminder that real-world mechanical behavior is full of complexity, especially in 3D printing, where interlayer adhesion and tool path decisions can affect performance in unexpected ways.”

A lean, green, eco-friendly printing machine

SustainaPrint starts off by letting a user upload their 3D model into a custom interface. By selecting fixed regions and areas where forces will be applied, the software then uses an approach called “Finite Element Analysis” to simulate how the object will deform under stress. It then creates a map showing pressure distribution inside the structure, highlighting areas under compression or tension, and applies heuristics to segment the object into two categories: those that need reinforcement, and those that don’t.

Recognizing the need for accessible and low-cost testing, the team also developed a DIY testing toolkit to help users assess strength before printing. The kit has a 3D-printable device with modules for measuring both tensile and flexural strength. Users can pair the device with common items like pull-up bars or digital scales to get rough, but reliable performance metrics. The team benchmarked their results against manufacturer data and found that their measurements consistently fell within one standard deviation, even for filaments that had undergone multiple recycling cycles.

Although the current system is designed for dual-extrusion printers, the researchers believe that with some manual filament swapping and calibration, it could be adapted for single-extruder setups, too. In current form, the system simplifies the modeling process by allowing just one force and one fixed boundary per simulation. While this covers a wide range of common use cases, the team sees future work expanding the software to support more complex and dynamic loading conditions. The team also sees potential in using AI to infer the object’s intended use based on its geometry, which could allow for fully automated stress modeling without manual input of forces or boundaries.

3D for free

The researchers plan to release SustainaPrint open-source, making both the software and testing toolkit available for public use and modification. Another initiative they aspire to bring to life in the future: education. “In a classroom, SustainaPrint isn’t just a tool, it’s a way to teach students about material science, structural engineering, and sustainable design, all in one project,” says Perroni-Scharf. “It turns these abstract concepts into something tangible.”

As 3D printing becomes more embedded in how we manufacture and prototype everything from consumer goods to emergency equipment, sustainability concerns will only grow. With tools like SustainaPrint, those concerns no longer need to come at the expense of performance. Instead, they can become part of the design process: built into the very geometry of the things we make.

Co-author Patrick Baudisch, who is a professor at the Hasso Plattner Institute, adds that “the project addresses a key question: What is the point of collecting material for the purpose of recycling, when there is no plan to actually ever use that material? Maxine presents the missing link between the theoretical/abstract idea of 3D printing material recycling and what it actually takes to make this idea relevant.”

Perroni-Scharf and Baudisch wrote the paper with CSAIL research assistant Jennifer Xiao; MIT Department of Electrical Engineering and Computer Science master’s student Cole Paulin ’24; master’s student Ray Wang SM ’25 and PhD student Ticha Sethapakdi SM ’19 (both CSAIL members); Hasso Plattner Institute PhD student Muhammad Abdullah; and Associate Professor Stefanie Mueller, lead of the Human-Computer Interaction Engineering Group at CSAIL.

The researchers’ work was supported by a Designing for Sustainability Grant from the Designing for Sustainability MIT-HPI Research Program. Their work will be presented at the ACM Symposium on User Interface Software and Technology in September.

Many attempts have been made to harness the power of new artificial intelligence and large language models (LLMs) to try to predict the outcomes of new chemical reactions. These have had limited success, in part because until now they have not been grounded in an understanding of fundamental physical principles, such as the laws of conservation of mass. Now, a team of researchers at MIT has come up with a way of incorporating these physical constraints on a reaction prediction model, and thus greatly improving the accuracy and reliability of its outputs.

The new work was reported Aug. 20 in the journal Nature, in a paper by recent postdoc Joonyoung Joung (now an assistant professor at Kookmin University, South Korea); former software engineer Mun Hong Fong (now at Duke University); chemical engineering graduate student Nicholas Casetti; postdoc Jordan Liles; physics undergraduate student Ne Dassanayake; and senior author Connor Coley, who is the Class of 1957 Career Development Professor in the MIT departments of Chemical Engineering and Electrical Engineering and Computer Science.

“The prediction of reaction outcomes is a very important task,” Joung explains. For example, if you want to make a new drug, “you need to know how to make it. So, this requires us to know what product is likely” to result from a given set of chemical inputs to a reaction. But most previous efforts to carry out such predictions look only at a set of inputs and a set of outputs, without looking at the intermediate steps or considering the constraints of ensuring that no mass is gained or lost in the process, which is not possible in actual reactions.

Joung points out that while large language models such as ChatGPT have been very successful in many areas of research, these models do not provide a way to limit their outputs to physically realistic possibilities, such as by requiring them to adhere to conservation of mass. These models use computational “tokens,” which in this case represent individual atoms, but “if you don’t conserve the tokens, the LLM model starts to make new atoms, or deletes atoms in the reaction.” Instead of being grounded in real scientific understanding, “this is kind of like alchemy,” he says. While many attempts at reaction prediction only look at the final products, “we want to track all the chemicals, and how the chemicals are transformed” throughout the reaction process from start to end, he says.

In order to address the problem, the team made use of a method developed back in the 1970s by chemist Ivar Ugi, which uses a bond-electron matrix to represent the electrons in a reaction. They used this system as the basis for their new program, called FlowER (Flow matching for Electron Redistribution), which allows them to explicitly keep track of all the electrons in the reaction to ensure that none are spuriously added or deleted in the process.

The system uses a matrix to represent the electrons in a reaction, and uses nonzero values to represent bonds or lone electron pairs and zeros to represent a lack thereof. “That helps us to conserve both atoms and electrons at the same time,” says Fong. This representation, he says, was one of the key elements to including mass conservation in their prediction system.

The system they developed is still at an early stage, Coley says. “The system as it stands is a demonstration — a proof of concept that this generative approach of flow matching is very well suited to the task of chemical reaction prediction.” While the team is excited about this promising approach, he says, “we’re aware that it does have specific limitations as far as the breadth of different chemistries that it’s seen.” Although the model was trained using data on more than a million chemical reactions, obtained from a U.S. Patent Office database, those data do not include certain metals and some kinds of catalytic reactions, he says.

“We’re incredibly excited about the fact that we can get such reliable predictions of chemical mechanisms” from the existing system, he says. “It conserves mass, it conserves electrons, but we certainly acknowledge that there’s a lot more expansion and robustness to work on in the coming years as well.”

But even in its present form, which is being made freely available through the online platform GitHub, “we think it will make accurate predictions and be helpful as a tool for assessing reactivity and mapping out reaction pathways,” Coley says. “If we’re looking toward the future of really advancing the state of the art of mechanistic understanding and helping to invent new reactions, we’re not quite there. But we hope this will be a steppingstone toward that.”

“It’s all open source,” says Fong. “The models, the data, all of them are up there,” including a previous dataset developed by Joung that exhaustively lists the mechanistic steps of known reactions. “I think we are one of the pioneering groups making this dataset, and making it available open-source, and making this usable for everyone,” he says.

The FlowER model matches or outperforms existing approaches in finding standard mechanistic pathways, the team says, and makes it possible to generalize to previously unseen reaction types. They say the model could potentially be relevant for predicting reactions for medicinal chemistry, materials discovery, combustion, atmospheric chemistry, and electrochemical systems.

In their comparisons with existing reaction prediction systems, Coley says, “using the architecture choices that we’ve made, we get this massive increase in validity and conservation, and we get a matching or a little bit better accuracy in terms of performance.”

He adds that “what’s unique about our approach is that while we are using these textbook understandings of mechanisms to generate this dataset, we’re anchoring the reactants and products of the overall reaction in experimentally validated data from the patent literature.” They are inferring the underlying mechanisms, he says, rather than just making them up. “We’re imputing them from experimental data, and that’s not something that has been done and shared at this kind of scale before.”

The next step, he says, is “we are quite interested in expanding the model’s understanding of metals and catalytic cycles. We’ve just scratched the surface in this first paper,” and most of the reactions included so far don’t include metals or catalysts, “so that’s a direction we’re quite interested in.”

In the long term, he says, “a lot of the excitement is in using this kind of system to help discover new complex reactions and help elucidate new mechanisms. I think that the long-term potential impact is big, but this is of course just a first step.”

The work was supported by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium and the National Science Foundation.

Synthetic data are artificially generated by algorithms to mimic the statistical properties of actual data, without containing any information from real-world sources. While concrete numbers are hard to pin down, some estimates suggest that more than 60 percent of data used for AI applications in 2024 was synthetic, and this figure is expected to grow across industries.

Because synthetic data don’t contain real-world information, they hold the promise of safeguarding privacy while reducing the cost and increasing the speed at which new AI models are developed. But using synthetic data requires careful evaluation, planning, and checks and balances to prevent loss of performance when AI models are deployed.       

To unpack some pros and cons of using synthetic data, MIT News spoke with Kalyan Veeramachaneni, a principal research scientist in the Laboratory for Information and Decision Systems and co-founder of DataCebo whose open-core platform, the Synthetic Data Vault, helps users generate and test synthetic data.

Q: How are synthetic data created?

A: Synthetic data are algorithmically generated but do not come from a real situation. Their value lies in their statistical similarity to real data. If we’re talking about language, for instance, synthetic data look very much as if a human had written those sentences. While researchers have created synthetic data for a long time, what has changed in the past few years is our ability to build generative models out of data and use them to create realistic synthetic data. We can take a little bit of real data and build a generative model from that, which we can use to create as much synthetic data as we want. Plus, the model creates synthetic data in a way that captures all the underlying rules and infinite patterns that exist in the real data.

There are essentially four different data modalities: language, video or images, audio, and tabular data. All four of them have slightly different ways of building the generative models to create synthetic data. An LLM, for instance, is nothing but a generative model from which you are sampling synthetic data when you ask it a question.      

A lot of language and image data are publicly available on the internet. But tabular data, which is the data collected when we interact with physical and social systems, is often locked up behind enterprise firewalls. Much of it is sensitive or private, such as customer transactions stored by a bank. For this type of data, platforms like the Synthetic Data Vault provide software that can be used to build generative models. Those models then create synthetic data that preserve customer privacy and can be shared more widely.      

One powerful thing about this generative modeling approach for synthesizing data is that enterprises can now build a customized, local model for their own data. Generative AI automates what used to be a manual process.

Q: What are some benefits of using synthetic data, and which use-cases and applications are they particularly well-suited for?

A: One fundamental application which has grown tremendously over the past decade is using synthetic data to test software applications. There is data-driven logic behind many software applications, so you need data to test that software and its functionality. In the past, people have resorted to manually generating data, but now we can use generative models to create as much data as we need.

Users can also create specific data for application testing. Say I work for an e-commerce company. I can generate synthetic data that mimics real customers who live in Ohio and made transactions pertaining to one particular product in February or March.

Because synthetic data aren’t drawn from real situations, they are also privacy-preserving. One of the biggest problems in software testing has been getting access to sensitive real data for testing software in non-production environments, due to privacy concerns. Another immediate benefit is in performance testing. You can create a billion transactions from a generative model and test how fast your system can process them.

Another application where synthetic data hold a lot of promise is in training machine-learning models. Sometimes, we want an AI model to help us predict an event that is less frequent. A bank may want to use an AI model to predict fraudulent transactions, but there may be too few real examples to train a model that can identify fraud accurately. Synthetic data provide data augmentation — additional data examples that are similar to the real data. These can significantly improve the accuracy of AI models.

Also, sometimes users don’t have time or the financial resources to collect all the data. For instance, collecting data about customer intent would require conducting many surveys. If you end up with limited data and then try to train a model, it won’t perform well. You can augment by adding synthetic data to train those models better.

Q. What are some of the risks or potential pitfalls of using synthetic data, and are there steps users can take to prevent or mitigate those problems?

A. One of the biggest questions people often have in their mind is, if the data are synthetically created, why should I trust them? Determining whether you can trust the data often comes down to evaluating the overall system where you are using them.

There are a lot of aspects of synthetic data we have been able to evaluate for a long time. For instance, there are existing methods to measure how close synthetic data are to real data, and we can measure their quality and whether they preserve privacy. But there are other important considerations if you are using those synthetic data to train a machine-learning model for a new use case. How would you know the data are going to lead to models that still make valid conclusions?

New efficacy metrics are emerging, and the emphasis is now on efficacy for a particular task. You must really dig into your workflow to ensure the synthetic data you add to the system still allow you to draw valid conclusions. That is something that must be done carefully on an application-by-application basis.

Bias can also be an issue. Since it is created from a small amount of real data, the same bias that exists in the real data can carry over into the synthetic data. Just like with real data, you would need to purposefully make sure the bias is removed through different sampling techniques, which can create balanced datasets. It takes some careful planning, but you can calibrate the data generation to prevent the proliferation of bias.

To help with the evaluation process, our group created the Synthetic Data Metrics Library. We worried that people would use synthetic data in their environment and it would give different conclusions in the real world. We created a metrics and evaluation library to ensure checks and balances. The machine learning community has faced a lot of challenges in ensuring models can generalize to new situations. The use of synthetic data adds a whole new dimension to that problem.

I expect that the old systems of working with data, whether to build software applications, answer analytical questions, or train models, will dramatically change as we get more sophisticated at building these generative models. A lot of things we have never been able to do before will now be possible.

If your hand lotion is a bit runnier than usual coming out of the bottle, it might have something to do with the goop’s “mechanical memory.”

Soft gels and lotions are made by mixing ingredients until they form a stable and uniform substance. But even after a gel has set, it can hold onto “memories,” or residual stress, from the mixing process. Over time, the material can give in to these embedded stresses and slide back into its former, premixed state. Mechanical memory is, in part, why hand lotion separates and gets runny over time. 

Now, an MIT engineer has devised a simple way to measure the degree of residual stress in soft materials after they have been mixed, and found that common products like hair gel and shaving cream have longer mechanical memories, holding onto residual stresses for longer periods of time than manufacturers might have assumed.

In a study appearing today in Physical Review Letters, Crystal Owens, a postdoc in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), presents a new protocol for measuring residual stress in soft, gel-like materials, using a standard benchtop rheometer.

Applying this protocol to everyday soft materials, Owens found that if a gel is made by mixing it in one direction, once it settles into a stable and uniform state, it effectively holds onto the memory of the direction in which it is mixed. Even after several days, the gel will hold some internal stress that, if released, will cause the gel to shift in the direction opposite to how it was initially mixed, reverting back to its earlier state.

“This is one reason different batches of cosmetics or food behave differently even if they underwent ‘identical’ manufacturing,” Owens says. “Understanding and measuring these hidden stresses during processing could help manufacturers design better products that last longer and perform more predictably.”

A soft glass

Hand lotion, hair gel, and shaving cream all fall under the category of “soft glassy materials” — materials that exhibit properties of both solids and liquids.

“Anything you can pour into your hand and it forms a soft mound is going to be considered a soft glass,” Owens explains. “In materials science, it’s considered a soft version of something that has the same amorphous structure as glass.”

In other words, a soft glassy material is a strange amalgam of a solid and a liquid. It can be poured out like a liquid, and it can hold its shape like a solid. Once they are made, these materials exist in a delicate balance between solid and liquid. And Owens wondered: For how long?

“What happens to these materials after very long times? Do they finally relax or do they never relax?” Owens says. “From a physics perspective, that’s a very interesting concept: What is the essential state of these materials?”

Twist and hold

In the manufacturing of soft glassy materials such as hair gel and shampoo, ingredients are first mixed into a uniform product. Quality control engineers then let a sample sit for about a minute — a period of time that they assume is enough to allow any residual stresses from the mixing process dissipate. In that time, the material should settle into a steady, stable state, ready for use.

But Owens suspected that the materials may hold some degree of stress from the production process long after they’ve appeared to settle.

“Residual stress is a low level of stress that’s trapped inside a material after it’s come to a steady state,” Owens says. “This sort of stress has not been measured in these sorts of materials.”

To test her hypothesis, she carried out experiments with two common soft glassy materials: hair gel and shaving cream. She made measurements of each material in a rheometer — an instrument consisting of two rotating plates that can twist and press a material together at precisely controlled pressures and forces that relate directly to the material’s internal stresses and strains.

In her experiments, she placed each material in the rheometer and spun the instrument’s top plate around to mix the material. Then she let the material settle, and then settle some more — much longer than one minute. During this time, she observed the amount of force it took the rheometer to hold the material in place. She reasoned that the greater the rheometer’s force, the more it must be counteracting any stress within the material that would otherwise cause it to shift out of its current state.

Over multiple experiments using this new protocol, Owens found that different types of soft glassy materials held a significant amount of residual stress, long after most researchers would assume the stress had dissipated. What’s more, she found that the degree of stress that a material retained was a reflection of the direction in which it was initially mixed, and when it was mixed.

“The material can effectively ‘remember’ which direction it was mixed, and how long ago,” Owens says. “And it turns out they hold this memory of their past, a lot longer than we used to think.”

In addition to the protocol she has developed to measure residual stress, Owens has developed a model to estimate how a material will change over time, given the degree of residual stress that it holds. Using this model, she says scientists might design materials with “short-term memory,” or very little residual stress, such that they remain stable over longer periods.

One material where she sees room for such improvement is asphalt — a substance that is first mixed, then poured in molten form over a surface where it then cools and settles over time. She suspects that residual stresses from the mixing of asphalt may contribute to cracks forming in pavement over time. Reducing these stresses at the start of the process could lead to longer-lasting, more resilient roads.

“People are inventing new types of asphalt all the time to be more eco-friendly, and all of these will have different levels of residual stress that will need some control,” she says. “There’s plenty of room to explore.”

This research was supported, in part, by MIT’s Postdoctoral Fellowship for Engineering Excellence and an MIT Mathworks Fellowship.

One in every eight people — 970 million globally — live with mental illness, according to the World Health Organization, with depression and anxiety being the most common mental health conditions worldwide. Existing therapies for complex psychiatric disorders like depression, anxiety, and schizophrenia have limitations, and federal funding to address these shortcomings is growing increasingly uncertain.

Patricia and James Poitras ’63 have committed $8 million to the Poitras Center for Psychiatric Disorders Research to launch pioneering research initiatives aimed at uncovering the brain basis of major mental illness and accelerating the development of novel treatments.

“Federal funding rarely supports the kind of bold, early-stage research that has the potential to transform our understanding of psychiatric illness. Pat and I want to help fill that gap — giving researchers the freedom to follow their most promising leads, even when the path forward isn’t guaranteed,” says James Poitras, who is chair of the McGovern Institute for Brain Research board.

Their latest gift builds upon their legacy of philanthropic support for psychiatric disorders research at MIT, which now exceeds $46 million.

“With deep gratitude for Jim and Pat’s visionary support, we are eager to launch a bold set of studies aimed at unraveling the neural and cognitive underpinnings of major mental illnesses,” says Professor Robert Desimone, director of the McGovern Institute, home to the Poitras Center. “Together, these projects represent a powerful step toward transforming how we understand and treat mental illness.”

A legacy of support

Soon after joining the McGovern Institute Leadership Board in 2006, the Poitrases made a $20 million commitment to establish the Poitras Center for Psychiatric Disorders Research at MIT. The center’s goal, to improve human health by addressing the root causes of complex psychiatric disorders, is deeply personal to them both.

“We had decided many years ago that our philanthropic efforts would be directed towards psychiatric research. We could not have imagined then that this perfect synergy between research at MIT’s McGovern Institute and our own philanthropic goals would develop,” recalls Patricia. 

The center supports research at the McGovern Institute and collaborative projects with institutions such as the Broad Institute of MIT and Harvard, McLean Hospital, Mass General Brigham, and other clinical research centers. Since its establishment in 2007, the center has enabled advances in psychiatric research including the development of a machine learning “risk calculator” for bipolar disorder, the use of brain imaging to predict treatment outcomes for anxiety, and studies demonstrating that mindfulness can improve mental health in adolescents.

For the past decade, the Poitrases have also fueled breakthroughs in the lab of McGovern investigator and MIT Professor Feng Zhang, backing the invention of powerful CRISPR systems and other molecular tools that are transforming biology and medicine. Their support has enabled the Zhang team to engineer new delivery vehicles for gene therapy, including vehicles capable of carrying genetic payloads that were once out of reach. The lab has also advanced innovative RNA-guided gene engineering tools such as NovaIscB, published in Nature Biotechnology in May 2025. These revolutionary genome editing and delivery technologies hold promise for the next generation of therapies needed for serious psychiatric illness.

In addition to fueling research in the center, the Poitras family has gifted two endowed professorships — the James and Patricia Poitras Professor of Neuroscience at MIT, currently held by Feng Zhang, and the James W. (1963) and Patricia T. Poitras Professor of Brain and Cognitive Sciences at MIT, held by Guoping Feng — and an annual postdoctoral fellowship at the McGovern Institute.

New initiatives at the Poitras Center

The Poitras family’s latest commitment to the Poitras Center will launch an ambitious set of new projects that bring together neuroscientists, clinicians, and computational experts to probe underpinnings of complex psychiatric disorders including schizophrenia, anxiety, and depression. These efforts reflect the center’s core mission: to speed scientific discovery and therapeutic innovation in the field of psychiatric brain disorders research.

McGovern cognitive neuroscientists Evelina Fedorenko PhD ’07, an associate professor, and Nancy Kanwisher ’80, PhD ’86, the Walter A. Rosenblith Professor of Cognitive Neuroscience — in collaboration with psychiatrist Ann Shinn of McLean Hospital — will explore how altered inner speech and reasoning contribute to the symptoms of schizophrenia. They will collect functional MRI data from individuals diagnosed with schizophrenia and matched controls as they perform reasoning tasks. The goal is to identify the brain activity patterns that underlie impaired reasoning in schizophrenia, a core cognitive disruption in the disorder.

A complementary line of investigation will focus on the role of inner speech — the “voice in our head” that shapes thought and self-awareness. The team will conduct a large-scale online behavioral study of neurotypical individuals to analyze how inner speech characteristics correlate with schizophrenia-spectrum traits. This will be followed by neuroimaging work comparing brain architecture among individuals with strong or weak inner voices and people with schizophrenia, with the aim of discovering neural markers linked to self-talk and disrupted cognition.

A different project led by McGovern neuroscientist and MIT Associate Professor Mark Harnett and 2024–2026 Poitras Center Postdoctoral Fellow Cynthia Rais focuses on how ketamine — an increasingly used antidepressant — alters brain circuits to produce rapid and sustained improvements in mood. Despite its clinical success, ketamine’s mechanisms of action remain poorly understood. The Harnett lab is using sophisticated tools to track how ketamine affects synaptic communication and large-scale brain network dynamics, particularly in models of treatment-resistant depression. By mapping these changes at both the cellular and systems levels, the team hopes to reveal how ketamine lifts mood so quickly — and inform the development of safer, longer-lasting antidepressants.

Guoping Feng is leveraging a new animal model of depression to uncover the brain circuits that drive major depressive disorder. The new animal model provides a powerful system for studying the intricacies of mood regulation. Feng’s team is using state-of-the-art molecular tools to identify the specific genes and cell types involved in this circuit, with the goal of developing targeted treatments that can fine-tune these emotional pathways.

“This is one of the most promising models we have for understanding depression at a mechanistic level,” says Feng, who is also associate director of the McGovern Institute. “It gives us a clear target for future therapies.”

Another novel approach to treating mood disorders comes from the lab of James DiCarlo, the Peter de Florez Professor of Neuroscience at MIT, who is exploring the brain’s visual-emotional interface as a therapeutic tool for anxiety. The amygdala, a key emotional center in the brain, is heavily influenced by visual input. DiCarlo’s lab is using advanced computational models to design visual scenes that may subtly shift emotional processing in the brain — essentially using sight to regulate mood. Unlike traditional therapies, this strategy could offer a noninvasive, drug-free option for individuals suffering from anxiety.

Together, these projects exemplify the kind of interdisciplinary, high-impact research that the Poitras Center was established to support.

“Mental illness affects not just individuals, but entire families who often struggle in silence and uncertainty,” adds Patricia Poitras. “Our hope is that Poitras Center scientists will continue to make important advancements and spark novel treatments for complex mental health disorders and, most of all, give families living with these conditions a renewed sense of hope for the future.”

A new and powerful particle detector just passed a critical test in its goal to decipher the ingredients of the early universe.

The sPHENIX detector is the newest experiment at Brookhaven National Laboratory’s Relativistic Heavy Ion Collider (RHIC) and is designed to precisely measure products of high-speed particle collisions. From the aftermath, scientists hope to reconstruct the properties of quark-gluon plasma (QGP) — a white-hot soup of subatomic particles known as quarks and gluons that is thought to have sprung into existence in the few microseconds following the Big Bang. Just as quickly, the mysterious plasma disappeared, cooling and combining to form the protons and neutrons that make up today’s ordinary matter.

Now, the sPHENIX detector has made a key measurement that proves it has the precision to help piece together the primordial properties of quark-gluon plasma.

In a paper in the Journal of High Energy Physics, scientists including physicists at MIT report that sPHENIX precisely measured the number and energy of particles that streamed out from gold ions that collided at close to the speed of light.

Straight ahead

This test is considered in physics to be a “standard candle,” meaning that the measurement is a well-established constant that can be used to gauge a detector’s precision.

In particular, sPHENIX successfully measured the number of charged particles that are produced when two gold ions collide, and determined how this number changes when the ions collide head-on, versus just glancing by. The detector’s measurements revealed that head-on collisions produced 10 times more charged particles, which were also 10 times more energetic, compared to less straight-on collisions.

“This indicates the detector works as it should,” says Gunther Roland, professor of physics at MIT, who is a member and former spokesperson for the sPHENIX Collaboration. “It’s as if you sent a new telescope up in space after you’ve spent 10 years building it, and it snaps the first picture. It’s not necessarily a picture of something completely new, but it proves that it’s now ready to start doing new science.”

“With this strong foundation, sPHENIX is well-positioned to advance the study of the quark-gluon plasma with greater precision and improved resolution,” adds Hao-Ren Jheng, a graduate student in physics at MIT and a lead co-author of the new paper. “Probing the evolution, structure, and properties of the QGP will help us reconstruct the conditions of the early universe.”

The paper’s co-authors are all members of the sPHENIX Collaboration, which comprises over 300 scientists from multiple institutions around the world, including Roland, Jheng, and physicists at MIT’s Bates Research and Engineering Center.

“Gone in an instant”

Particle colliders such as Brookhaven’s RHIC are designed to accelerate particles at “relativistic” speeds, meaning close to the speed of light. When these particles are flung around in opposite, circulating beams and brought back together, any smash-ups that occur can release an enormous amount of energy. In the right conditions, this energy can very briefly exist in the form of quark-gluon plasma — the same stuff that sprung out of the Big Bang.

Just as in the early universe, quark-gluon plasma doesn’t hang around for very long in particle colliders. If and when QGP is produced, it exists for just 10 to the minus 22, or about a sextillionth, of a second. In this moment, quark-gluon plasma is incredibly hot, up to several trillion degrees Celsius, and behaves as a “perfect fluid,” moving as one entity rather than as a collection of random particles. Almost immediately, this exotic behavior disappears, and the plasma cools and transitions into more ordinary particles such as protons and neutrons, which stream out from the main collision.

“You never see the QGP itself — you just see its ashes, so to speak, in the form of the particles that come from its decay,” Roland says. “With sPHENIX, we want to measure these particles to reconstruct the properties of the QGP, which is essentially gone in an instant.”

“One in a billion”

The sPHENIX detector is the next generation of Brookhaven’s original Pioneering High Energy Nuclear Interaction eXperiment, or PHENIX, which measured collisions of heavy ions generated by RHIC. In 2021, sPHENIX was installed in place of its predecessor, as a faster and more powerful version, designed to detect quark-gluon plasma’s more subtle and ephemeral signatures.

The detector itself is about the size of a two-story house and weighs around 1,000 tons. It sits at the intersection of RHIC’s two main collider beams, where relativistic particles, accelerated from opposite directions, meet and collide, producing particles that fly out into the detector. The sPHENIX detector is able to catch and measure 15,000 particle collisions per second, thanks to its novel, layered components, including the MVTX, or micro-vertex — a subdetector that was designed, built, and installed by scientists at MIT’s Bates Research and Engineering Center.

Together, the detector’s systems enable sPHENIX to act as a giant 3D camera that can track the number, energy, and paths of individual particles during an explosion of particles generated by a single collision.

“SPHENIX takes advantage of developments in detector technology since RHIC switched on 25 years ago, to collect data at the fastest possible rate,” says MIT postdoc Cameron Dean, who was a main contributor to the new study’s analysis. “This allows us to probe incredibly rare processes for the first time.”

In the fall of 2024, scientists ran the detector through the “standard candle” test to gauge its speed and precision. Over three weeks, they gathered data from sPHENIX as the main collider accelerated and smashed together beams of gold ions traveling at the speed of light. Their analysis of the data showed that sPHENIX accurately measured the number of charged particles produced in individual gold ion collisions, as well as the particles’ energies. What’s more, the detector was sensitive to a collision’s “head-on-ness,” and could observe that head-on collisions produced more particles with greater energy, compared to less direct collisions.

“This measurement provides clear evidence that the detector is functioning as intended,” Jheng says.

“The fun for sPHENIX is just beginning,” Dean adds. “We are currently back colliding particles and expect to do so for several more months. With all our data, we can look for the one-in-a-billion rare process that could give us insights on things like the density of QGP, the diffusion of particles through ultra-dense matter, and how much energy it takes to bind different particles together.”

This work was supported, in part, by the U.S. Department of Energy Office of Science, and the National Science Foundation.

Every year, global health experts are faced with a high-stakes decision: Which influenza strains should go into the next seasonal vaccine? The choice must be made months in advance, long before flu season even begins, and it can often feel like a race against the clock. If the selected strains match those that circulate, the vaccine will likely be highly effective. But if the prediction is off, protection can drop significantly, leading to (potentially preventable) illness and strain on health care systems.

This challenge became even more familiar to scientists in the years during the Covid-19 pandemic. Think back to the time (and time and time again), when new variants emerged just as vaccines were being rolled out. Influenza behaves like a similar, rowdy cousin, mutating constantly and unpredictably. That makes it hard to stay ahead, and therefore harder to design vaccines that remain protective.

To reduce this uncertainty, scientists at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Abdul Latif Jameel Clinic for Machine Learning in Health set out to make vaccine selection more accurate and less reliant on guesswork. They created an AI system called VaxSeer, designed to predict dominant flu strains and identify the most protective vaccine candidates, months ahead of time. The tool uses deep learning models trained on decades of viral sequences and lab test results to simulate how the flu virus might evolve and how the vaccines will respond.

Traditional evolution models often analyze the effect of single amino acid mutations independently. “VaxSeer adopts a large protein language model to learn the relationship between dominance and the combinatorial effects of mutations,” explains Wenxian Shi, a PhD student in MIT’s Department of Electrical Engineering and Computer Science, researcher at CSAIL, and lead author of a new paper on the work. “Unlike existing protein language models that assume a static distribution of viral variants, we model dynamic dominance shifts, making it better suited for rapidly evolving viruses like influenza.”

An open-access report on the study was published today in Nature Medicine.

The future of flu

VaxSeer has two core prediction engines: one that estimates how likely each viral strain is to spread (dominance), and another that estimates how effectively a vaccine will neutralize that strain (antigenicity). Together, they produce a predicted coverage score: a forward-looking measure of how well a given vaccine is likely to perform against future viruses.

The scale of the score could be from an infinite negative to 0. The closer the score to 0, the better the antigenic match of vaccine strains to the circulating viruses. (You can imagine it as the negative of some kind of “distance.”)

In a 10-year retrospective study, the researchers evaluated VaxSeer’s recommendations against those made by the World Health Organization (WHO) for two major flu subtypes: A/H3N2 and A/H1N1. For A/H3N2, VaxSeer’s choices outperformed the WHO’s in nine out of 10 seasons, based on retrospective empirical coverage scores (a surrogate metric of the vaccine effectiveness, calculated from the observed dominance from past seasons and experimental HI test results). The team used this to evaluate vaccine selections, as the effectiveness is only available for vaccines actually given to the population. 

For A/H1N1, it outperformed or matched the WHO in six out of 10 seasons. In one notable case, for the 2016 flu season, VaxSeer identified a strain that wasn’t chosen by the WHO until the following year. The model’s predictions also showed strong correlation with real-world vaccine effectiveness estimates, as reported by the CDC, Canada’s Sentinel Practitioner Surveillance Network, and Europe’s I-MOVE program. VaxSeer’s predicted coverage scores aligned closely with public health data on flu-related illnesses and medical visits prevented by vaccination.

So how exactly does VaxSeer make sense of all these data? Intuitively, the model first estimates how rapidly a viral strain spreads over time using a protein language model, and then determines its dominance by accounting for competition among different strains.

Once the model has calculated its insights, they’re plugged into a mathematical framework based on something called ordinary differential equations to simulate viral spread over time. For antigenicity, the system estimates how well a given vaccine strain will perform in a common lab test called the hemagglutination inhibition assay. This measures how effectively antibodies can inhibit the virus from binding to human red blood cells, which is a widely used proxy for antigenic match/antigenicity. 

Outpacing evolution

“By modeling how viruses evolve and how vaccines interact with them, AI tools like VaxSeer could help health officials make better, faster decisions — and stay one step ahead in the race between infection and immunity,” says Shi. 

VaxSeer currently focuses only on the flu virus’s HA (hemagglutinin) protein,the major antigen of influenza. Future versions could incorporate other proteins like NA (neuraminidase), and factors like immune history, manufacturing constraints, or dosage levels. Applying the system to other viruses would also require large, high-quality datasets that track both viral evolution and immune responses — data that aren’t always publicly available. The team, however is currently working on the methods that can predict viral evolution in low-data regimes building on relations between viral families

“Given the speed of viral evolution, current therapeutic development often lags behind. VaxSeer is our attempt to catch up,” says Regina Barzilay, the School of Engineering Distinguished Professor for AI and Health at MIT, AI lead of Jameel Clinic, and CSAIL principal investigator. 

“This paper is impressive, but what excites me perhaps even more is the team’s ongoing work on predicting viral evolution in low-data settings,” says Assistant Professor Jon Stokes of the Department of Biochemistry and Biomedical Sciences at McMaster University in Hamilton, Ontario. “The implications go far beyond influenza. Imagine being able to anticipate how antibiotic-resistant bacteria or drug-resistant cancers might evolve, both of which can adapt rapidly. This kind of predictive modeling opens up a powerful new way of thinking about how diseases change, giving us the opportunity to stay one step ahead and design clinical interventions before escape becomes a major problem.”

Shi and Barzilay wrote the paper with MIT CSAIL postdoc Jeremy Wohlwend ’16, MEng ’17, PhD ’25 and recent CSAIL affiliate Menghua Wu ’19, MEng ’20, PhD ’25. Their work was supported, in part, by the U.S. Defense Threat Reduction Agency and MIT Jameel Clinic.

Today’s electric vehicle boom is tomorrow’s mountain of electronic waste. And while myriad efforts are underway to improve battery recycling, many EV batteries still end up in landfills.

A research team from MIT wants to help change that with a new kind of self-assembling battery material that quickly breaks apart when submerged in a simple organic liquid. In a new paper published in Nature Chemistry, the researchers showed the material can work as the electrolyte in a functioning, solid-state battery cell and then revert back to its original molecular components in minutes.

The approach offers an alternative to shredding the battery into a mixed, hard-to-recycle mass. Instead, because the electrolyte serves as the battery’s connecting layer, when the new material returns to its original molecular form, the entire battery disassembles to accelerate the recycling process.

“So far in the battery industry, we’ve focused on high-performing materials and designs, and only later tried to figure out how to recycle batteries made with complex structures and hard-to-recycle materials,” says the paper’s first author Yukio Cho PhD ’23. “Our approach is to start with easily recyclable materials and figure out how to make them battery-compatible. Designing batteries for recyclability from the beginning is a new approach.”

Joining Cho on the paper are PhD candidate Cole Fincher, Ty Christoff-Tempesta PhD ’22, Kyocera Professor of Ceramics Yet-Ming Chiang, Visiting Associate Professor Julia Ortony, Xiaobing Zuo, and Guillaume Lamour.

Better batteries

There’s a scene in one of the “Harry Potter” films where Professor Dumbledore cleans a dilapidated home with the flick of the wrist and a spell. Cho says that image stuck with him as a kid. (What better way to clean your room?) When he saw a talk by Ortony on engineering molecules so that they could assemble into complex structures and then revert back to their original form, he wondered if it could be used to make battery recycling work like magic.

That would be a paradigm shift for the battery industry. Today, batteries require harsh chemicals, high heat, and complex processing to recycle. There are three main parts of a battery: the positively charged cathode, the negatively charged electrode, and the electrolyte that shuttles lithium ions between them. The electrolytes in most lithium-ion batteries are highly flammable and degrade over time into toxic byproducts that require specialized handling.

To simplify the recycling process, the researchers decided to make a more sustainable electrolyte. For that, they turned to a class of molecules that self-assemble in water, named aramid amphiphiles (AAs), whose chemical structures and stability mimic that of Kevlar. The researchers further designed the AAs to contain polyethylene glycol (PEG), which can conduct lithium ions, on one end of each molecule. When the molecules are exposed to water, they spontaneously form nanoribbons with ion-conducting PEG surfaces and bases that imitate the robustness of Kevlar through tight hydrogen bonding. The result is a mechanically stable nanoribbon structure that conducts ions across its surface.

“The material is composed of two parts,” Cho explains. “The first part is this flexible chain that gives us a nest, or host, for lithium ions to jump around. The second part is this strong organic material component that is used in the Kevlar, which is a bulletproof material. Those make the whole structure stable.”

When added to water, the nanoribbons self-assemble to form millions of nanoribbons that can be hot-pressed into a solid-state material.

“Within five minutes of being added to water, the solution becomes gel-like, indicating there are so many nanofibers formed in the liquid that they start to entangle each other,” Cho says. “What’s exciting is we can make this material at scale because of the self-assembly behavior.”

The team tested the material’s strength and toughness, finding it could endure the stresses associated with making and running the battery. They also constructed a solid-state battery cell that used lithium iron phosphate for the cathode and lithium titanium oxide as the anode, both common materials in today’s batteries. The nanoribbons moved lithium ions successfully between the electrodes, but a side-effect known as polarization limited the movement of lithium ions into the battery’s electrodes during fast bouts of charging and discharging, hampering its performance compared to today’s gold-standard commercial batteries.

“The lithium ions moved along the nanofiber all right, but getting the lithium ion from the nanofibers to the metal oxide seems to be the most sluggish point of the process,” Cho says.

When they immersed the battery cell into organic solvents, the material immediately dissolved, with each part of the battery falling away for easier recycling. Cho compared the materials’ reaction to cotton candy being submerged in water.

“The electrolyte holds the two battery electrodes together and provides the lithium-ion pathways,” Cho says. “So, when you want to recycle the battery, the entire electrolyte layer can fall off naturally and you can recycle the electrodes separately.”

Validating a new approach

Cho says the material is a proof of concept that demonstrates the recycle-first approach.

“We don’t want to say we solved all the problems with this material,” Cho says. “Our battery performance was not fantastic because we used only this material as the entire electrolyte for the paper, but what we’re picturing is using this material as one layer in the battery electrolyte. It doesn’t have to be the entire electrolyte to kick off the recycling process.”

Cho also sees a lot of room for optimizing the material’s performance with further experiments.

Now, the researchers are exploring ways to integrate these kinds of materials into existing battery designs as well as implementing the ideas into new battery chemistries.

“It’s very challenging to convince existing vendors to do something very differently,” Cho says. “But with new battery materials that may come out in five or 10 years, it could be easier to integrate this into new designs in the beginning.”

Cho also believes the approach could help reshore lithium supplies by reusing materials from batteries that are already in the U.S.

“People are starting to realize how important this is,” Cho says. “If we can start to recycle lithium-ion batteries from battery waste at scale, it’ll have the same effect as opening lithium mines in the U.S. Also, each battery requires a certain amount of lithium, so extrapolating out the growth of electric vehicles, we need to reuse this material to avoid massive lithium price spikes.”

The work was supported, in part, by the National Science Foundation and the U.S. Department of Energy. This work was performed, in part, using the MIT.nano Characterization facilities.

In World War II, Britain was fighting for its survival against German aerial bombardment. Yet Britain was importing dyes from Germany at the same time. This sounds curious, to put it mildly. How can two countries at war with each other also be trading goods?

Examples of this abound, actually. Britain also traded with its enemies for almost all of World War I. India and Pakistan conducted trade with each other during the First Kashmir War, from 1947 to 1949, and during the India-Pakistan War of 1965. Croatia and then-Yugoslavia traded with each other while fighting in 1992.

“States do in fact trade with their enemies during wars,” says MIT political scientist Mariya Grinberg. “There is a lot of variation in which products get traded, and in which wars, and there are differences in how long trade lasts into a war. But it does happen.”

Indeed, as Grinberg has found, state leaders tend to calculate whether trade can give them an advantage by boosting their own economies while not supplying their enemies with anything too useful in the near term.

“At its heart, wartime trade is all about the tradeoff between military benefits and economic costs,” Grinberg says. “Severing trade denies the enemy access to your products that could increase their military capabilities, but it also incurs a cost to you because you’re losing trade and neutral states could take over your long-term market share.” Therefore, many countries try trading with their wartime foes.

Grinberg explores this topic in a groundbreaking new book, the first one on the subject, “Trade in War: Economic Cooperation Across Enemy Lines,” published this month by Cornell University Press. It is also the first book by Grinberg, an assistant professor of political science at MIT.

Calculating time and utility

“Trade in War” has its roots in research Grinberg started as a doctoral student at the University of Chicago, where she noticed that wartime trade was a phenomenon not yet incorporated into theories of state behavior.

Grinberg wanted to learn about it comprehensively, so, as she quips, “I did what academics usually do: I went to the work of historians and said, ‘Historians, what have you got for me?’”

Modern wartime trading began during the Crimean War, which pitted Russia against France, Britain, the Ottoman Empire, and other allies. Before the war’s start in 1854, France had paid for many Russian goods that could not be shipped because ice in the Baltic Sea was late to thaw. To rescue its produce, France then persuaded Britain and Russia to adopt “neutral rights,” codified in the 1856 Declaration of Paris, which formalized the idea that goods in wartime could be shipped via neutral parties (sometimes acting as intermediaries for warring countries).

“This mental image that everyone has, that we don’t trade with our enemies during war, is actually an artifact of the world without any neutral rights,” Grinberg says. “Once we develop neutral rights, all bets are off, and now we have wartime trade.”

Overall, Grinberg’s systematic analysis of wartime trade shows that it needs to be understood on the level of particular goods. During wartime, states calculate how much it would hurt their own economies to stop trade of certain items; how useful specific products would be to enemies during war, and in what time frame; and how long a war is going to last.

“There are two conditions under which we can see wartime trade,” Grinberg says. “Trade is permitted when it does not help the enemy win the war, and it’s permitted when ending it would damage the state’s long-term economic security, beyond the current war.”

Therefore a state might export diamonds, knowing an adversary would need to resell such products over time to finance any military activities. Conversely, states will not trade products that can quickly convert into military use.

“The tradeoff is not the same for all products,” Grinberg says. “All products can be converted into something of military utility, but they vary in how long that takes. If I’m expecting to fight a short war, things that take a long time for my opponent to convert into military capabilities won’t help them win the current war, so they’re safer to trade.” Moreover, she adds, “States tend to prioritize maintaining their long-term economic stability, as long as the stakes don’t hit too close to home.”

This calculus helps explain some seemingly inexplicable wartime trade decisions. In 1917, three years into World War I, Germany started trading dyes to Britain. As it happens, dyes have military uses, for example as coatings for equipment. And World War I, infamously, was lasting far beyond initial expectations. But as of 1917, German planners thought the introduction of unrestricted submarine warfare would bring the war to a halt in their favor within a few months, so they approved the dye exports. That calculation was wrong, but it fits the framework Grinberg has developed.

States: Usually wrong about the length of wars

“Trade in War” has received praise from other scholars in the field. Michael Mastanduno of Dartmouth College has said the book “is a masterful contribution to our understanding of how states manage trade-offs across economics and security in foreign policy.”

For her part, Grinberg notes that her work holds multiple implications for international relations — one being that trade relationships do not prevent hostilities from unfolding, as some have theorized.

“We can’t expect even strong trade relations to deter a conflict,” Grinberg says. “On the other hand, when we learn our assumptions about the world are not necessarily correct, we can try to find different levers to deter war.”

Grinberg has also observed that states are not good, by any measure, at projecting how long they will be at war.

“States very infrequently get forecasts about the length of war right,” Grinberg says. That fact has formed the basis of a second, ongoing Grinberg book project.

“Now I’m studying why states go to war unprepared, why they think their wars are going to end quickly,” Grinberg says. “If people just read history, they will learn almost all of human history works against this assumption.”

At the same time, Grinberg thinks there is much more that scholars could learn specifically about trade and economic relations among warring countries — and hopes her book will spur additional work on the subject.

“I’m almost certain that I’ve only just begun to scratch the surface with this book,” she says. 

Our cells produce a variety of proteins, each with a specific role that, in many cases, means that they need to be in a particular part of the cell where that role is needed. One of the ways that cells ensure certain proteins end up in the right location at the right time is through localized translation, a process that ensures that proteins are made — or translated — close to where they will be needed. MIT professor of biology and Whitehead Institute for Biomedical Research member Jonathan Weissman and colleagues have studied localized translation in order to understand how it affects cell functions and allows cells to quickly respond to changing conditions.

Now, Weissman, who is also a Howard Hughes Medical Institute Investigator, and postdoc in his lab Jingchuan Luo have expanded our knowledge of localized translation at mitochondria, structures that generate energy for the cell. In an open-access paper published today in Cell, they share a new tool, LOCL-TL, for studying localized translation in close detail, and describe the discoveries it enabled about two classes of proteins that are locally translated at mitochondria.

The importance of localized translation at mitochondria relates to their unusual origin. Mitochondria were once bacteria that lived within our ancestors’ cells. Over time, the bacteria lost their autonomy and became part of the larger cells, which included migrating most of their genes into the larger cell’s genome in the nucleus. Cells evolved processes to ensure that proteins needed by mitochondria that are encoded in genes in the larger cell’s genome get transported to the mitochondria. Mitochondria retain a few genes in their own genome, so production of proteins from the mitochondrial genome and that of the larger cell’s genome must be coordinated to avoid mismatched production of mitochondrial parts. Localized translation may help cells to manage the interplay between mitochondrial and nuclear protein production — among other purposes.

How to detect local protein production

For a protein to be made, genetic code stored in DNA is read into RNA, and then the RNA is read or translated by a ribosome, a cellular machine that builds a protein according to the RNA code. Weissman’s lab previously developed a method to study localized translation by tagging ribosomes near a structure of interest, and then capturing the tagged ribosomes in action and observing the proteins they are making. This approach, called proximity-specific ribosome profiling, allows researchers to see what proteins are being made where in the cell. The challenge that Luo faced was how to tweak this method to capture only ribosomes at work near mitochondria.

Ribosomes work quickly, so a ribosome that gets tagged while making a protein at the mitochondria can move on to making other proteins elsewhere in the cell in a matter of minutes. The only way researchers can guarantee that the ribosomes they capture are still working on proteins made near the mitochondria is if the experiment happens very quickly.

Weissman and colleagues had previously solved this time sensitivity problem in yeast cells with a ribosome-tagging tool called BirA that is activated by the presence of the molecule biotin. BirA is fused to the cellular structure of interest, and tags ribosomes it can touch — but only once activated. Researchers keep the cell depleted of biotin until they are ready to capture the ribosomes, to limit the time when tagging occurs. However, this approach does not work with mitochondria in mammalian cells because they need biotin to function normally, so it cannot be depleted.

Luo and Weissman adapted the existing tool to respond to blue light instead of biotin. The new tool, LOV-BirA, is fused to the mitochondrion’s outer membrane. Cells are kept in the dark until the researchers are ready. Then they expose the cells to blue light, activating LOV-BirA to tag ribosomes. They give it a few minutes and then quickly extract the ribosomes. This approach proved very accurate at capturing only ribosomes working at mitochondria.

The researchers then used a method originally developed by the Weissman lab to extract the sections of RNA inside of the ribosomes. This allows them to see exactly how far along in the process of making a protein the ribosome is when captured, which can reveal whether the entire protein is made at the mitochondria, or whether it is partly produced elsewhere and only gets completed at the mitochondria.

“One advantage of our tool is the granularity it provides,” Luo says. “Being able to see what section of the protein is locally translated helps us understand more about how localized translation is regulated, which can then allow us to understand its dysregulation in disease and to control localized translation in future studies.”

Two protein groups are made at mitochondria

Using these approaches, the researchers found that about 20 percent of the genes needed in mitochondria that are located in the main cellular genome are locally translated at mitochondria. These proteins can be divided into two distinct groups with different evolutionary histories and mechanisms for localized translation.

One group consists of relatively long proteins, each containing more than 400 amino acids or protein building blocks. These proteins tend to be of bacterial origin — present in the ancestor of mitochondria — and they are locally translated in both mammalian and yeast cells, suggesting that their localized translation has been maintained through a long evolutionary history.

Like many mitochondrial proteins encoded in the nucleus, these proteins contain a mitochondrial targeting sequence (MTS), a ZIP code that tells the cell where to bring them. The researchers discovered that most proteins containing an MTS also contain a nearby inhibitory sequence that prevents transportation until they are done being made. This group of locally translated proteins lacks the inhibitory sequence, so they are brought to the mitochondria during their production.

Production of these longer proteins begins anywhere in the cell, and then after approximately the first 250 amino acids are made, they get transported to the mitochondria. While the rest of the protein gets made, it is simultaneously fed into a channel that brings it inside the mitochondrion. This ties up the channel for a long time, limiting import of other proteins, so cells can only afford to do this simultaneous production and import for select proteins. The researchers hypothesize that these bacterial-origin proteins are given priority as an ancient mechanism to ensure that they are accurately produced and placed within mitochondria.

The second locally translated group consists of short proteins, each less than 200 amino acids long. These proteins are more recently evolved, and correspondingly, the researchers found that the mechanism for their localized translation is not shared by yeast. Their mitochondrial recruitment happens at the RNA level. Two sequences within regulatory sections of each RNA molecule that do not encode the final protein instead code for the cell’s machinery to recruit the RNAs to the mitochondria.

The researchers searched for molecules that might be involved in this recruitment, and identified the RNA binding protein AKAP1, which exists at mitochondria. When they eliminated AKAP1, the short proteins were translated indiscriminately around the cell. This provided an opportunity to learn more about the effects of localized translation, by seeing what happens in its absence. When the short proteins were not locally translated, this led to the loss of various mitochondrial proteins, including those involved in oxidative phosphorylation, our cells’ main energy generation pathway.

In future research, Weissman and Luo will delve deeper into how localized translation affects mitochondrial function and dysfunction in disease. The researchers also intend to use LOCL-TL to study localized translation in other cellular processes, including in relation to embryonic development, neural plasticity, and disease.

“This approach should be broadly applicable to different cellular structures and cell types, providing many opportunities to understand how localized translation contributes to biological processes,” Weissman says. “We’re particularly interested in what we can learn about the roles it may play in diseases including neurodegeneration, cardiovascular diseases, and cancers.”

MIT researchers have developed a technique that enables real-time, 3D monitoring of corrosion, cracking, and other material failure processes inside a nuclear reactor environment.

This could allow engineers and scientists to design safer nuclear reactors that also deliver higher performance for applications like electricity generation and naval vessel propulsion.

During their experiments, the researchers utilized extremely powerful X-rays to mimic the behavior of neutrons interacting with a material inside a nuclear reactor.

They found that adding a buffer layer of silicon dioxide between the material and its substrate, and keeping the material under the X-ray beam for a longer period of time, improves the stability of the sample. This allows for real-time monitoring of material failure processes.

By reconstructing 3D image data on the structure of a material as it fails, researchers could design more resilient materials that can better withstand the stress caused by irradiation inside a nuclear reactor.

“If we can improve materials for a nuclear reactor, it means we can extend the life of that reactor. It also means the materials will take longer to fail, so we can get more use out of a nuclear reactor than we do now. The technique we’ve demonstrated here allows to push the boundary in understanding how materials fail in real-time,” says Ericmoore Jossou, who has shared appointments in the Department of Nuclear Science and Engineering (NSE), where he is the John Clark Hardwick Professor, and the Department of Electrical Engineering and Computer Science (EECS), and the MIT Schwarzman College of Computing.

Jossou, senior author of a study on this technique, is joined on the paper by lead author David Simonne, an NSE postdoc; Riley Hultquist, a graduate student in NSE; Jiangtao Zhao, of the European Synchrotron; and Andrea Resta, of Synchrotron SOLEIL. The research was published Tuesday by the journal Scripta Materiala.

“Only with this technique can we measure strain with a nanoscale resolution during corrosion processes. Our goal is to bring such novel ideas to the nuclear science community while using synchrotrons both as an X-ray probe and radiation source,” adds Simonne.

Real-time imaging

Studying real-time failure of materials used in advanced nuclear reactors has long been a goal of Jossou’s research group.

Usually, researchers can only learn about such material failures after the fact, by removing the material from its environment and imaging it with a high-resolution instrument.

“We are interested in watching the process as it happens. If we can do that, we can follow the material from beginning to end and see when and how it fails. That helps us understand a material much better,” he says.

They simulate the process by firing an extremely focused X-ray beam at a sample to mimic the environment inside a nuclear reactor. The researchers must use a special type of high-intensity X-ray, which is only found in a handful of experimental facilities worldwide.

For these experiments they studied nickel, a material incorporated into alloys that are commonly used in advanced nuclear reactors. But before they could start the X-ray equipment, they had to prepare a sample.

To do this, the researchers used a process called solid state dewetting, which involves putting a thin film of the material onto a substrate and heating it to an extremely high temperature in a furnace until it transforms into single crystals.

“We thought making the samples was going to be a walk in the park, but it wasn’t,” Jossou says.

As the nickel heated up, it interacted with the silicon substrate and formed a new chemical compound, essentially derailing the entire experiment. After much trial-and-error, the researchers found that adding a thin layer of silicon dioxide between the nickel and substrate prevented this reaction.

But when crystals formed on top of the buffer layer, they were highly strained. This means the individual atoms had moved slightly to new positions, causing distortions in the crystal structure.

Phase retrieval algorithms can typically recover the 3D size and shape of a crystal in real-time, but if there is too much strain in the material, the algorithms will fail.

However, the team was surprised to find that keeping the X-ray beam trained on the sample for a longer period of time caused the strain to slowly relax, due to the silicon buffer layer. After a few extra minutes of X-rays, the sample was stable enough that they could utilize phase retrieval algorithms to accurately recover the 3D shape and size of the crystal.

“No one had been able to do that before. Now that we can make this crystal, we can image electrochemical processes like corrosion in real time, watching the crystal fail in 3D under conditions that are very similar to inside a nuclear reactor. This has far-reaching impacts,” he says.

They experimented with a different substrate, such as niobium doped strontium titanate, and found that only a silicon dioxide buffered silicon wafer created this unique effect.

An unexpected result

As they fine-tuned the experiment, the researchers discovered something else.

They could also use the X-ray beam to precisely control the amount of strain in the material, which could have implications for the development of microelectronics.

In the microelectronics community, engineers often introduce strain to deform a material’s crystal structure in a way that boosts its electrical or optical properties.

“With our technique, engineers can use X-rays to tune the strain in microelectronics while they are manufacturing them. While this was not our goal with these experiments, it is like getting two results for the price of one,” he adds.

In the future, the researchers want to apply this technique to more complex materials like steel and other metal alloys used in nuclear reactors and aerospace applications. They also want to see how changing the thickness of the silicon dioxide buffer layer impacts their ability to control the strain in a crystal sample.

“This discovery is significant for two reasons. First, it provides fundamental insight into how nanoscale materials respond to radiation — a question of growing importance for energy technologies, microelectronics, and quantum materials. Second, it highlights the critical role of the substrate in strain relaxation, showing that the supporting surface can determine whether particles retain or release strain when exposed to focused X-ray beams,” says Edwin Fohtung, an associate professor at the Rensselaer Polytechnic Institute, who was not involved with this work.

This work was funded, in part, by the MIT Faculty Startup Fund and the U.S. Department of Energy. The sample preparation was carried out, in part, at the MIT.nano facilities.

Crime and public safety are among the most pressing concerns across communities in the United States. Violence fractures lives and carries staggering costs; the economic burden of gun violence alone tops $100 billion each year. More than 5 million people live under supervision through incarceration, probation, or parole, while countless more experience the collateral consequences of arrests and criminal charges. Achieving lasting public safety requires confronting both crime itself and the collateral consequences of the U.S. criminal justice system.

To help meet these dual challenges, J-PAL North America — a regional office of MIT’s Abdul Latif Jameel Poverty Action Lab (J-PAL) — with generous grant support from Arnold Ventures, launched the Initiative for Effective US Crime Policy (IECP). This initiative will generate rigorous evidence on strategies to make communities safer, reduce discrimination, and improve outcomes at every stage of the criminal justice process.

“There are a lot of open questions. We desperately need to be trying new solutions, but we need to try them in a way that enables us to learn whether they work,” notes Jennifer Doleac, executive vice president of criminal justice at Arnold Ventures. “There is a path forward for us to step up and make a concerted effort to make sure that we are being very strategic in how we spend our time and where we are directing our resources.”

Building on more than a decade of pioneering randomized evaluations, J‑PAL North America’s IECP will fund rigorous new studies in the criminal justice space, offer hands-on technical assistance, and connect researchers with practitioners. By reviewing both established and emerging evidence, the initiative will also help decision-makers focus resources on interventions that demonstrably improve public safety.

“Through this initiative, we aim to expand the use of rigorous existing evidence and help scale interventions that are proven to improve outcomes, from prevention to reintegration,” says Sara Heller, associate professor of economics at the University of Michigan and co-chair of IECP. “At the same time, IECP seeks to fill critical gaps in the evidence base by supporting new research on what works to improve the criminal justice system in the United States.”

A platform for collaboration

In June at the MIT Museum, IECP convened over 70 researchers, policymakers, and practitioners to identify research priorities and catalyze collaboration. Speakers explored the structural drivers of violence, effective pathways for translating evidence into policy, and strategies for establishing successful partnerships between researchers and practitioners.

Speakers also reflected on the value and limits of existing evidence and discussed areas in which randomized evaluations can help address the most pressing questions. Randomized evaluations have contributed powerful insights in areas such as summer youth employment programs, reminders to increase court appearances, hot-spot policing, and the use of body-worn cameras. Yet many important questions remain unanswered.

“We know randomized evaluations can answer hard policy questions, but only if we ask the right questions, with the right lens, at the right scale,” says Amanda Agan, associate professor at Cornell University and co-chair of IECP. “This convening was a call to push further: to design studies that are not only rigorous, but also relevant to the lived experiences of communities and the structural forces that shape public safety."

How to take part

Are you a practitioner with a promising idea in the criminal justice space, a policymaker planning a new program, a researcher developing a real-world intervention, or a funder investing in rigorous empirical evidence? IECP supports research partnerships to advance scalable, evidence-based solutions in the criminal legal system by funding impact evaluations, connecting researchers and practitioners, and supporting the design of randomized evaluations and the dissemination of evidence.   

To learn more about this initiative, please contact iecp@povertyactionlab.org.

Environmental scientists are increasingly using enormous artificial intelligence models to make predictions about changes in weather and climate, but a new study by MIT researchers shows that bigger models are not always better.

The team demonstrates that, in certain climate scenarios, much simpler, physics-based models can generate more accurate predictions than state-of-the-art deep-learning models.

Their analysis also reveals that a benchmarking technique commonly used to evaluate machine-learning techniques for climate predictions can be distorted by natural variations in the data, like fluctuations in weather patterns. This could lead someone to believe a deep-learning model makes more accurate predictions when that is not the case.

The researchers developed a more robust way of evaluating these techniques, which shows that, while simple models are more accurate when estimating regional surface temperatures, deep-learning approaches can be the best choice for estimating local rainfall.

They used these results to enhance a simulation tool known as a climate emulator, which can rapidly simulate the effect of human activities onto a future climate.

The researchers see their work as a “cautionary tale” about the risk of deploying large AI models for climate science. While deep-learning models have shown incredible success in domains such as natural language, climate science contains a proven set of physical laws and approximations, and the challenge becomes how to incorporate those into AI models.

“We are trying to develop models that are going to be useful and relevant for the kinds of things that decision-makers need going forward when making climate policy choices. While it might be attractive to use the latest, big-picture machine-learning model on a climate problem, what this study shows is that stepping back and really thinking about the problem fundamentals is important and useful,” says study senior author Noelle Selin, a professor in the MIT Institute for Data, Systems, and Society (IDSS) and the Department of Earth, Atmospheric and Planetary Sciences (EAPS), and director of the Center for Sustainability Science and Strategy.

Selin’s co-authors are lead author Björn Lütjens, a former EAPS postdoc who is now a research scientist at IBM Research; senior author Raffaele Ferrari, the Cecil and Ida Green Professor of Oceanography in EAPS and co-director of the Lorenz Center; and Duncan Watson-Parris, assistant professor at the University of California at San Diego. Selin and Ferrari are also co-principal investigators of the Bringing Computation to the Climate Challenge project, out of which this research emerged. The paper appears today in the Journal of Advances in Modeling Earth Systems.

Comparing emulators

Because the Earth’s climate is so complex, running a state-of-the-art climate model to predict how pollution levels will impact environmental factors like temperature can take weeks on the world’s most powerful supercomputers.

Scientists often create climate emulators, simpler approximations of a state-of-the art climate model, which are faster and more accessible. A policymaker could use a climate emulator to see how alternative assumptions on greenhouse gas emissions would affect future temperatures, helping them develop regulations.

But an emulator isn’t very useful if it makes inaccurate predictions about the local impacts of climate change. While deep learning has become increasingly popular for emulation, few studies have explored whether these models perform better than tried-and-true approaches.

The MIT researchers performed such a study. They compared a traditional technique called linear pattern scaling (LPS) with a deep-learning model using a common benchmark dataset for evaluating climate emulators.

Their results showed that LPS outperformed deep-learning models on predicting nearly all parameters they tested, including temperature and precipitation.

“Large AI methods are very appealing to scientists, but they rarely solve a completely new problem, so implementing an existing solution first is necessary to find out whether the complex machine-learning approach actually improves upon it,” says Lütjens.

Some initial results seemed to fly in the face of the researchers’ domain knowledge. The powerful deep-learning model should have been more accurate when making predictions about precipitation, since those data don’t follow a linear pattern.

They found that the high amount of natural variability in climate model runs can cause the deep learning model to perform poorly on unpredictable long-term oscillations, like El Niño/La Niña. This skews the benchmarking scores in favor of LPS, which averages out those oscillations.

Constructing a new evaluation

From there, the researchers constructed a new evaluation with more data that address natural climate variability. With this new evaluation, the deep-learning model performed slightly better than LPS for local precipitation, but LPS was still more accurate for temperature predictions.

“It is important to use the modeling tool that is right for the problem, but in order to do that you also have to set up the problem the right way in the first place,” Selin says.

Based on these results, the researchers incorporated LPS into a climate emulation platform to predict local temperature changes in different emission scenarios.

“We are not advocating that LPS should always be the goal. It still has limitations. For instance, LPS doesn’t predict variability or extreme weather events,” Ferrari adds.

Rather, they hope their results emphasize the need to develop better benchmarking techniques, which could provide a fuller picture of which climate emulation technique is best suited for a particular situation.

“With an improved climate emulation benchmark, we could use more complex machine-learning methods to explore problems that are currently very hard to address, like the impacts of aerosols or estimations of extreme precipitation,” Lütjens says.

Ultimately, more accurate benchmarking techniques will help ensure policymakers are making decisions based on the best available information.

The researchers hope others build on their analysis, perhaps by studying additional improvements to climate emulation methods and benchmarks. Such research could explore impact-oriented metrics like drought indicators and wildfire risks, or new variables like regional wind speeds.

This research is funded, in part, by Schmidt Sciences, LLC, and is part of the MIT Climate Grand Challenges team for “Bringing Computation to the Climate Challenge.”

Back in the 17th century, German astronomer Johannes Kepler figured out the laws of motion that made it possible to accurately predict where our solar system’s planets would appear in the sky as they orbit the sun. But it wasn’t until decades later, when Isaac Newton formulated the universal laws of gravitation, that the underlying principles were understood. Although they were inspired by Kepler’s laws, they went much further, and made it possible to apply the same formulas to everything from the trajectory of a cannon ball to the way the moon’s pull controls the tides on Earth — or how to launch a satellite from Earth to the surface of the moon or planets.

Today’s sophisticated artificial intelligence systems have gotten very good at making the kind of specific predictions that resemble Kepler’s orbit predictions. But do they know why these predictions work, with the kind of deep understanding that comes from basic principles like Newton’s laws? As the world grows ever-more dependent on these kinds of AI systems, researchers are struggling to try to measure just how they do what they do, and how deep their understanding of the real world actually is.

Now, researchers in MIT’s Laboratory for Information and Decision Systems (LIDS) and at Harvard University have devised a new approach to assessing how deeply these predictive systems understand their subject matter, and whether they can apply knowledge from one domain to a slightly different one. And by and large the answer at this point, in the examples they studied, is — not so much.

The findings were presented at the International Conference on Machine Learning, in Vancouver, British Columbia, last month by Harvard postdoc Keyon Vafa, MIT graduate student in electrical engineering and computer science and LIDS affiliate Peter G. Chang, MIT assistant professor and LIDS principal investigator Ashesh Rambachan, and MIT professor, LIDS principal investigator, and senior author Sendhil Mullainathan.

“Humans all the time have been able to make this transition from good predictions to world models,” says Vafa, the study’s lead author. So the question their team was addressing was, “have foundation models — has AI — been able to make that leap from predictions to world models? And we’re not asking are they capable, or can they, or will they. It’s just, have they done it so far?” he says.

“We know how to test whether an algorithm predicts well. But what we need is a way to test for whether it has understood well,” says Mullainathan, the Peter de Florez Professor with dual appointments in the MIT departments of Economics and Electrical Engineering and Computer Science and the senior author on the study. “Even defining what understanding means was a challenge.” 

In the Kepler versus Newton analogy, Vafa says, “they both had models that worked really well on one task, and that worked essentially the same way on that task. What Newton offered was ideas that were able to generalize to new tasks.” That capability, when applied to the predictions made by various AI systems, would entail having it develop a world model so it can “transcend the task that you’re working on and be able to generalize to new kinds of problems and paradigms.”

Another analogy that helps to illustrate the point is the difference between centuries of accumulated knowledge of how to selectively breed crops and animals, versus Gregor Mendel’s insight into the underlying laws of genetic inheritance.

“There is a lot of excitement in the field about using foundation models to not just perform tasks, but to learn something about the world,” for example in the natural sciences, he says. “It would need to adapt, have a world model to adapt to any possible task.”

Are AI systems anywhere near the ability to reach such generalizations? To test the question, the team looked at different examples of predictive AI systems, at different levels of complexity. On the very simplest of examples, the systems succeeded in creating a realistic model of the simulated system, but as the examples got more complex that ability faded fast.

The team developed a new metric, a way of measuring quantitatively how well a system approximates real-world conditions. They call the measurement inductive bias — that is, a tendency or bias toward responses that reflect reality, based on inferences developed from looking at vast amounts of data on specific cases.

The simplest level of examples they looked at was known as a lattice model. In a one-dimensional lattice, something can move only along a line. Vafa compares it to a frog jumping between lily pads in a row. As the frog jumps or sits, it calls out what it’s doing — right, left, or stay. If it reaches the last lily pad in the row, it can only stay or go back. If someone, or an AI system, can just hear the calls, without knowing anything about the number of lily pads, can it figure out the configuration? The answer is yes: Predictive models do well at reconstructing the “world” in such a simple case. But even with lattices, as you increase the number of dimensions, the systems no longer can make that leap.

“For example, in a two-state or three-state lattice, we showed that the model does have a pretty good inductive bias toward the actual state,” says Chang. “But as we increase the number of states, then it starts to have a divergence from real-world models.”

A more complex problem is a system that can play the board game Othello, which involves players alternately placing black or white disks on a grid. The AI models can accurately predict what moves are allowable at a given point, but it turns out they do badly at inferring what the overall arrangement of pieces on the board is, including ones that are currently blocked from play.

The team then looked at five different categories of predictive models actually in use, and again, the more complex the systems involved, the more poorly the predictive modes performed at matching the true underlying world model.

With this new metric of inductive bias, “our hope is to provide a kind of test bed where you can evaluate different models, different training approaches, on problems where we know what the true world model is,” Vafa says. If it performs well on these cases where we already know the underlying reality, then we can have greater faith that its predictions may be useful even in cases “where we don’t really know what the truth is,” he says.

People are already trying to use these kinds of predictive AI systems to aid in scientific discovery, including such things as properties of chemical compounds that have never actually been created, or of potential pharmaceutical compounds, or for predicting the folding behavior and properties of unknown protein molecules. “For the more realistic problems,” Vafa says, “even for something like basic mechanics, we found that there seems to be a long way to go.”

Chang says, “There’s been a lot of hype around foundation models, where people are trying to build domain-specific foundation models — biology-based foundation models, physics-based foundation models, robotics foundation models, foundation models for other types of domains where people have been collecting a ton of data” and training these models to make predictions, “and then hoping that it acquires some knowledge of the domain itself, to be used for other downstream tasks.”

This work shows there’s a long way to go, but it also helps to show a path forward. “Our paper suggests that we can apply our metrics to evaluate how much the representation is learning, so that we can come up with better ways of training foundation models, or at least evaluate the models that we’re training currently,” Chang says. “As an engineering field, once we have a metric for something, people are really, really good at optimizing that metric.”

For both research and medical purposes, researchers have spent decades pushing the limits of microscopy to produce ever deeper and sharper images of brain activity, not only in the cortex but also in regions underneath, such as the hippocampus. In a new study, a team of MIT scientists and engineers demonstrates a new microscope system capable of peering exceptionally deep into brain tissues to detect the molecular activity of individual cells by using sound.

“The major advance here is to enable us to image deeper at single-cell resolution,” says neuroscientist Mriganka Sur, a corresponding author along with mechanical engineering professor Peter So and principal research scientist Brian Anthony. Sur is the Paul and Lilah Newton Professor in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT.

In the journal Light: Science and Applications, the team demonstrates that they could detect NAD(P)H, a molecule tightly associated with cell metabolism in general and electrical activity in neurons in particular, all the way through samples such as a 1.1-millimeter “cerebral organoid,” a 3D-mini brain-like tissue generated from human stem cells, and a 0.7-milimeter-thick slice of mouse brain tissue.

In fact, says co-lead author and mechanical engineering postdoc W. David Lee, who conceived the microscope’s innovative design, the system could have peered far deeper, but the test samples weren’t big enough to demonstrate that.

“That’s when we hit the glass on the other side,” he says. “I think we’re pretty confident about going deeper.”

Still, a depth of 1.1 milimeters is more than five times deeper than other microscope technologies can resolve NAD(P)H within dense brain tissue. The new system achieved the depth and sharpness by combining several advanced technologies to precisely and efficiently excite the molecule and then to detect the resulting energy, all without having to add any external labels, either via added chemicals or genetically engineered fluorescence.

Rather than focusing the required NAD(P)H excitation energy on a neuron with near ultraviolet light at its normal peak absorption, the scope accomplishes the excitation by focusing an intense, extremely short burst of light (a quadrillionth of a second long) at three times the normal absorption wavelength. Such “three-photon” excitation penetrates deep into tissue with less scattering by brain tissue because of the longer wavelength of the light (“like fog lamps,” Sur says). Meanwhile, although the excitation produces a weak fluorescent signal of light from NAD(P)H, most of the absorbed energy produces a localized (about 10 microns) thermal expansion within the cell, which produces sound waves that travel relatively easily through tissue compared to the fluorescence emission. A sensitive ultrasound microphone in the microscope detects those waves and, with enough sound data, software turns them into high-resolution images (much like a sonogram does). Imaging created in this way is “three-photon photoacoustic imaging.”

“We merged all these techniques — three-photon, label-free, photoacoustic detection,” says co-lead author Tatsuya Osaki, a research scientist in the Picower Institute in Sur’s lab. “We integrated all these cutting-edge techniques into one process to establish this ‘Multiphoton-In and Acoustic-Out’ platform.”

Lee and Osaki combined with research scientist Xiang Zhang and postdoc Rebecca Zubajlo to lead the study, in which the team demonstrated reliable detection of the sound signal through the samples. So far, the team has produced visual images from the sound at various depths as they refine their signal processing.

In the study, the team also shows simultaneous “third-harmonic generation” imaging, which comes from the three-photon stimulation and finely renders cellular structures, alongside their photoacoustic imaging, which detects NAD(P)H. They also note that their photoacoustic method could detect other molecules such as the genetically encoded calcium indicator GCaMP, that neuroscientists use to signal neural electrical activity.

With the concept of label-free, multiphoton, photoacoustic microscopy (LF-MP-PAM) established in the paper, the team is now looking ahead to neuroscience and clinical applications.

For instance, through the company Precision Healing, Inc., which he founded and sold, Lee has already established that NAD(P)H imaging can inform wound care. In the brain, levels of the molecule are known to vary in conditions such as Alzheimer’s disease, Rett syndrome, and seizures, making it a potentially valuable biomarker. Because the new system is label-free (i.e., no added chemicals or altered genes), it could be used in humans, for instance, during brain surgeries.

The next step for the team is to demonstrate it in a living animal, rather than just in in vitro and ex-vivo tissues. The technical challenge there is that the microphone can no longer be on the opposite side of the sample from the light source (as it was in the current study). It has to be on top, just like the light source.

Lee says he expects that full imaging at depths of 2 milimeters in live brains is entirely feasible, given the results in the new study.

“In principle, it should work,” he says.

Mercedes Balcells and Elazer Edelman are also authors of the paper. Funding for the research came from sources including the National Institutes of Health, the Simon Center for the Social Brain, the lab of Peter So, The Picower Institute for Learning and Memory, and the Freedom Together Foundation.

A fast radio burst is an immense flash of radio emission that lasts for just a few milliseconds, during which it can momentarily outshine every other radio source in its galaxy. These flares can be so bright that their light can be seen from halfway across the universe, several billion light years away.

The sources of these brief and dazzling signals are unknown. But scientists now have a chance to study a fast radio burst (FRB) in unprecedented detail. An international team of scientists including physicists at MIT have detected a near and ultrabright fast radio burst some 130 million light-years from Earth in the constellation Ursa Major. It is one of the closest FRBs detected to date. It is also the brightest — so bright that the signal has garnered the informal moniker, RBFLOAT, for “radio brightest flash of all time.”

The burst’s brightness, paired with its proximity, is giving scientists the closest look yet at FRBs and the environments from which they emerge.

“Cosmically speaking, this fast radio burst is just in our neighborhood,” says Kiyoshi Masui, associate professor of physics and affiliate of MIT’s Kavli Institute for Astrophysics and Space Research. “This means we get this chance to study a pretty normal FRB in exquisite detail.”

Masui and his colleagues report their findings today in the Astrophysical Journal Letters.

Diverse bursts

The clarity of the new detection is thanks to a significant upgrade to The Canadian Hydrogen Intensity Mapping Experiment (CHIME), a large array of halfpipe-shaped antennae based in British Columbia. CHIME was originally designed to detect and map the distribution of hydrogen across the universe. The telescope is also sensitive to ultrafast and bright radio emissions. Since it started observations in 2018, CHIME has detected about 4,000 fast radio bursts, from all parts of the sky. But the telescope had not been able to precisely pinpoint the location of each fast radio burst, until now.

CHIME recently got a significant boost in precision, in the form of CHIME Outriggers — three miniature versions of CHIME, each sited in different parts of North America. Together, the telescopes work as one continent-sized system that can focus in on any bright flash that CHIME detects, to pin down its location in the sky with extreme precision.

“Imagine we are in New York and there’s a firefly in Florida that is bright for a thousandth of a second, which is usually how quick FRBs are,” says MIT Kavli graduate student Shion Andrew. “Localizing an FRB to a specific part of its host galaxy is analogous to figuring out not just what tree the firefly came from, but which branch it’s sitting on.”

The new fast radio burst is the first detection made using the combination of CHIME and the completed CHIME Outriggers. Together, the telescope array identified the FRB and determined not only the specific galaxy, but also the region of the galaxy from where the burst originated. It appears that the burst arose from the edge of the galaxy, just outside of a star-forming region. The precise localization of the FRB is allowing scientists to study the environment around the signal for clues to what brews up such bursts.

“As we’re getting these much more precise looks at FRBs, we’re better able to see the diversity of environments they’re coming from,” says MIT physics postdoc Adam Lanman.

Lanman, Andrew, and Masui are members of the CHIME Collaboration — which includes scientists from multiple institutions around the world — and are authors of the new paper detailing the discovery of the new FRB detection.

An older edge

Each of CHIME’s Outrigger stations continuously monitors the same swath of sky as the parent CHIME array. Both CHIME and the Outriggers “listen” for radio flashes, at incredibly short, millisecond timescales. Even over several minutes, such precision monitoring can amount to a huge amount of data. If CHIME detects no FRB signal, the Outriggers automatically delete the last 40 seconds of data to make room for the next span of measurements.

On March 16, 2025, CHIME detected an ultrabright flash of radio emissions, which automatically triggered the CHIME Outriggers to record the data. Initially, the flash was so bright that astronomers were unsure whether it was an FRB or simply a terrestrial event caused, for instance, by a burst of cellular communications.

That notion was put to rest as the CHIME Outrigger telescopes focused in on the flash and pinned down its location to NGC4141 — a spiral galaxy in the constellation Ursa Major about 130 million light years away, which happens to be surprisingly close to our own Milky Way. The detection is one of the closest and brightest fast radio bursts detected to date.

Follow-up observations in the same region revealed that the burst came from the very edge of an active region of star formation. While it’s still a mystery as to what source could produce FRBs, scientists’ leading hypothesis points to magnetars — young neutron stars with extremely powerful magnetic fields that can spin out high-energy flares across the electromagnetic spectrum, including in the radio band. Physicists suspect that magnetars are found in the center of star formation regions, where the youngest, most active stars are forged. The location of the new FRB, just outside a star-forming region in its galaxy, may suggest that the source of the burst is a slightly older magnetar.

“These are mostly hints,” Masui says. “But the precise localization of this burst is letting us dive into the details of how old an FRB source could be. If it were right in the middle, it would only be thousands of years old — very young for a star. This one, being on the edge, may have had a little more time to bake.”

No repeats

In addition to pinpointing where the new FRB was in the sky, the scientists also looked back through CHIME data to see whether any similar flares occurred in the same region in the past. Since the first FRB was discovered in 2007, astronomers have detected over 4,000 radio flares. Most of these bursts are one-offs. But a few percent have been observed to repeat, flashing every so often. And an even smaller fraction of these repeaters flash in a pattern, like a rhythmic heartbeat, before flaring out. A central question surrounding fast radio bursts is whether repeaters and nonrepeaters come from different origins.

The scientists looked through CHIME’s six years of data and came up empty: This new FRB appears to be a one-off, at least in the last six years. The findings are particularly exciting, given the burst’s proximity. Because it is so close and so bright, scientists can probe the environment in and around the burst for clues to what might produce a nonrepeating FRB.

“Right now we’re in the middle of this story of whether repeating and nonrepeating FRBs are different. These observations are putting together bits and pieces of the puzzle,” Masui says.

“There’s evidence to suggest that not all FRB progenitors are the same,” Andrew adds. “We’re on track to localize hundreds of FRBs every year. The hope is that a larger sample of FRBs localized to their host environments can help reveal the full diversity of these populations.”

The construction of the CHIME Outriggers was funded by the Gordon and Betty Moore Foundation and the U.S. National Science Foundation. The construction of CHIME was funded by the Canada Foundation for Innovation and provinces of Quebec, Ontario, and British Columbia.

Rising global temperatures affect human activity in many ways. Now, a new study illuminates an important dimension of the problem: Very hot days are associated with more negative moods, as shown by a large-scale look at social media postings.

Overall, the study examines 1.2 billion social media posts from 157 countries over the span of a year. The research finds that when the temperature rises above 95 degrees Fahrenheit, or 35 degrees Celsius, expressed sentiments become about 25 percent more negative in lower-income countries and about 8 percent more negative in better-off countries. Extreme heat affects people emotionally, not just physically.

“Our study reveals that rising temperatures don’t just threaten physical health or economic productivity — they also affect how people feel, every day, all over the world,” says Siqi Zheng, a professor in MIT’s Department of Urban Studies and Planning (DUSP) and Center for Real Estate (CRE), and co-author of a new paper detailing the results. “This work opens up a new frontier in understanding how climate stress is shaping human well-being at a planetary scale.”

The paper, “Unequal Impacts of Rising Temperatures on Global Human Sentiment,” is published today in the journal One Earth. The authors are Jianghao Wang, of the Chinese Academy of Sciences; Nicolas Guetta-Jeanrenaud SM ’22, a graduate of MIT’s Technology and Policy Program (TPP) and Institute for Data, Systems, and Society; Juan Palacios, a visiting assistant professor at MIT’s Sustainable Urbanization Lab (SUL) and an assistant professor Maastricht University; Yichun Fan, of SUL and Duke University; Devika Kakkar, of Harvard University; Nick Obradovich, of SUL and the Laureate Institute for Brain Research in Tulsa; and Zheng, who is the STL Champion Professor of Urban and Real Estate Sustainability at CRE and DUSP. Zheng is also the faculty director of CRE and founded the Sustainable Urbanization Lab in 2019.

Social media as a window

To conduct the study, the researchers evaluated 1.2 billion posts from the social media platforms Twitter and Weibo, all of which appeared in 2019. They used a natural language processing technique called Bidirectional Encoder Representations from Transformers (BERT), to analyze 65 languages across the 157 countries in the study.

Each social media post was given a sentiment rating from 0.0 (for very negative posts) to 1.0 (for very positive posts). The posts were then aggregated geographically to 2,988 locations and evaluated in correlation with area weather. From this method, the researchers could then deduce the connection between extreme temperatures and expressed sentiment.

“Social media data provides us with an unprecedented window into human emotions across cultures and continents,” Wang says. “This approach allows us to measure emotional impacts of climate change at a scale that traditional surveys simply cannot achieve, giving us real-time insights into how temperature affects human sentiment worldwide.”

To assess the effects of temperatures on sentiment in higher-income and middle-to-lower-income settings, the scholars also used a World Bank cutoff level of gross national income per-capita annual income of $13,845, finding that in places with incomes below that, the effects of heat on mood were triple those found in economically more robust settings.

“Thanks to the global coverage of our data, we find that people in low- and middle-income countries experience sentiment declines from extreme heat that are three times greater than those in high-income countries,” Fan says. “This underscores the importance of incorporating adaptation into future climate impact projections.”

In the long run

Using long-term global climate models, and expecting some adaptation to heat, the researchers also produced a long-range estimate of the effects of extreme temperatures on sentiment by the year 2100. Extending the current findings to that time frame, they project a 2.3 percent worsening of people’s emotional well-being based on high temperatures alone by then — although that is a far-range projection.

“It's clear now, with our present study adding to findings from prior studies, that weather alters sentiment on a global scale,” Obradovich says. “And as weather and climates change, helping individuals become more resilient to shocks to their emotional states will be an important component of overall societal adaptation.”

The researchers note that there are many nuances to the subject, and room for continued research in this area. For one thing, social media users are not likely to be a perfectly representative portion of the population, with young children and the elderly almost certainly using social media less than other people. However, as the researchers observe in the paper, the very young and elderly are probably particularly vulnerable to heat shocks, making the response to hot weather possible even larger than their study can capture.

The research is part of the Global Sentiment project led by the MIT Sustainable Urbanization Lab, and the study’s dataset is publicly available. Zheng and other co-authors have previously investigated these dynamics using social media, although never before at this scale.

“We hope this resource helps researchers, policymakers, and communities better prepare for a warming world,” Zheng says.

The research was supported, in part, by Zheng’s chaired professorship research fund, and grants Wang received from the National Natural Science Foundation of China and the Chinese Academy of Sciences. 

A multiyear program at MIT Lincoln Laboratory to characterize how biological and chemical vapors and aerosols disperse through the New York City subway system is coming to a close. The program, part of the U.S. Department of Homeland Security (DHS) Science and Technology Directorate's Urban Area Security Initiative, builds on other efforts at Lincoln Laboratory to detect chemical and biological threats, validate air dispersion models, and improve emergency protocols in urban areas in case of an airborne attack. The results of this program will inform the New York Metropolitan Transportation Authority (MTA) on how best to install an efficient, cost-effective system for airborne threat detection and mitigation throughout the subway. On a broader scale, the study will help the national security community understand pragmatic chemical and biological defense options for mass transit, critical facilities, and special events.

Trina Vian from the laboratory's Counter–Weapons of Mass Destruction (WMD) Systems Group led this project, which she says had as much to do with air flow and sensors as it did with MTA protocols and NYC commuters. "There are real dangers associated with panic during an alarm. People can get hurt during mass evacuation, or lose trust in a system and the authorities that administer that system, if there are false alarms," she says. "A novel aspect of our project was to investigate effective low-regret response options, meaning those with little operational consequence to responding to a false alarm."

Currently, depending on the severity of the alarm, the MTA's response can include stopping service and evacuating passengers and employees.

A complex environment for testing

For the program, which started in 2019, Vian and her team collected data on how chemical and biological sensors performed in the subway, what factors affected sensor accuracy, and how different mitigation protocols fared in stopping an airborne threat from spreading and removing the threat from a contaminated location. For their tests, they released batches of a safe, custom-developed aerosol simulant within Grand Central Station that they could track with DNA barcodes. Each batch had a different barcode, which allowed the team to differentiate among them and quantitatively assess different combinations of mitigation strategies.

To control and isolate air flow, the team tested static air curtains as well as air filtration systems. They also tested a spray knockdown system developed by Sandia National Laboratories designed to reduce and isolate particulate hazards in large volume areas. The system sprays a fine water mist into the tunnels that attaches to threat particulates and uses gravity to rain out the threat material. The spray contains droplets of a particular size and concentration, and with an applied electrostatic field. The original idea for the system was adapted from the coal mining industry, which used liquid sprayers to reduce the amount of inhalable soot.

The tests were done in a busy environment, and the team was required to complete trainings on MTA protocols such as track safety and how to interact with the public.

"We had long and sometimes very dirty days," says Jason Han of the Counter–WMD Systems Group, who collected measurements in the tunnels and analyzed the data. "We all wore bright orange contractor safety vests, which made people think we were official employees of the MTA. We would often get approached by people asking for directions!"

At times, issues such as power outages or database errors could disrupt data capture.

"We learned fairly early on that we had to capture daily data backups and keep a daily evolving master list of unique sensor identifiers and locations," says fellow team member Cassie Smith. "We developed workflows and wrote scripts to help automate the process, which ensured successful sensor data capture and attribution."

The team also worked closely with the MTA to make sure their tests and data capture ran smoothly. "The MTA was great at helping us maintain the test bed, doing as much as they could in our physical absence," Vian says.

Calling on industry

Another crucial aspect of the program was to connect with the greater chemical and biological industrial community to solicit their sensors for testing. These partnerships reduced the cost for DHS to bring new sensing technologies into the project, and, in return, participants gained a testing and data collection opportunity within the challenging NYC subway environment.

The team ultimately fielded 16 different sensors, each with varying degrees of maturity, that operated through a range of methods, such as ultraviolet laser–induced fluorescence, polymerase chain reaction, and long-wave infrared spectrometry.

"The partners appreciated the unique data they got and the opportunity to work with the MTA and experience an environment and customer base that they may not have anticipated before," Vian says.

The team finished testing in 2024 and has delivered the final report to the DHS. The MTA will use the report to help expand their PROTECT chemical detection system (originally developed by Argonne National Laboratory) from Grand Central Station into adjacent stations. They expect to complete this work in 2026.

"The value of this program cannot be overstated. This partnership with DHS and MIT Lincoln Laboratory has led to the identification of the best-suited systems for the MTA’s unique operating environment," says Michael Gemelli, director of chemical, biological, radiological, and nuclear/WMD detection and mitigation at the New York MTA.

"Other transit authorities can leverage these results to start building effective chemical and biological defense systems for their own specific spaces and threat priorities," adds Benjamin Ervin, leader of Lincoln Laboratory's Counter–WMD Systems Group. "Specific test and evaluation within the operational environment of interest, however, is always recommended to ensure defense system objectives are met."

Building these types of decision-making reports for airborne chemical and biological sensing has been a part of Lincoln Laboratory's mission since the mid-1990s. The laboratory also helped to define priorities in the field when DHS was forming in the early 2000s.

Beyond this study, the Lincoln Laboratory is leading several other projects focused on forecasting the impact of novel chemical and biological threats within multiple domains — military, space, agriculture, health, etc. — and on prototyping rapid, autonomous, high-confidence biological identification capabilities for the homeland to provide actionable evidence of hazardous environments.

From toddlers’ timeouts to criminals’ prison sentences, punishment reinforces social norms, making it known that an offender has done something unacceptable. At least, that is usually the intent — but the strategy can backfire. When a punishment is perceived as too harsh, observers can be left with the impression that an authority figure is motivated by something other than justice.

It can be hard to predict what people will take away from a particular punishment, because everyone makes their own inferences not just about the acceptability of the act that led to the punishment, but also the legitimacy of the authority who imposed it. A new computational model developed by scientists at MIT’s McGovern Institute for Brain Research makes sense of these complicated cognitive processes, recreating the ways people learn from punishment and revealing how their reasoning is shaped by their prior beliefs.

Their work, reported Aug. 4 in the journal PNAS, explains how a single punishment can send different messages to different people, and even strengthen the opposing viewpoints of groups who hold different opinions about authorities or social norms.

“The key intuition in this model is the fact that you have to be evaluating simultaneously both the norm to be learned and the authority who’s punishing,” says McGovern investigator and John W. Jarve Professor of Brain and Cognitive Sciences Rebecca Saxe, who led the research. “One really important consequence of that is even where nobody disagrees about the facts — everybody knows what action happened, who punished it, and what they did to punish it — different observers of the same situation could come to different conclusions.”

For example, she says, a child who is sent to timeout after biting a sibling might interpret the event differently than the parent. One might see the punishment as proportional and important, teaching the child not to bite. But if the biting, to the toddler, seemed a reasonable tactic in the midst of a squabble, the punishment might be seen as unfair, and the lesson will be lost.

People draw on their own knowledge and opinions when they evaluate these situations — but to study how the brain interprets punishment, Saxe and graduate student Setayesh Radkani wanted to take those personal ideas out of the equation. They needed a clear understanding of the beliefs that people held when they observed a punishment, so they could learn how different kinds of information altered their perceptions. So Radkani set up scenarios in imaginary villages where authorities punished individuals for actions that had no obvious analog in the real world.

Participants observed these scenarios in a series of experiments, with different information offered in each one. In some cases, for example, participants were told that the person being punished was either an ally or a competitor of the authority, whereas in other cases, the authority’s possible bias was left ambiguous.

“That gives us a really controlled setup to vary prior beliefs,” Radkani explains. “We could ask what people learn from observing punitive decisions with different severities, in response to acts that vary in their level of wrongness, by authorities that vary in their level of different motives.”

For each scenario, participants were asked to evaluate four factors: how much the authority figure cared about justice; the selfishness of the authority; the authority’s bias for or against the individual being punished; and the wrongness of the punished act. The research team asked these questions when participants were first introduced to the hypothetical society, then tracked how their responses changed after they observed the punishment. Across the scenarios, participants’ initial beliefs about the authority and the wrongness of the act shaped the extent to which those beliefs shifted after they observed the punishment.

Radkani was able to replicate these nuanced interpretations using a cognitive model framed around an idea that Saxe’s team has long used to think about how people interpret the actions of others. That is, to make inferences about others’ intentions and beliefs, we assume that people choose actions that they expect will help them achieve their goals.

To apply that concept to the punishment scenarios, Radkani developed a model that evaluates the meaning of a punishment (an action aimed at achieving a goal of the authority) by considering the harm associated with that punishment; its costs or benefits to the authority; and its proportionality to the violation. By assessing these factors, along with prior beliefs about the authority and the punished act, the model was able to predict people’s responses to the hypothetical punishment scenarios, supporting the idea that people use a similar mental model. “You need to have them consider those things, or you can’t make sense of how people understand punishment when they observe it,” Saxe says.

Even though the team designed their experiments to preclude preconceived ideas about the people and actions in their imaginary villages, not everyone drew the same conclusions from the punishments they observed. Saxe’s group found that participants’ general attitudes toward authority influenced their interpretation of events. Those with more authoritarian attitudes — assessed through a standard survey — tended to judge punished acts as more wrong and authorities as more motivated by justice than other observers.

“If we differ from other people, there’s a knee-jerk tendency to say, ‘either they have different evidence from us, or they’re crazy,’” Saxe says. Instead, she says, “It’s part of the way humans think about each other’s actions.”

“When a group of people who start out with different prior beliefs get shared evidence, they will not end up necessarily with shared beliefs. That’s true even if everybody is behaving rationally,” says Saxe.

This way of thinking also means that the same action can simultaneously strengthen opposing viewpoints. The Saxe lab’s modeling and experiments showed that when those viewpoints shape individuals’ interpretations of future punishments, the groups’ opinions will continue to diverge. For instance, a punishment that seems too harsh to a group who suspects an authority is biased can make that group even more skeptical of the authority’s future actions. Meanwhile, people who see the same punishment as fair and the authority as just will be more likely to conclude that the authority figure’s future actions are also just. 

“You will get a vicious cycle of polarization, staying and actually spreading to new things,” says Radkani.

The researchers say their findings point toward strategies for communicating social norms through punishment. “It is exactly sensible in our model to do everything you can to make your action look like it’s coming out of a place of care for the long-term outcome of this individual, and that it’s proportional to the norm violation they did,” Saxe says. “That is your best shot at getting a punishment interpreted pedagogically, rather than as evidence that you’re a bully.”

Nevertheless, she says that won’t always be enough. “If the beliefs are strong the other way, it’s very hard to punish and still sustain a belief that you were motivated by justice.”

Joining Saxe and Radkani on the paper is Joshua Tenenbaum, MIT professor of brain and cognitive sciences. The study was funded, in part, by the Patrick J McGovern Foundation.

The U.S. Food and Drug Administration’s recent approval of the first CRISPR-Cas9–based gene therapy has marked a major milestone in biomedicine, validating genome editing as a promising treatment strategy for disorders like sickle cell disease, muscular dystrophy, and certain cancers.

CRISPR-Cas9, often likened to “molecular scissors,” allows scientists to cut DNA at targeted sites to snip, repair, or replace genes. But despite its power, Cas9 poses a critical safety risk: The active enzyme can linger in cells and cause unintended DNA breaks — so-called off-target effects — which may trigger harmful mutations in healthy genes.

Now, researchers in the labs of Ronald T. Raines, MIT professor of chemistry, and Amit Choudhary, professor of medicine at Harvard Medical School, have engineered a precise way to turn Cas9 off after its job is done — significantly reducing off-target effects and improving the clinical safety of gene editing. Their findings are detailed in a new paper published in the Proceedings of the National Academy of Sciences (PNAS).

“To ‘turn off’ Cas9 after it achieves its intended genome-editing outcome, we developed the first cell-permeable anti-CRISPR protein system,” says Raines, the Roger and Georges Firmenich Professor of Natural Products Chemistry. “Our technology reduces the off-target activity of Cas9 and increases its genome-editing specificity and clinical utility.”

The new tool — called LF_N-Acr/PA — uses a protein-based delivery system to ferry anti-CRISPR proteins into human cells rapidly and efficiently. While natural Type II anti-CRISPR proteins (Acrs) are known to inhibit Cas9, their use in therapy has been limited because they’re often too bulky or charged to enter cells, and conventional delivery methods are too slow or ineffective.

LF_N-Acr/PA overcomes these hurdles using a component derived from anthrax toxin to introduce Acrs into cells within minutes. Even at picomolar concentrations, the system shuts down Cas9 activity with remarkable speed and precision — boosting genome-editing specificity up to 40 percent.

Bradley L. Pentelute, MIT professor of chemistry, is an expert on the anthrax delivery system, and is also an author of the paper.

The implications of this advance are wide-ranging. With patent applications filed, LF_N-Acr/PA represents a faster, safer, and more controllable means of harnessing CRISPR-Cas9, opening the door to more-refined gene therapies with fewer unintended consequences.

The research was supported by the National Institutes of Health and a Gilliam Fellowship from the Howard Hughes Medical Institute awarded to lead author Axel O. Vera, a graduate student in the Department of Chemistry.

Materials research thrives across MIT, spanning disciplines and departments. Recent breakthroughs include strategies for securing sustainable supplies of nickel — critical to clean-energy technologies (Department of Materials Science and Engineering); the discovery of unexpected magnetism in atomically thin quantum materials (Department of Physics); and the development of adhesive coatings that reduce scarring around medical implants (departments of Mechanical Engineering and Civil and Environmental Engineering).

At the center of these efforts is the Materials Research Laboratory (MRL), a hub that connects and supports the Institute’s materials research community. “MRL serves as a home for the entire materials research community at MIT,” says C. Cem Tasan, who became director in April 2025. “Our goal is to make it easier for our faculty to conduct their extraordinary research,” adds Tasan, the POSCO Associate Professor of Metallurgy in the Department of Materials Science and Engineering.

A storied history

Established in 2017, the MRL brings together more than 30 researchers and builds on a 48-year legacy of innovation. It was formed through the merger of the MIT Materials Processing Center (MPC) and the Center for Materials Science and Engineering (CMSE), two institutions that helped lay the foundation for MIT’s global leadership in materials science.

Over the years, research supported by MPC and CMSE has led to transformative technologies and successful spinout companies. Notable examples include amsc, based on advances in superconductivity; OmniGuide, which developed cutting-edge optical fiber technologies; and QD Vision, a pioneer in quantum dot technology acquired by Samsung in 2016. Another landmark achievement was the development of the first germanium laser to operate at room temperature — a breakthrough now used in optical communications.

Enabling research through partnership and support

MRL is launching targeted initiatives to connect MIT researchers with industry partners around specific technical challenges. Each initiative will be led by a junior faculty member working closely with MRL to identify a problem that aligns with their research expertise and is relevant to industry needs.

Through multi-year collaborations with participating companies, faculty can explore early-stage solutions in partnership with postdocs or graduate students. These initiatives are designed to be agile and interdisciplinary, with the potential to grow into major, long-term research programs.

Behind-the-scenes support, front-line impact

MRL provides critical infrastructure that enables faculty to focus on discovery, not logistics. “MRL works silently in the background, where every problem a principal investigator has related to the administration of materials research is solved with efficiency, good organization, and minimum effort,” says Tasan.

This quiet but powerful support spans multiple areas:

• The finance team manages grants and helps secure new funding opportunities.
• The human resources team supports the hiring of postdocs.
• The communications team amplifies the lab’s impact through compelling stories shared with the public and funding agencies.
• The events team plans and coordinates conferences, seminars, and symposia that foster collaboration within the MIT community and with external partners.

Together, these functions ensure that research at MRL runs smoothly and effectively — from initial idea to lasting innovation.

Leadership with a vision

Tasan, who also leads a research group focused on metallurgy, says he took on the directorship because “I thrive on new challenges.” He also saw the role as an opportunity to contribute more broadly to MIT. 

“I believe MRL can play an even greater role in advancing materials research across the Institute, and I’m excited to help make that happen,” he says.

Optical frequency combs are specially designed lasers that act like rulers to accurately and rapidly measure specific frequencies of light. They can be used to detect and identify chemicals and pollutants with extremely high precision.

Frequency combs would be ideal for remote sensors or portable spectrometers because they can enable accurate, real-time monitoring of multiple chemicals without complex moving parts or external equipment.

But developing frequency combs with high enough bandwidth for these applications has been a challenge. Often, researchers must add bulky components that limit scalability and performance.

Now, a team of MIT researchers has demonstrated a compact, fully integrated device that uses a carefully crafted mirror to generate a stable frequency comb with very broad bandwidth. The mirror they developed, along with an on-chip measurement platform, offers the scalability and flexibility needed for mass-producible remote sensors and portable spectrometers. This development could enable more accurate environmental monitors that can identify multiple harmful chemicals from trace gases in the atmosphere.

“The broader the bandwidth a spectrometer has, the more powerful it is, but dispersion is in the way. Here we took the hardest problem that limits bandwidth and made it the centerpiece of our study, addressing every step to ensure robust frequency comb operation,” says Qing Hu, Distinguished Professor in Electrical Engineering and Computer Science at MIT, principal investigator in the Research Laboratory of Electronics, and senior author on an open-access paper describing the work.

He is joined on the paper by lead author Tianyi Zeng PhD ’23; as well as Yamac Dikmelik of General Dynamics Mission Systems; Feng Xie and Kevin Lascola of Thorlabs Quantum Electronics; and David Burghoff SM ’09, PhD ’14, an assistant professor at the University of Texas at Austin. The research appears today in Light: Science and Applications.

Broadband combs

An optical frequency comb produces a spectrum of equally spaced laser lines, which resemble the teeth of a comb.

Scientists can generate frequency combs using several types of lasers for different wavelengths. By using a laser that produces long wave infrared radiation, such as a quantum cascade laser, they can use frequency combs for high-resolution sensing and spectroscopy.

In dual-comb spectroscopy (DCS), the beam of one frequency comb travels straight through the system and strikes a detector at the other end. The beam of the second frequency comb passes through a chemical sample before striking the same detector. Using the results from both combs, scientists can faithfully replicate the chemical features of the sample at much lower frequencies, where signals can be easily analyzed.

The frequency combs must have high bandwidth, or they will only be able to detect a small frequency range of chemical compounds, which could lead to false alarms or inaccurate results.

Dispersion is the most important factor that limits a frequency comb’s bandwidth. If there is dispersion, the laser lines are not evenly spaced, which is incompatible with the formation of frequency combs.

“With long wave infrared radiation, the dispersion will be very high. There is no way to get around it, so we have to find a way to compensate for it or counteract it by engineering our system,” Hu says.

Many existing approaches aren’t flexible enough to be used in different scenarios or don’t enable high enough bandwidth.

Hu’s group previously solved this problem in a different type of frequency comb, one that used terahertz waves, by developing a double-chirped mirror (DCM).

A DCM is a special type of optical mirror that has multiple layers with thicknesses that change gradually from one end to the other. They found that this DCM, which has a corrugated structure, could effectively compensate for dispersion when used with a terahertz laser.

“We tried to borrow this trick and apply it to an infrared comb, but we ran into lots of challenges,” Hu says.

Because infrared waves are 10 times shorter than terahertz waves, fabricating the new mirror required an extreme level of precision. At the same time, they needed to coat the entire DCM in a thick layer of gold to remove the heat under laser operation. Plus, their dispersion measurement system, designed for terahertz waves, wouldn’t work with infrared waves, which have frequencies that are about 10 times higher than terahertz.

“After more than two years of trying to implement this scheme, we reached a dead end,” Hu says.

A new solution

Ready to throw in the towel, the team realized something they had missed. They had designed the mirror with corrugation to compensate for the lossy terahertz laser, but infrared radiation sources aren’t as lossy.

This meant they could use a standard DCM design to compensate for dispersion, which is compatible with infrared radiation. However, they still needed to create curved mirror layers to capture the beam of the laser, which made fabrication much more difficult than usual.

“The adjacent layers of mirror differ only by tens of nanometers. That level of precision precludes standard photolithography techniques. On top of that, we still had to etch very deeply into the notoriously stubborn material stacks. Achieving those critical dimensions and etch depths was key to unlocking broadband comb performance,” Zeng says. In addition to precisely fabricating the DCM, they integrated the mirror directly onto the laser, making the device extremely compact. The team also developed a high-resolution, on-chip dispersion measurement platform that doesn’t require bulky external equipment.

“Our approach is flexible. As long as we can use our platform to measure the dispersion, we can design and fabricate a DCM that compensates for it,” Hu adds.

Taken together, the DCM and on-chip measurement platform enabled the team to generate stable infrared laser frequency combs that had far greater bandwidth than can usually be achieved without a DCM.

In the future, the researchers want to extend their approach to other laser platforms that could generate combs with even greater bandwidth and higher power for more demanding applications.

“These researchers developed an ingenious nanophotonic dispersion compensation scheme based on an integrated air–dielectric double-chirped mirror. This approach provides unprecedented control over dispersion, enabling broadband comb formation at room temperature in the long-wave infrared. Their work opens the door to practical, chip-scale frequency combs for applications ranging from chemical sensing to free-space communications,” says Jacob B. Khurgin, a professor at the Johns Hopkins University Whiting School of Engineering, who was not involved with this paper.

This work is funded, in part, by the U.S. Defense Advanced Research Projects Agency (DARPA) and the Gordon and Betty Moore Foundation. This work was carried out, in part, using facilities at MIT.nano.

Using machine learning, MIT chemical engineers have created a computational model that can predict how well any given molecule will dissolve in an organic solvent — a key step in the synthesis of nearly any pharmaceutical. This type of prediction could make it much easier to develop new ways to produce drugs and other useful molecules.

The new model, which predicts how much of a solute will dissolve in a particular solvent, should help chemists to choose the right solvent for any given reaction in their synthesis, the researchers say. Common organic solvents include ethanol and acetone, and there are hundreds of others that can also be used in chemical reactions.

“Predicting solubility really is a rate-limiting step in synthetic planning and manufacturing of chemicals, especially drugs, so there’s been a longstanding interest in being able to make better predictions of solubility,” says Lucas Attia, an MIT graduate student and one of the lead authors of the new study.

The researchers have made their model freely available, and many companies and labs have already started using it. The model could be particularly useful for identifying solvents that are less hazardous than some of the most commonly used industrial solvents, the researchers say.

“There are some solvents which are known to dissolve most things. They’re really useful, but they’re damaging to the environment, and they’re damaging to people, so many companies require that you have to minimize the amount of those solvents that you use,” says Jackson Burns, an MIT graduate student who is also a lead author of the paper. “Our model is extremely useful in being able to identify the next-best solvent, which is hopefully much less damaging to the environment.”

William Green, the Hoyt Hottel Professor of Chemical Engineering and director of the MIT Energy Initiative, is the senior author of the study, which appears today in Nature Communications. Patrick Doyle, the Robert T. Haslam Professor of Chemical Engineering, is also an author of the paper.

Solving solubility

The new model grew out of a project that Attia and Burns worked on together in an MIT course on applying machine learning to chemical engineering problems. Traditionally, chemists have predicted solubility with a tool known as the Abraham Solvation Model, which can be used to estimate a molecule’s overall solubility by adding up the contributions of chemical structures within the molecule. While these predictions are useful, their accuracy is limited.

In the past few years, researchers have begun using machine learning to try to make more accurate solubility predictions. Before Burns and Attia began working on their new model, the state-of-the-art model for predicting solubility was a model developed in Green’s lab in 2022.

That model, known as SolProp, works by predicting a set of related properties and combining them, using thermodynamics, to ultimately predict the solubility. However, the model has difficulty predicting solubility for solutes that it hasn’t seen before.

“For drug and chemical discovery pipelines where you’re developing a new molecule, you want to be able to predict ahead of time what its solubility looks like,” Attia says.

Part of the reason that existing solubility models haven’t worked well is because there wasn’t a comprehensive dataset to train them on. However, in 2023 a new dataset called BigSolDB was released, which compiled data from nearly 800 published papers, including information on solubility for about 800 molecules dissolved about more than 100 organic solvents that are commonly used in synthetic chemistry.

Attia and Burns decided to try training two different types of models on this data. Both of these models represent the chemical structures of molecules using numerical representations known as embeddings, which incorporate information such as the number of atoms in a molecule and which atoms are bound to which other atoms. Models can then use these representations to predict a variety of chemical properties.

One of the models used in this study, known as FastProp and developed by Burns and others in Green’s lab, incorporates “static embeddings.” This means that the model already knows the embedding for each molecule before it starts doing any kind of analysis.

The other model, ChemProp, learns an embedding for each molecule during the training, at the same time that it learns to associate the features of the embedding with a trait such as solubility. This model, developed across multiple MIT labs, has already been used for tasks such as antibiotic discovery, lipid nanoparticle design, and predicting chemical reaction rates.

The researchers trained both types of models on over 40,000 data points from BigSolDB, including information on the effects of temperature, which plays a significant role in solubility. Then, they tested the models on about 1,000 solutes that had been withheld from the training data. They found that the models’ predictions were two to three times more accurate than those of SolProp, the previous best model, and the new models were especially accurate at predicting variations in solubility due to temperature.

“Being able to accurately reproduce those small variations in solubility due to temperature, even when the overarching experimental noise is very large, was a really positive sign that the network had correctly learned an underlying solubility prediction function,” Burns says.

Accurate predictions

The researchers had expected that the model based on ChemProp, which is able to learn new representations as it goes along, would be able to make more accurate predictions. However, to their surprise, they found that the two models performed essentially the same. That suggests that the main limitation on their performance is the quality of the data, and that the models are performing as well as theoretically possible based on the data that they’re using, the researchers say.

“ChemProp should always outperform any static embedding when you have sufficient data,” Burns says. “We were blown away to see that the static and learned embeddings were statistically indistinguishable in performance across all the different subsets, which indicates to us that that the data limitations that are present in this space dominated the model performance.”

The models could become more accurate, the researchers say, if better training and testing data were available — ideally, data obtained by one person or a group of people all trained to perform the experiments the same way.

“One of the big limitations of using these kinds of compiled datasets is that different labs use different methods and experimental conditions when they perform solubility tests. That contributes to this variability between different datasets,” Attia says.

Because the model based on FastProp makes its predictions faster and has code that is easier for other users to adapt, the researchers decided to make that one, known as FastSolv, available to the public. Multiple pharmaceutical companies have already begun using it.

“There are applications throughout the drug discovery pipeline,” Burns says. “We’re also excited to see, outside of formulation and drug discovery, where people may use this model.”

The research was funded, in part, by the U.S. Department of Energy.

Within the past few years, models that can predict the structure or function of proteins have been widely used for a variety of biological applications, such as identifying drug targets and designing new therapeutic antibodies.

These models, which are based on large language models (LLMs), can make very accurate predictions of a protein’s suitability for a given application. However, there’s no way to determine how these models make their predictions or which protein features play the most important role in those decisions.

In a new study, MIT researchers have used a novel technique to open up that “black box” and allow them to determine what features a protein language model takes into account when making predictions. Understanding what is happening inside that black box could help researchers to choose better models for a particular task, helping to streamline the process of identifying new drugs or vaccine targets.

“Our work has broad implications for enhanced explainability in downstream tasks that rely on these representations,” says Bonnie Berger, the Simons Professor of Mathematics, head of the Computation and Biology group in MIT’s Computer Science and Artificial Intelligence Laboratory, and the senior author of the study. “Additionally, identifying features that protein language models track has the potential to reveal novel biological insights from these representations.”

Onkar Gujral, an MIT graduate student, is the lead author of the open-access study, which appears this week in the Proceedings of the National Academy of Sciences. Mihir Bafna, an MIT graduate student in electrical engineering and computer science, and Eric Alm, an MIT professor of biological engineering, are also authors of the paper.

Opening the black box

In 2018, Berger and former MIT graduate student Tristan Bepler PhD ’20 introduced the first protein language model. Their model, like subsequent protein models that accelerated the development of AlphaFold, such as ESM2 and OmegaFold, was based on LLMs. These models, which include ChatGPT, can analyze huge amounts of text and figure out which words are most likely to appear together.

Protein language models use a similar approach, but instead of analyzing words, they analyze amino acid sequences. Researchers have used these models to predict the structure and function of proteins, and for applications such as identifying proteins that might bind to particular drugs.

In a 2021 study, Berger and colleagues used a protein language model to predict which sections of viral surface proteins are less likely to mutate in a way that enables viral escape. This allowed them to identify possible targets for vaccines against influenza, HIV, and SARS-CoV-2.

However, in all of these studies, it has been impossible to know how the models were making their predictions.

“We would get out some prediction at the end, but we had absolutely no idea what was happening in the individual components of this black box,” Berger says.

In the new study, the researchers wanted to dig into how protein language models make their predictions. Just like LLMs, protein language models encode information as representations that consist of a pattern of activation of different “nodes” within a neural network. These nodes are analogous to the networks of neurons that store memories and other information within the brain.

The inner workings of LLMs are not easy to interpret, but within the past couple of years, researchers have begun using a type of algorithm known as a sparse autoencoder to help shed some light on how those models make their predictions. The new study from Berger’s lab is the first to use this algorithm on protein language models.

Sparse autoencoders work by adjusting how a protein is represented within a neural network. Typically, a given protein will be represented by a pattern of activation of a constrained number of neurons, for example, 480. A sparse autoencoder will expand that representation into a much larger number of nodes, say 20,000.

When information about a protein is encoded by only 480 neurons, each node lights up for multiple features, making it very difficult to know what features each node is encoding. However, when the neural network is expanded to 20,000 nodes, this extra space along with a sparsity constraint gives the information room to “spread out.” Now, a feature of the protein that was previously encoded by multiple nodes can occupy a single node.

“In a sparse representation, the neurons lighting up are doing so in a more meaningful manner,” Gujral says. “Before the sparse representations are created, the networks pack information so tightly together that it's hard to interpret the neurons.”

Interpretable models

Once the researchers obtained sparse representations of many proteins, they used an AI assistant called Claude (related to the popular Anthropic chatbot of the same name), to analyze the representations. In this case, they asked Claude to compare the sparse representations with the known features of each protein, such as molecular function, protein family, or location within a cell.

By analyzing thousands of representations, Claude can determine which nodes correspond to specific protein features, then describe them in plain English. For example, the algorithm might say, “This neuron appears to be detecting proteins involved in transmembrane transport of ions or amino acids, particularly those located in the plasma membrane.”

This process makes the nodes far more “interpretable,” meaning the researchers can tell what each node is encoding. They found that the features most likely to be encoded by these nodes were protein family and certain functions, including several different metabolic and biosynthetic processes.

“When you train a sparse autoencoder, you aren’t training it to be interpretable, but it turns out that by incentivizing the representation to be really sparse, that ends up resulting in interpretability,” Gujral says.

Understanding what features a particular protein model is encoding could help researchers choose the right model for a particular task, or tweak the type of input they give the model, to generate the best results. Additionally, analyzing the features that a model encodes could one day help biologists to learn more about the proteins that they are studying.

“At some point when the models get a lot more powerful, you could learn more biology than you already know, from opening up the models,” Gujral says.

The research was funded by the National Institutes of Health. 

MIT researchers have developed a reconfigurable antenna that dynamically adjusts its frequency range by changing its physical shape, making it more versatile for communications and sensing than static antennas.

A user can stretch, bend, or compress the antenna to make reversible changes to its radiation properties, enabling a device to operate in a wider frequency range without the need for complex, moving parts. With an adjustable frequency range, a reconfigurable antenna could adapt to changing environmental conditions and reduce the need for multiple antennas.

The word “antenna” may draw to mind metal rods like the “bunny ears” on top of old television sets, but the MIT team instead worked with metamaterials — engineered materials whose mechanical properties, such as stiffness and strength, depend on the geometric arrangement of the material’s components.

The result is a simplified design for a reconfigurable antenna that could be used for applications like energy transfer in wearable devices, motion tracking and sensing for augmented reality, or wireless communication across a wide range of network protocols.

In addition, the researchers developed an editing tool so users can generate customized metamaterial antennas, which can be fabricated using a laser cutter.

“Usually, when we think of antennas, we think of static antennas — they are fabricated to have specific properties and that is it. However, by using auxetic metamaterials, which can deform into three different geometric states, we can seamlessly change the properties of the antenna by changing its geometry, without fabricating a new structure. In addition, we can use changes in the antenna’s radio frequency properties, due to changes in the metamaterial geometry, as a new method of sensing for interaction design,” says lead author Marwa AlAlawi, a mechanical engineering graduate student at MIT.

Her co-authors include Regina Zheng and Katherine Yan, both MIT undergraduate students; Ticha Sethapakdi, an MIT graduate student in electrical engineering and computer science; Soo Yeon Ahn of the Gwangju Institute of Science and Technology in Korea; and co-senior authors Junyi Zhu, assistant professor at the University of Michigan; and Stefanie Mueller, the TIBCO Career Development Associate Professor in MIT’s departments of Electrical Engineering and Computer Science and Mechanical Engineering and leader of the Human-Computer Interaction Group at the Computer Science and Artificial Intelligence Lab. The research will be presented at the ACM Symposium on User Interface Software and Technology.

Making sense of antennas

While traditional antennas radiate and receive radio signals, in this work, the researchers looked at how the devices can act as sensors. The team’s goal was to develop a mechanical element that can also be used as an antenna for sensing.

To do this, they leveraged the antenna’s “resonance frequency,” which is the frequency at which the antenna is most efficient.

An antenna’s resonance frequency will shift due to changes in its shape. (Think about extending the left “bunny ear” to reduce TV static.) Researchers can capture these shifts for sensing. For instance, a reconfigurable antenna could be used in this way to detect the expansion of a person’s chest, to monitor their respiration.

To design a versatile reconfigurable antenna, the researchers used metamaterials. These engineered materials, which can be programmed to adopt different shapes, are composed of a periodic arrangement of unit cells that can be rotated, compressed, stretched, or bent.

By deforming the metamaterial structure, one can shift the antenna’s resonance frequency.

“In order to trigger changes in resonance frequency, we either need to change the antenna’s effective length or introduce slits and holes into it. Metamaterials allow us to get those different states from only one structure,” AlAlawi says.

The device, dubbed the meta-antenna, is composed of a dielectric layer of material sandwiched between two conductive layers.

To fabricate a meta-antenna, the researchers cut the dielectric laser out of a rubber sheet with a laser cutter. Then they added a patch on top of the dielectric layer using conductive spray paint, creating a resonating “patch antenna.”

But they found that even the most flexible conductive material couldn’t withstand the amount of deformation the antenna would experience.

“We did a lot of trial and error to determine that, if we coat the structure with flexible acrylic paint, it protects the hinges so they don’t break prematurely,” AlAlawi explains.

A means for makers

With the fabrication problem solved, the researchers built a tool that enables users to design and produce metamaterial antennas for specific applications.

The user can define the size of the antenna patch, choose a thickness for the dielectric layer, and set the length to width ratio of the metamaterial unit cells. Then the system automatically simulates the antenna’s resonance frequency range.

“The beauty of metamaterials is that, because it is an interconnected system of linkages, the geometric structure allows us to reduce the complexity of a mechanical system,” AlAlawi says.

Using the design tool, the researchers incorporated meta-antennas into several smart devices, including a curtain that dynamically adjusts household lighting and headphones that seamlessly transition between noise-cancelling and transparent modes.

For the smart headphone, for instance, when the meta-antenna expands and bends, it shifts the resonance frequency by 2.6 percent, which switches the headphone mode. The team’s experiments also showed that meta-antenna structures are durable enough to withstand more than 10,000 compressions.

Because the antenna patch can be patterned onto any surface, it could be used with more complex structures. For instance, the antenna could be incorporated into smart textiles that perform noninvasive biomedical sensing or temperature monitoring.

In the future, the researchers want to design three-dimensional meta-antennas for a wider range of applications. They also want to add more functions to the design tool, improve the durability and flexibility of the metamaterial structure, experiment with different symmetric metamaterial patterns, and streamline some manual fabrication steps.

This research was funded, in part, by the Bahrain Crown Prince International Scholarship and the Gwangju Institute of Science and Technology.

Using artificial intelligence, MIT researchers have come up with a new way to design nanoparticles that can more efficiently deliver RNA vaccines and other types of RNA therapies.

After training a machine-learning model to analyze thousands of existing delivery particles, the researchers used it to predict new materials that would work even better. The model also enabled the researchers to identify particles that would work well in different types of cells, and to discover ways to incorporate new types of materials into the particles.

“What we did was apply machine-learning tools to help accelerate the identification of optimal ingredient mixtures in lipid nanoparticles to help target a different cell type or help incorporate different materials, much faster than previously was possible,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, and the senior author of the study.

This approach could dramatically speed the process of developing new RNA vaccines, as well as therapies that could be used to treat obesity, diabetes, and other metabolic disorders, the researchers say.

Alvin Chan, a former MIT postdoc who is now an assistant professor at Nanyang Technological University, and Ameya Kirtane, a former MIT postdoc who is now an assistant professor at the University of Minnesota, are the lead authors of the new open-access study, which appears today in Nature Nanotechnology.

Particle predictions

RNA vaccines, such as the vaccines for SARS-CoV-2, are usually packaged in lipid nanoparticles (LNPs) for delivery. These particles protect mRNA from being broken down in the body and help it to enter cells once injected.

Creating particles that handle these jobs more efficiently could help researchers to develop even more effective vaccines. Better delivery vehicles could also make it easier to develop mRNA therapies that encode genes for proteins that could help to treat a variety of diseases.

In 2024, Traverso’s lab launched a multiyear research program, funded by the U.S. Advanced Research Projects Agency for Health (ARPA-H), to develop new ingestible devices that could achieve oral delivery of RNA treatments and vaccines.

“Part of what we’re trying to do is develop ways of producing more protein, for example, for therapeutic applications. Maximizing the efficiency is important to be able to boost how much we can have the cells produce,” Traverso says.

A typical LNP consists of four components — a cholesterol, a helper lipid, an ionizable lipid, and a lipid that is attached to polyethylene glycol (PEG). Different variants of each of these components can be swapped in to create a huge number of possible combinations. Changing up these formulations and testing each one individually is very time-consuming, so Traverso, Chan, and their colleagues decided to turn to artificial intelligence to help speed up the process.

“Most AI models in drug discovery focus on optimizing a single compound at a time, but that approach doesn’t work for lipid nanoparticles, which are made of multiple interacting components,” Chan says. “To tackle this, we developed a new model called COMET, inspired by the same transformer architecture that powers large language models like ChatGPT. Just as those models understand how words combine to form meaning, COMET learns how different chemical components come together in a nanoparticle to influence its properties — like how well it can deliver RNA into cells.”

To generate training data for their machine-learning model, the researchers created a library of about 3,000 different LNP formulations. The team tested each of these 3,000 particles in the lab to see how efficiently they could deliver their payload to cells, then fed all of this data into a machine-learning model.

After the model was trained, the researchers asked it to predict new formulations that would work better than existing LNPs. They tested those predictions by using the new formulations to deliver mRNA encoding a fluorescent protein to mouse skin cells grown in a lab dish. They found that the LNPs predicted by the model did indeed work better than the particles in the training data, and in some cases better than LNP formulations that are used commercially.

Accelerated development

Once the researchers showed that the model could accurately predict particles that would efficiently deliver mRNA, they began asking additional questions. First, they wondered if they could train the model on nanoparticles that incorporate a fifth component: a type of polymer known as branched poly beta amino esters (PBAEs).

Research by Traverso and his colleagues has shown that these polymers can effectively deliver nucleic acids on their own, so they wanted to explore whether adding them to LNPs could improve LNP performance. The MIT team created a set of about 300 LNPs that also include these polymers, which they used to train the model. The resulting model could then predict additional formulations with PBAEs that would work better.

Next, the researchers set out to train the model to make predictions about LNPs that would work best in different types of cells, including a type of cell called Caco-2, which is derived from colorectal cancer cells. Again, the model was able to predict LNPs that would efficiently deliver mRNA to these cells.

Lastly, the researchers used the model to predict which LNPs could best withstand lyophilization — a freeze-drying process often used to extend the shelf-life of medicines.

“This is a tool that allows us to adapt it to a whole different set of questions and help accelerate development. We did a large training set that went into the model, but then you can do much more focused experiments and get outputs that are helpful on very different kinds of questions,” Traverso says.

He and his colleagues are now working on incorporating some of these particles into potential treatments for diabetes and obesity, which are two of the primary targets of the ARPA-H funded project. Therapeutics that could be delivered using this approach include GLP-1 mimics with similar effects to Ozempic.

This research was funded by the GO Nano Marble Center at the Koch Institute, the Karl van Tassel Career Development Professorship, the MIT Department of Mechanical Engineering, Brigham and Women’s Hospital, and ARPA-H.

Graphite is a key structural component in some of the world’s oldest nuclear reactors and many of the next-generation designs being built today. But it also condenses and swells in response to radiation — and the mechanism behind those changes has proven difficult to study.

Now, MIT researchers and collaborators have uncovered a link between properties of graphite and how the material behaves in response to radiation. The findings could lead to more accurate, less destructive ways of predicting the lifespan of graphite materials used in reactors around the world.

“We did some basic science to understand what leads to swelling and, eventually, failure in graphite structures,” says MIT Research Scientist Boris Khaykovich, senior author of the new study. “More research will be needed to put this into practice, but the paper proposes an attractive idea for industry: that you might not need to break hundreds of irradiated samples to understand their failure point.”

Specifically, the study shows a connection between the size of the pores within graphite and the way the material swells and shrinks in volume, leading to degradation.

“The lifetime of nuclear graphite is limited by irradiation-induced swelling,” says co-author and MIT Research Scientist Lance Snead. “Porosity is a controlling factor in this swelling, and while graphite has been extensively studied for nuclear applications since the Manhattan Project, we still do not have a clear understanding of the porosity in both mechanical properties and swelling. This work addresses that.”

The open-access paper appears this week in Interdisciplinary Materials. It is co-authored by Khaykovich, Snead, MIT Research Scientist Sean Fayfar, former MIT research fellow Durgesh Rai, Stony Brook University Assistant Professor David Sprouster, Oak Ridge National Laboratory Staff Scientist Anne Campbell, and Argonne National Laboratory Physicist Jan Ilavsky.

A long-studied, complex material

Ever since 1942, when physicists and engineers built the world’s first nuclear reactor on a converted squash court at the University of Chicago, graphite has played a central role in the generation of nuclear energy. That first reactor, dubbed the Chicago Pile, was constructed from about 40,000 graphite blocks, many of which contained nuggets of uranium.

Today graphite is a vital component of many operating nuclear reactors and is expected to play a central role in next-generation reactor designs like molten-salt and high-temperature gas reactors. That’s because graphite is a good neutron moderator, slowing down the neutrons released by nuclear fission so they are more likely to create fissions themselves and sustain a chain reaction.

“The simplicity of graphite makes it valuable,” Khaykovich explains. “It’s made of carbon, and it’s relatively well-known how to make it cleanly. Graphite is a very mature technology. It’s simple, stable, and we know it works.”

But graphite also has its complexities.

“We call graphite a composite even though it’s made up of only carbon atoms,” Khaykovich says. “It includes ‘filler particles’ that are more crystalline, then there is a matrix called a ‘binder’ that is less crystalline, then there are pores that span in length from nanometers to many microns.”

Each graphite grade has its own composite structure, but they all contain fractals, or shapes that look the same at different scales.

Those complexities have made it hard to predict how graphite will respond to radiation in microscopic detail, although it’s been known for decades that when graphite is irradiated, it first densifies, reducing its volume by up to 10 percent, before swelling and cracking. The volume fluctuation is caused by changes to graphite’s porosity and lattice stress.

“Graphite deteriorates under radiation, as any material does,” Khaykovich says. “So, on the one hand we have a material that’s extremely well-known, and on the other hand, we have a material that is immensely complicated, with a behavior that’s impossible to predict through computer simulations.”

For the study, the researchers received irradiated graphite samples from Oak Ridge National Laboratory. Co-authors Campbell and Snead were involved in irradiating the samples some 20 years ago. The samples are a grade of graphite known as G347A.

The research team used an analysis technique known as X-ray scattering, which uses the scattered intensity of an X-ray beam to analyze the properties of material. Specifically, they looked at the distribution of sizes and surface areas of the sample’s pores, or what are known as the material’s fractal dimensions.

“When you look at the scattering intensity, you see a large range of porosity,” Fayfar says. “Graphite has porosity over such large scales, and you have this fractal self-similarity: The pores in very small sizes look similar to pores spanning microns, so we used fractal models to relate different morphologies across length scales.”

Fractal models had been used on graphite samples before, but not on irradiated samples to see how the material’s pore structures changed. The researchers found that when graphite is first exposed to radiation, its pores get filled as the material degrades.

“But what was quite surprising to us is the [size distribution of the pores] turned back around,” Fayfar says. “We had this recovery process that matched our overall volume plots, which was quite odd. It seems like after graphite is irradiated for so long, it starts recovering. It’s sort of an annealing process where you create some new pores, then the pores smooth out and get slightly bigger. That was a big surprise.”

The researchers found that the size distribution of the pores closely follows the volume change caused by radiation damage.

“Finding a strong correlation between the [size distribution of pores] and the graphite’s volume changes is a new finding, and it helps connect to the failure of the material under irradiation,” Khaykovich says. “It’s important for people to know how graphite parts will fail when they are under stress and how failure probability changes under irradiation.”

From research to reactors

The researchers plan to study other graphite grades and explore further how pore sizes in irradiated graphite correlate with the probability of failure. They speculate that a statistical technique known as the Weibull Distribution could be used to predict graphite’s time until failure. The Weibull Distribution is already used to describe the probability of failure in ceramics and other porous materials like metal alloys.

Khaykovich also speculated that the findings could contribute to our understanding of why materials densify and swell under irradiation.

“There’s no quantitative model of densification that takes into account what’s happening at these tiny scales in graphite,” Khaykovich says. “Graphite irradiation densification reminds me of sand or sugar, where when you crush big pieces into smaller grains, they densify. For nuclear graphite, the crushing force is the energy that neutrons bring in, causing large pores to get filled with smaller, crushed pieces. But more energy and agitation create still more pores, and so graphite swells again. It’s not a perfect analogy, but I believe analogies bring progress for understanding these materials.”

The researchers describe the paper as an important step toward informing graphite production and use in nuclear reactors of the future.

“Graphite has been studied for a very long time, and we’ve developed a lot of strong intuitions about how it will respond in different environments, but when you’re building a nuclear reactor, details matter,” Khaykovich says. “People want numbers. They need to know how much thermal conductivity will change, how much cracking and volume change will happen. If components are changing volume, at some point you need to take that into account.”

This work was supported, in part, by the U.S. Department of Energy.

With help from artificial intelligence, MIT researchers have designed novel antibiotics that can combat two hard-to-treat infections: drug-resistant Neisseria gonorrhoeae and multi-drug-resistant Staphylococcus aureus (MRSA).

Using generative AI algorithms, the research team designed more than 36 million possible compounds and computationally screened them for antimicrobial properties. The top candidates they discovered are structurally distinct from any existing antibiotics, and they appear to work by novel mechanisms that disrupt bacterial cell membranes.

This approach allowed the researchers to generate and evaluate theoretical compounds that have never been seen before — a strategy that they now hope to apply to identify and design compounds with activity against other species of bacteria.

“We’re excited about the new possibilities that this project opens up for antibiotics development. Our work shows the power of AI from a drug design standpoint, and enables us to exploit much larger chemical spaces that were previously inaccessible,” says James Collins, the Termeer Professor of Medical Engineering and Science in MIT’s Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering, and a member of the Broad Institute.

Collins is the senior author of the study, which appears today in Cell. The paper’s lead authors are MIT postdoc Aarti Krishnan, former postdoc Melis Anahtar ’08, and Jacqueline Valeri PhD ’23.

Exploring chemical space

Over the past 45 years, a few dozen new antibiotics have been approved by the FDA, but most of these are variants of existing antibiotics. At the same time, bacterial resistance to many of these drugs has been growing. Globally, it is estimated that drug-resistant bacterial infections cause nearly 5 million deaths per year.

In hopes of finding new antibiotics to fight this growing problem, Collins and others at MIT’s Antibiotics-AI Project have harnessed the power of AI to screen huge libraries of existing chemical compounds. This work has yielded several promising drug candidates, including halicin and abaucin.

To build on that progress, Collins and his colleagues decided to expand their search into molecules that can’t be found in any chemical libraries. By using AI to generate hypothetically possible molecules that don’t exist or haven’t been discovered, they realized that it should be possible to explore a much greater diversity of potential drug compounds.

In their new study, the researchers employed two different approaches: First, they directed generative AI algorithms to design molecules based on a specific chemical fragment that showed antimicrobial activity, and second, they let the algorithms freely generate molecules, without having to include a specific fragment.

For the fragment-based approach, the researchers sought to identify molecules that could kill N. gonorrhoeae, a Gram-negative bacterium that causes gonorrhea. They began by assembling a library of about 45 million known chemical fragments, consisting of all possible combinations of 11 atoms of carbon, nitrogen, oxygen, fluorine, chlorine, and sulfur, along with fragments from Enamine’s REadily AccessibLe (REAL) space.

Then, they screened the library using machine-learning models that Collins’ lab has previously trained to predict antibacterial activity against N. gonorrhoeae. This resulted in nearly 4 million fragments. They narrowed down that pool by removing any fragments predicted to be cytotoxic to human cells, displayed chemical liabilities, and were known to be similar to existing antibiotics. This left them with about 1 million candidates.

“We wanted to get rid of anything that would look like an existing antibiotic, to help address the antimicrobial resistance crisis in a fundamentally different way. By venturing into underexplored areas of chemical space, our goal was to uncover novel mechanisms of action,” Krishnan says.

Through several rounds of additional experiments and computational analysis, the researchers identified a fragment they called F1 that appeared to have promising activity against N. gonorrhoeae. They used this fragment as the basis for generating additional compounds, using two different generative AI algorithms.

One of those algorithms, known as chemically reasonable mutations (CReM), works by starting with a particular molecule containing F1 and then generating new molecules by adding, replacing, or deleting atoms and chemical groups. The second algorithm, F-VAE (fragment-based variational autoencoder), takes a chemical fragment and builds it into a complete molecule. It does so by learning patterns of how fragments are commonly modified, based on its pretraining on more than 1 million molecules from the ChEMBL database.

Those two algorithms generated about 7 million candidates containing F1, which the researchers then computationally screened for activity against N. gonorrhoeae. This screen yielded about 1,000 compounds, and the researchers selected 80 of those to see if they could be produced by chemical synthesis vendors. Only two of these could be synthesized, and one of them, named NG1, was very effective at killing N. gonorrhoeae in a lab dish and in a mouse model of drug-resistant gonorrhea infection.

Additional experiments revealed that NG1 interacts with a protein called LptA, a novel drug target involved in the synthesis of the bacterial outer membrane. It appears that the drug works by interfering with membrane synthesis, which is fatal to cells.

Unconstrained design

In a second round of studies, the researchers explored the potential of using generative AI to freely design molecules, using Gram-positive bacteria, S. aureus as their target.

Again, the researchers used CReM and VAE to generate molecules, but this time with no constraints other than the general rules of how atoms can join to form chemically plausible molecules. Together, the models generated more than 29 million compounds. The researchers then applied the same filters that they did to the N. gonorrhoeae candidates, but focusing on S. aureus, eventually narrowing the pool down to about 90 compounds.

They were able to synthesize and test 22 of these molecules, and six of them showed strong antibacterial activity against multi-drug-resistant S. aureus grown in a lab dish. They also found that the top candidate, named DN1, was able to clear a methicillin-resistant S. aureus (MRSA) skin infection in a mouse model. These molecules also appear to interfere with bacterial cell membranes, but with broader effects not limited to interaction with one specific protein.

Phare Bio, a nonprofit that is also part of the Antibiotics-AI Project, is now working on further modifying NG1 and DN1 to make them suitable for additional testing.

“In a collaboration with Phare Bio, we are exploring analogs, as well as working on advancing the best candidates preclinically, through medicinal chemistry work,” Collins says. “We are also excited about applying the platforms that Aarti and the team have developed toward other bacterial pathogens of interest, notably Mycobacterium tuberculosis and Pseudomonas aeruginosa.”

The research was funded, in part, by the U.S. Defense Threat Reduction Agency, the National Institutes of Health, the Audacious Project, Flu Lab, the Sea Grape Foundation, Rosamund Zander and Hansjorg Wyss for the Wyss Foundation, and an anonymous donor.

Is this movie review a rave or a pan? Is this news story about business or technology? Is this online chatbot conversation veering off into giving financial advice? Is this online medical information site giving out misinformation?

These kinds of automated conversations, whether they involve seeking a movie or restaurant review or getting information about your bank account or health records, are becoming increasingly prevalent. More than ever, such evaluations are being made by highly sophisticated algorithms, known as text classifiers, rather than by human beings. But how can we tell how accurate these classifications really are?

Now, a team at MIT’s Laboratory for Information and Decision Systems (LIDS) has come up with an innovative approach to not only measure how well these classifiers are doing their job, but then go one step further and show how to make them more accurate.

The new evaluation and remediation software was led and developed by Lei Xu alongside the research conducted by Sarah Alnegheimish, Kalyan Veeramachaneni, a principal research scientist at LIDS and senior author, with two others. The software package is being made freely available for download by anyone who wants to use it.

A standard method for testing these classification systems is to create what are known as synthetic examples — sentences that closely resemble ones that have already been classified. For example, researchers might take a sentence that has already been tagged by a classifier program as being a rave review, and see if changing a word or a few words while retaining the same meaning could fool the classifier into deeming it a pan. Or a sentence that was determined to be misinformation might get misclassified as accurate. This ability to fool the classifiers makes these adversarial examples.

People have tried various ways to find the vulnerabilities in these classifiers, Veeramachaneni says. But existing methods of finding these vulnerabilities have a hard time with this task and miss many examples that they should catch, he says.

Increasingly, companies are trying to use such evaluation tools in real time, monitoring the output of chatbots used for various purposes to try to make sure they are not putting out improper responses. For example, a bank might use a chatbot to respond to routine customer queries such as checking account balances or applying for a credit card, but it wants to ensure that its responses could never be interpreted as financial advice, which could expose the company to liability. “Before showing the chatbot’s response to the end user, they want to use the text classifier to detect whether it’s giving financial advice or not,” Veeramachaneni says. But then it’s important to test that classifier to see how reliable its evaluations are.

“These chatbots, or summarization engines or whatnot are being set up across the board,” he says, to deal with external customers and within an organization as well, for example providing information about HR issues. It’s important to put these text classifiers into the loop to detect things that they are not supposed to say, and filter those out before the output gets transmitted to the user.

That’s where the use of adversarial examples comes in — those sentences that have already been classified but then produce a different response when they are slightly modified while retaining the same meaning. How can people confirm that the meaning is the same? By using another large language model (LLM) that interprets and compares meanings. So, if the LLM says the two sentences mean the same thing, but the classifier labels them differently, “that is a sentence that is adversarial — it can fool the classifier,” Veeramachaneni says. And when the researchers examined these adversarial sentences, “we found that most of the time, this was just a one-word change,” although the people using LLMs to generate these alternate sentences often didn’t realize that.

Further investigation, using LLMs to analyze many thousands of examples, showed that certain specific words had an outsized influence in changing the classifications, and therefore the testing of a classifier’s accuracy could focus on this small subset of words that seem to make the most difference. They found that one-tenth of 1 percent of all the 30,000 words in the system’s vocabulary could account for almost half of all these reversals of classification, in some specific applications.

Lei Xu PhD ’23, a recent graduate from LIDS who performed much of the analysis as part of his thesis work, “used a lot of interesting estimation techniques to figure out what are the most powerful words that can change the overall classification, that can fool the classifier,” Veeramachaneni says. The goal is to make it possible to do much more narrowly targeted searches, rather than combing through all possible word substitutions, thus making the computational task of generating adversarial examples much more manageable. “He’s using large language models, interestingly enough, as a way to understand the power of a single word.”

Then, also using LLMs, he searches for other words that are closely related to these powerful words, and so on, allowing for an overall ranking of words according to their influence on the outcomes. Once these adversarial sentences have been found, they can be used in turn to retrain the classifier to take them into account, increasing the robustness of the classifier against those mistakes.

Making classifiers more accurate may not sound like a big deal if it’s just a matter of classifying news articles into categories, or deciding whether reviews of anything from movies to restaurants are positive or negative. But increasingly, classifiers are being used in settings where the outcomes really do matter, whether preventing the inadvertent release of sensitive medical, financial, or security information, or helping to guide important research, such as into properties of chemical compounds or the folding of proteins for biomedical applications, or in identifying and blocking hate speech or known misinformation.

As a result of this research, the team introduced a new metric, which they call p, which provides a measure of how robust a given classifier is against single-word attacks. And because of the importance of such misclassifications, the research team has made its products available as open access for anyone to use. The package consists of two components: SP-Attack, which generates adversarial sentences to test classifiers in any particular application, and SP-Defense, which aims to improve the robustness of the classifier by generating and using adversarial sentences to retrain the model.

In some tests, where competing methods of testing classifier outputs allowed a 66 percent success rate by adversarial attacks, this team’s system cut that attack success rate almost in half, to 33.7 percent. In other applications, the improvement was as little as a 2 percent difference, but even that can be quite important, Veeramachaneni says, since these systems are being used for so many billions of interactions that even a small percentage can affect millions of transactions.

The team’s results were published on July 7 in the journal Expert Systems in a paper by Xu, Veeramachaneni, and Alnegheimish of LIDS, along with Laure Berti-Equille at IRD in Marseille, France, and Alfredo Cuesta-Infante at the Universidad Rey Juan Carlos, in Spain. 

Around the world, about 2 billion people suffer from iron deficiency, which can lead to anemia, impaired brain development in children, and increased infant mortality.

To combat that problem, MIT researchers have come up with a new way to fortify foods and beverages with iron, using small crystalline particles. These particles, known as metal-organic frameworks, could be sprinkled on food, added to staple foods such as bread, or incorporated into drinks like coffee and tea.

“We’re creating a solution that can be seamlessly added to staple foods across different regions,” says Ana Jaklenec, a principal investigator at MIT’s Koch Institute for Integrative Cancer Research. “What’s considered a staple in Senegal isn’t the same as in India or the U.S., so our goal was to develop something that doesn’t react with the food itself. That way, we don’t have to reformulate for every context — it can be incorporated into a wide range of foods and beverages without compromise.”

The particles designed in this study can also carry iodine, another critical nutrient. The particles could also be adapted to carry important minerals such as zinc, calcium, or magnesium.

“We are very excited about this new approach and what we believe is a novel application of metal-organic frameworks to potentially advance nutrition, particularly in the developing world,” says Robert Langer, the David H. Koch Institute Professor at MIT and a member of the Koch Institute.

Jaklenec and Langer are the senior authors of the study, which appears today in the journal Matter. MIT postdoc Xin Yang and Linzixuan (Rhoda) Zhang PhD ’24 are the lead authors of the paper.

Iron stabilization

Food fortification can be a successful way to combat nutrient deficiencies, but this approach is often challenging because many nutrients are fragile and break down during storage or cooking. When iron is added to foods, it can react with other molecules in the food, giving the food a metallic taste.

In previous work, Jaklenec’s lab has shown that encapsulating nutrients in polymers can protect them from breaking down or reacting with other molecules. In a small clinical trial, the researchers found that women who ate bread fortified with encapsulated iron were able to absorb the iron from the food.

However, one drawback to this approach is that the polymer adds a lot of bulk to the material, limiting the amount of iron or other nutrients that end up in the food.

“Encapsulating iron in polymers significantly improves its stability and reactivity, making it easier to add to food,” Jaklenec says. “But to be effective, it requires a substantial amount of polymer. That limits how much iron you can deliver in a typical serving, making it difficult to meet daily nutritional targets through fortified foods alone.”

To overcome that challenge, Yang came up with a new idea: Instead of encapsulating iron in a polymer, they could use iron itself as a building block for a crystalline particle known as a metal-organic framework, or MOF (pronounced “moff”).

MOFs consist of metal atoms joined by organic molecules called ligands to create a rigid, cage-like structure. Depending on the combination of metals and ligands chosen, they can be used for a wide variety of applications.

“We thought maybe we could synthesize a metal-organic framework with food-grade ligands and food-grade micronutrients,” Yang says. “Metal-organic frameworks have very high porosity, so they can load a lot of cargo. That’s why we thought we could leverage this platform to make a new metal-organic framework that could be used in the food industry.”

In this case, the researchers designed a MOF consisting of iron bound to a ligand called fumaric acid, which is often used as a food additive to enhance flavor or help preserve food.

This structure prevents iron from reacting with polyphenols — compounds commonly found in foods such as whole grains and nuts, as well as coffee and tea. When iron does react with those compounds, it forms a metal polyphenol complex that cannot be absorbed by the body.

The MOFs’ structure also allows them to remain stable until they reach an acidic environment, such as the stomach, where they break down and release their iron payload.

Double-fortified salts

The researchers also decided to include iodine in their MOF particle, which they call NuMOF. Iodized salt has been very successful at preventing iodine deficiency, and many efforts are now underway to create “double-fortified salts” that would also contain iron.

Delivering these nutrients together has proven difficult because iron and iodine can react with each other, making each one less likely to be absorbed by the body. In this study, the MIT team showed that once they formed their iron-containing MOF particles, they could load them with iodine, in a way that the iron and iodine do not react with each other.

In tests of the particles’ stability, the researchers found that the NuMOFs could withstand long-term storage, high heat and humidity, and boiling water.

Throughout these tests, the particles maintained their structure. When the researchers then fed the particles to mice, they found that both iron and iodine became available in the bloodstream within several hours of the NuMOF consumption.

The researchers are now working on launching a company that is developing coffee and other beverages fortified with iron and iodine. They also hope to continue working toward a double-fortified salt that could be consumed on its own or incorporated into staple food products.

The research was partially supported by J-WAFS Fellowships for Water and Food Solutions. 

Other authors of the paper include Fangzheng Chen, Wenhao Gao, Zhiling Zheng, Tian Wang, Erika Yan Wang, Behnaz Eshaghi, and Sydney MacDonald.

This research was conducted, in part, using MIT.nano’s facilities.

Water is essential for life on Earth. So, the liquid must be a requirement for life on other worlds. For decades, scientists’ definition of habitability on other planets has rested on this assumption.

But what makes some planets habitable might have very little to do with water. In fact, an entirely different type of liquid could conceivably support life in worlds where water can barely exist. That’s a possibility that MIT scientists raise in a study appearing this week in the Proceedings of the National Academy of Sciences.

From lab experiments, the researchers found that a type of fluid known as an ionic liquid can readily form from chemical ingredients that are also expected to be found on the surface of some rocky planets and moons. Ionic liquids are salts that exist in liquid form below about 100 degrees Celsius. The team’s experiments showed that a mixture of sulfuric acid and certain nitrogen-containing organic compounds produced such a liquid. On rocky planets, sulfuric acid may be a byproduct of volcanic activity, while nitrogen-containing compounds have been detected on several asteroids and planets in our solar system, suggesting the compounds may be present in other planetary systems.

The scientists propose that, even on planets that are too warm or that have atmospheres are too low-pressure to support liquid water, there could still be pockets of ionic liquid. And where there is liquid, there may be potential for life, though likely not anything that resembles Earth’s water-based beings.

Ionic liquids have extremely low vapor pressure and do not evaporate; they can form and persist at higher temperatures and lower pressures than what liquid water can tolerate. The researchers note that ionic liquid can be a hospitable environment for some biomolecules, such as certain proteins that can remain stable in the fluid.

“We consider water to be required for life because that is what’s needed for Earth life. But if we look at a more general definition, we see that what we need is a liquid in which metabolism for life can take place,” says Rachana Agrawal, who led the study as a postdoc in MIT’s Department of Earth, Atmospheric and Planetary Sciences. “Now if we include ionic liquid as a possibility, this can dramatically increase the habitability zone for all rocky worlds.”

The study’s MIT co-authors are Sara Seager, the Class of 1941 Professor of Planetary Sciences in the Department of Earth, Atmospheric and Planetary Sciences and a professor in the departments of Physics and of Aeronautics and Astronautics, along with Iaroslav Iakubivskyi, Weston Buchanan, Ana Glidden, and Jingcheng Huang. Co-authors also include Maxwell Seager of Worcester Polytechnic Institute, William Bains of Cardiff University, and Janusz Petkowski of Wroclaw University of Science and Technology, in Poland.

A liquid leap

The team’s work with ionic liquid grew out of an effort to search for signs of life on Venus, where clouds of sulfuric acid envelope the planet in a noxious haze. Despite its toxicity, Venus’ clouds may contain signs of life — a notion that scientists plan to test with upcoming missions to the planet’s atmosphere.

Agrawal and Seager, who is leading the Morning Star Missions to Venus, were investigating ways to collect and evaporate sulfuric acid. If a mission collects samples from Venus’ clouds, sulfuric acid would have to be evaporated away in order to reveal any residual organic compounds that could then be analyzed for signs of life.

The researchers were using their custom, low-pressure system designed to evaporate away excess sulfuric acid, to test evaporation of a solution of the acid and an organic compound, glycine. They found that in every case, while most of the liquid sulfuric acid evaporated, a stubborn layer of liquid always remained. They soon realized that sulfuric acid was chemically reacting with glycine, resulting in an exchange of hydrogen atoms from the acid to the organic compound. The result was a fluid mixture of salts, or ions, known as an ionic liquid, that persists as a liquid across a wide range of temperatures and pressures.

This accidental finding kickstarted an idea: Could ionic liquid form on planets that are too warm and host atmospheres too thin for water to exist?

“From there, we took the leap of imagination of what this could mean,” Agrawal says. “Sulfuric acid is found on Earth from volcanoes, and organic compounds have been found on asteroids and other planetary bodies. So, this led us to wonder if ionic liquids could potentially form and exist naturally on exoplanets.”

Rocky oases

On Earth, ionic liquids are mainly synthesized for industrial purposes. They do not occur naturally, except for in one specific case, in which the liquid is generated from the mixing of venoms produced by two rival species of ants.

The team set out to investigate what conditions ionic liquid could be naturally produced in, and over what range of temperatures and pressures. In the lab, they mixed sulfuric acid with various nitrogen-containing organic compounds. In previous work, Seager’s team had found that the compounds, some of which can be considered ingredients associated with life, are surprisingly stable in sulfuric acid.

“In high school, you learn that an acid wants to donate a proton,” Seager says. “And oddly enough, we knew from our past work with sulfuric acid (the main component of Venus’ clouds) and nitrogen-containing compounds, that a nitrogen wants to receive a hydrogen. It’s like one person’s trash is another person’s treasure.”

The reaction could produce a bit of ionic liquid if the sulfuric acid and nitrogen-containing organics were in a one-to-one ratio — a ratio that was not a focus of the prior work. For their new study, Seager and Agrawal mixed sulfuric acid with over 30 different nitrogen-containing organic compounds, across a range of temperatures and pressures, then observed whether ionic liquid formed when they evaporated away the sulfuric acid in various vials. They also mixed the ingredients onto basalt rocks, which are known to exist on the surface of many rocky planets.

The team found that the reactions produced ionic liquid at temperatures up to 180 degrees Celsius and at extremely low pressures — much lower than that of the Earth’s atmosphere. Their results suggest that ionic liquid could naturally form on other planets where liquid water cannot exist, under the right conditions.

“We were just astonished that the ionic liquid forms under so many different conditions,” Seager says. “If you put the sulfuric acid and the organic on a rock, the excess sulfuric acid seeps into the rock pores, but you’re still left with a drop of ionic liquid on the rock. Whatever we tried, ionic liquid still formed.”

“We’re envisioning a planet warmer than Earth, that doesn’t have water, and at some point in its past or currently, it has to have had sulfuric acid, formed from volcanic outgassing,” Seager says. “This sulfuric acid has to flow over a little pocket of organics. And organic deposits are extremely common in the solar system.”

Then, she says, the resulting pockets of liquid could stay on the planet’s surface, potentially for years or millenia, where they could theoretically serve as small oases for simple forms of ionic-liquid-based life. Going forward, Seager’s team plans to investigate further, to see what biomolecules, and ingredients for life, might survive, and thrive, in ionic liquid.

“We just opened up a Pandora’s box of new research,” Seager says. “It’s been a real journey.”

This research was supported, in part, by the Sloan Foundation and the Volkswagen Foundation.

The cost of solar panels has dropped by more than 99 percent since the 1970s, enabling widespread adoption of photovoltaic systems that convert sunlight into electricity.

A new MIT study drills down on specific innovations that enabled such dramatic cost reductions, revealing that technical advances across a web of diverse research efforts and industries played a pivotal role.

The findings could help renewable energy companies make more effective R&D investment decisions and aid policymakers in identifying areas to prioritize to spur growth in manufacturing and deployment.

The researchers’ modeling approach shows that key innovations often originated outside the solar sector, including advances in semiconductor fabrication, metallurgy, glass manufacturing, oil and gas drilling, construction processes, and even legal domains.

“Our results show just how intricate the process of cost improvement is, and how much scientific and engineering advances, often at a very basic level, are at the heart of these cost reductions. A lot of knowledge was drawn from different domains and industries, and this network of knowledge is what makes these technologies improve,” says study senior author Jessika Trancik, a professor in MIT’s Institute for Data, Systems, and Society.

Trancik is joined on the paper by co-lead authors Goksin Kavlak, a former IDSS graduate student and postdoc who is now a senior energy associate at the Brattle Group; Magdalena Klemun, a former IDSS graduate student and postdoc who is now an assistant professor at Johns Hopkins University; former MIT postdoc Ajinkya Kamat; as well as Brittany Smith and Robert Margolis of the National Renewable Energy Laboratory. The research appears today in PLOS ONE.

Identifying innovations

This work builds on mathematical models that the researchers previously developed that tease out the effects of engineering technologies on the cost of photovoltaic (PV) modules and systems.

In this study, the researchers aimed to dig even deeper into the scientific advances that drove those cost declines.

They combined their quantitative cost model with a detailed, qualitative analysis of innovations that affected the costs of PV system materials, manufacturing steps, and deployment processes.

“Our quantitative cost model guided the qualitative analysis, allowing us to look closely at innovations in areas that are hard to measure due to a lack of quantitative data,” Kavlak says.

Building on earlier work identifying key cost drivers — such as the number of solar cells per module, wiring efficiency, and silicon wafer area — the researchers conducted a structured scan of the literature for innovations likely to affect these drivers. Next, they grouped these innovations to identify patterns, revealing clusters that reduced costs by improving materials or prefabricating components to streamline manufacturing and installation. Finally, the team tracked industry origins and timing for each innovation, and consulted domain experts to zero in on the most significant innovations.

All told, they identified 81 unique innovations that affected PV system costs since 1970, from improvements in antireflective coated glass to the implementation of fully online permitting interfaces.

“With innovations, you can always go to a deeper level, down to things like raw materials processing techniques, so it was challenging to know when to stop. Having that quantitative model to ground our qualitative analysis really helped,” Trancik says.

They chose to separate PV module costs from so-called balance-of-system (BOS) costs, which cover things like mounting systems, inverters, and wiring.

PV modules, which are wired together to form solar panels, are mass-produced and can be exported, while many BOS components are designed, built, and sold at the local level.

“By examining innovations both at the BOS level and within the modules, we identify the different types of innovations that have emerged in these two parts of PV technology,” Kavlak says.

BOS costs depend more on soft technologies, nonphysical elements such as permitting procedures, which have contributed significantly less to PV’s past cost improvement compared to hardware innovations.

“Often, it comes down to delays. Time is money, and if you have delays on construction sites and unpredictable processes, that affects these balance-of-system costs,” Trancik says.

Innovations such as automated permitting software, which flags code-compliant systems for fast-track approval, show promise. Though not yet quantified in this study, the team’s framework could support future analysis of their economic impact and similar innovations that streamline deployment processes.

Interconnected industries

The researchers found that innovations from the semiconductor, electronics, metallurgy, and petroleum industries played a major role in reducing both PV and BOS costs, but BOS costs were also impacted by innovations in software engineering and electric utilities.

Noninnovation factors, like efficiency gains from bulk purchasing and the accumulation of knowledge in the solar power industry, also reduced some cost variables.

In addition, while most PV panel innovations originated in research organizations or industry, many BOS innovations were developed by city governments, U.S. states, or professional associations.

“I knew there was a lot going on with this technology, but the diversity of all these fields and how closely linked they are, and the fact that we can clearly see that network through this analysis, was interesting,” Trancik says.

“PV was very well-positioned to absorb innovations from other industries — thanks to the right timing, physical compatibility, and supportive policies to adapt innovations for PV applications,” Klemun adds.

The analysis also reveals the role greater computing power could play in reducing BOS costs through advances like automated engineering review systems and remote site assessment software.

“In terms of knowledge spillovers, what we've seen so far in PV may really just be the beginning,” Klemun says, pointing to the expanding role of robotics and AI-driven digital tools in driving future cost reductions and quality improvements.

In addition to their qualitative analysis, the researchers demonstrated how this methodology could be used to estimate the quantitative impact of a particular innovation if one has the numerical data to plug into the cost equation.

For instance, using information about material prices and manufacturing procedures, they estimate that wire sawing, a technique which was introduced in the 1980s, led to an overall PV system cost decrease of $5 per watt by reducing silicon losses and increasing throughput during fabrication.

“Through this retrospective analysis, you learn something valuable for future strategy because you can see what worked and what didn’t work, and the models can also be applied prospectively. It is also useful to know what adjacent sectors may help support improvement in a particular technology,” Trancik says.

Moving forward, the researchers plan to apply this methodology to a wide range of technologies, including other renewable energy systems. They also want to further study soft technology to identify innovations or processes that could accelerate cost reductions.

“Although the process of technological innovation may seem like a black box, we’ve shown that you can study it just like any other phenomena,” Trancik says.

This research is funded, in part, by the U.S. Department of Energy Solar Energy Technologies Office.

Any motorist who has ever waited through multiple cycles for a traffic light to turn green knows how annoying signalized intersections can be. But sitting at intersections isn’t just a drag on drivers’ patience — unproductive vehicle idling could contribute as much as 15 percent of the carbon dioxide emissions from U.S. land transportation.

A large-scale modeling study led by MIT researchers reveals that eco-driving measures, which can involve dynamically adjusting vehicle speeds to reduce stopping and excessive acceleration, could significantly reduce those CO₂ emissions.

Using a powerful artificial intelligence method called deep reinforcement learning, the researchers conducted an in-depth impact assessment of the factors affecting vehicle emissions in three major U.S. cities.

Their analysis indicates that fully adopting eco-driving measures could cut annual city-wide intersection carbon emissions by 11 to 22 percent, without slowing traffic throughput or affecting vehicle and traffic safety.

Even if only 10 percent of vehicles on the road employ eco-driving, it would result in 25 to 50 percent of the total reduction in CO2 emissions, the researchers found.

In addition, dynamically optimizing speed limits at about 20 percent of intersections provides 70 percent of the total emission benefits. This indicates that eco-driving measures could be implemented gradually while still having measurable, positive impacts on mitigating climate change and improving public health.

“Vehicle-based control strategies like eco-driving can move the needle on climate change reduction. We’ve shown here that modern machine-learning tools, like deep reinforcement learning, can accelerate the kinds of analysis that support sociotechnical decision making. This is just the tip of the iceberg,” says senior author Cathy Wu, the Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS) at MIT, and a member of the Laboratory for Information and Decision Systems (LIDS).

She is joined on the paper by lead author Vindula Jayawardana, an MIT graduate student; as well as MIT graduate students Ao Qu, Cameron Hickert, and Edgar Sanchez; MIT undergraduate Catherine Tang; Baptiste Freydt, a graduate student at ETH Zurich; and Mark Taylor and Blaine Leonard of the Utah Department of Transportation. The research appears in Transportation Research Part C: Emerging Technologies.

A multi-part modeling study

Traffic control measures typically call to mind fixed infrastructure, like stop signs and traffic signals. But as vehicles become more technologically advanced, it presents an opportunity for eco-driving, which is a catch-all term for vehicle-based traffic control measures like the use of dynamic speeds to reduce energy consumption.

In the near term, eco-driving could involve speed guidance in the form of vehicle dashboards or smartphone apps. In the longer term, eco-driving could involve intelligent speed commands that directly control the acceleration of semi-autonomous and fully autonomous vehicles through vehicle-to-infrastructure communication systems.

“Most prior work has focused on how to implement eco-driving. We shifted the frame to consider the question of should we implement eco-driving. If we were to deploy this technology at scale, would it make a difference?” Wu says.

To answer that question, the researchers embarked on a multifaceted modeling study that would take the better part of four years to complete.

They began by identifying 33 factors that influence vehicle emissions, including temperature, road grade, intersection topology, age of the vehicle, traffic demand, vehicle types, driver behavior, traffic signal timing, road geometry, etc.

“One of the biggest challenges was making sure we were diligent and didn’t leave out any major factors,” Wu says.

Then they used data from OpenStreetMap, U.S. geological surveys, and other sources to create digital replicas of more than 6,000 signalized intersections in three cities — Atlanta, San Francisco, and Los Angeles — and simulated more than a million traffic scenarios.

The researchers used deep reinforcement learning to optimize each scenario for eco-driving to achieve the maximum emissions benefits.

Reinforcement learning optimizes the vehicles’ driving behavior through trial-and-error interactions with a high-fidelity traffic simulator, rewarding vehicle behaviors that are more energy-efficient while penalizing those that are not.

The researchers cast the problem as a decentralized cooperative multi-agent control problem, where the vehicles cooperate to achieve overall energy efficiency, even among non-participating vehicles, and they act in a decentralized manner, avoiding the need for costly communication between vehicles.

However, training vehicle behaviors that generalize across diverse intersection traffic scenarios was a major challenge. The researchers observed that some scenarios are more similar to one another than others, such as scenarios with the same number of lanes or the same number of traffic signal phases.

As such, the researchers trained separate reinforcement learning models for different clusters of traffic scenarios, yielding better emission benefits overall.

But even with the help of AI, analyzing citywide traffic at the network level would be so computationally intensive it could take another decade to unravel, Wu says.

Instead, they broke the problem down and solved each eco-driving scenario at the individual intersection level.

“We carefully constrained the impact of eco-driving control at each intersection on neighboring intersections. In this way, we dramatically simplified the problem, which enabled us to perform this analysis at scale, without introducing unknown network effects,” she says.

Significant emissions benefits

When they analyzed the results, the researchers found that full adoption of eco-driving could result in intersection emissions reductions of between 11 and 22 percent.

These benefits differ depending on the layout of a city’s streets. A denser city like San Francisco has less room to implement eco-driving between intersections, offering a possible explanation for reduced emission savings, while Atlanta could see greater benefits given its higher speed limits.

Even if only 10 percent of vehicles employ eco-driving, a city could still realize 25 to 50 percent of the total emissions benefit because of car-following dynamics: Non-eco-driving vehicles would follow controlled eco-driving vehicles as they optimize speed to pass smoothly through intersections, reducing their carbon emissions as well.

In some cases, eco-driving could also increase vehicle throughput by minimizing emissions. However, Wu cautions that increasing throughput could result in more drivers taking to the roads, reducing emissions benefits.

And while their analysis of widely used safety metrics known as surrogate safety measures, such as time to collision, suggest that eco-driving is as safe as human driving, it could cause unexpected behavior in human drivers. More research is needed to fully understand potential safety impacts, Wu says.

Their results also show that eco-driving could provide even greater benefits when combined with alternative transportation decarbonization solutions. For instance, 20 percent eco-driving adoption in San Francisco would cut emission levels by 7 percent, but when combined with the projected adoption of hybrid and electric vehicles, it would cut emissions by 17 percent.

“This is a first attempt to systematically quantify network-wide environmental benefits of eco-driving. This is a great research effort that will serve as a key reference for others to build on in the assessment of eco-driving systems,” says Hesham Rakha, the Samuel L. Pritchard Professor of Engineering at Virginia Tech, who was not involved with this research.

And while the researchers focus on carbon emissions, the benefits are highly correlated with improvements in fuel consumption, energy use, and air quality.

“This is almost a free intervention. We already have smartphones in our cars, and we are rapidly adopting cars with more advanced automation features. For something to scale quickly in practice, it must be relatively simple to implement and shovel-ready. Eco-driving fits that bill,” Wu says.

This work is funded, in part, by Amazon and the Utah Department of Transportation.

A new strategy for strengthening polymer materials could lead to more durable plastics and cut down on plastic waste, according to researchers at MIT and Duke University.

Using machine learning, the researchers identified crosslinker molecules that can be added to polymer materials, allowing them to withstand more force before tearing. These crosslinkers belong to a class of molecules known as mechanophores, which change their shape or other properties in response to mechanical force.

“These molecules can be useful for making polymers that would be stronger in response to force. You apply some stress to them, and rather than cracking or breaking, you instead see something that has higher resilience,” says Heather Kulik, the Lammot du Pont Professor of Chemical Engineering at MIT, who is also a professor of chemistry and the senior author of the study.

The crosslinkers that the researchers identified in this study are iron-containing compounds known as ferrocenes, which until now had not been broadly explored for their potential as mechanophores. Experimentally evaluating a single mechanophore can take weeks, but the researchers showed that they could use a machine-learning model to dramatically speed up this process.

MIT postdoc Ilia Kevlishvili is the lead author of the open-access paper, which appeared Friday in ACS Central Science. Other authors include Jafer Vakil, a Duke graduate student; David Kastner and Xiao Huang, both MIT graduate students; and Stephen Craig, a professor of chemistry at Duke.

The weakest link

Mechanophores are molecules that respond to force in unique ways, typically by changing their color, structure, or other properties. In the new study, the MIT and Duke team wanted to investigate whether they could be used to help make polymers more resilient to damage.

The new work builds on a 2023 study from Craig and Jeremiah Johnson, the A. Thomas Guertin Professor of Chemistry at MIT, and their colleagues. In that work, the researchers found that, surprisingly, incorporating weak crosslinkers into a polymer network can make the overall material stronger. When materials with these weak crosslinkers are stretched to the breaking point, any cracks propagating through the material try to avoid the stronger bonds and go through the weaker bonds instead. This means the crack has to break more bonds than it would if all of the bonds were the same strength.

To find new ways to exploit that phenomenon, Craig and Kulik joined forces to try to identify mechanophores that could be used as weak crosslinkers.

“We had this new mechanistic insight and opportunity, but it came with a big challenge: Of all possible compositions of matter, how do we zero in on the ones with the greatest potential?” Craig says. “Full credit to Heather and Ilia for both identifying this challenge and devising an approach to meet it.”

Discovering and characterizing mechanophores is a difficult task that requires either time-consuming experiments or computationally intense simulations of molecular interactions. Most of the known mechanophores are organic compounds, such as cyclobutane, which was used as a crosslinker in the 2023 study.

In the new study, the researchers wanted to focus on molecules known as ferrocenes, which are believed to hold potential as mechanophores. Ferrocenes are organometallic compounds that have an iron atom sandwiched between two carbon-containing rings. Those rings can have different chemical groups added to them, which alter their chemical and mechanical properties.

Many ferrocenes are used as pharmaceuticals or catalysts, and a handful are known to be good mechanophores, but most have not been evaluated for that use. Experimental tests on a single potential mechanophore can take several weeks, and computational simulations, while faster, still take a couple of days. Evaluating thousands of candidates using these strategies is a daunting task.

Realizing that a machine-learning approach could dramatically speed up the characterization of these molecules, the MIT and Duke team decided to use a neural network to identify ferrocenes that could be promising mechanophores.

They began with information from a database known as the Cambridge Structural Database, which contains the structures of 5,000 different ferrocenes that have already been synthesized.

“We knew that we didn’t have to worry about the question of synthesizability, at least from the perspective of the mechanophore itself. This allowed us to pick a really large space to explore with a lot of chemical diversity, that also would be synthetically realizable,” Kevlishvili says.

First, the researchers performed computational simulations for about 400 of these compounds, allowing them to calculate how much force is necessary to pull atoms apart within each molecule. For this application, they were looking for molecules that would break apart quickly, as these weak links could make polymer materials more resistant to tearing.

Then they used this data, along with information on the structure of each compound, to train a machine-learning model. This model was able to predict the force needed to activate the mechanophore, which in turn influences resistance to tearing, for the remaining 4,500 compounds in the database, plus an additional 7,000 compounds that are similar to those in the database but have some atoms rearranged.

The researchers discovered two main features that seemed likely to increase tear resistance. One was interactions between the chemical groups that are attached to the ferrocene rings. Additionally, the presence of large, bulky molecules attached to both rings of the ferrocene made the molecule more likely to break apart in response to applied forces.

While the first of these features was not surprising, the second trait was not something a chemist would have predicted beforehand, and could not have been detected without AI, the researchers say. “This was something truly surprising,” Kulik says.

Tougher plastics

Once the researchers identified about 100 promising candidates, Craig’s lab at Duke synthesized a polymer material incorporating one of them, known as m-TMS-Fc. Within the material, m-TMS-Fc acts as a crosslinker, connecting the polymer strands that make up polyacrylate, a type of plastic.

By applying force to each polymer until it tore, the researchers found that the weak m-TMS-Fc linker produced a strong, tear-resistant polymer. This polymer turned out to be about four times tougher than polymers made with standard ferrocene as the crosslinker.

“That really has big implications because if we think of all the plastics that we use and all the plastic waste accumulation, if you make materials tougher, that means their lifetime will be longer. They will be usable for a longer period of time, which could reduce plastic production in the long term,” Kevlishvili says.

The researchers now hope to use their machine-learning approach to identify mechanophores with other desirable properties, such as the ability to change color or become catalytically active in response to force. Such materials could be used as stress sensors or switchable catalysts, and they could also be useful for biomedical applications such as drug delivery.

In those studies, the researchers plan to focus on ferrocenes and other metal-containing mechanophores that have already been synthesized but whose properties are not fully understood.

“Transition metal mechanophores are relatively underexplored, and they’re probably a little bit more challenging to make,” Kulik says. “This computational workflow can be broadly used to enlarge the space of mechanophores that people have studied.”

The research was funded by the National Science Foundation Center for the Chemistry of Molecularly Optimized Networks (MONET).

M.C. Escher’s artwork is a gateway into a world of depth-defying optical illusions, featuring “impossible objects” that break the laws of physics with convoluted geometries. What you perceive his illustrations to be depends on your point of view — for example, a person seemingly walking upstairs may be heading down the steps if you tilt your head sideways. 

Computer graphics scientists and designers can recreate these illusions in 3D, but only by bending or cutting a real shape and positioning it at a particular angle. This workaround has downsides, though: Changing the smoothness or lighting of the structure will expose that it isn’t actually an optical illusion, which also means you can’t accurately solve geometry problems on it.

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a unique approach to represent “impossible” objects in a more versatile way. Their “Meschers” tool converts images and 3D models into 2.5-dimensional structures, creating Escher-like depictions of things like windows, buildings, and even donuts. The approach helps users relight, smooth out, and study unique geometries while preserving their optical illusion.

This tool could assist geometry researchers with calculating the distance between two points on a curved impossible surface (“geodesics”) and simulating how heat dissipates over it (“heat diffusion”). It could also help artists and computer graphics scientists create physics-breaking designs in multiple dimensions.

Lead author and MIT PhD student Ana Dodik aims to design computer graphics tools that aren’t limited to replicating reality, enabling artists to express their intent independently of whether a shape can be realized in the physical world. “Using Meschers, we’ve unlocked a new class of shapes for artists to work with on the computer,” she says. “They could also help perception scientists understand the point at which an object truly becomes impossible.”

Dodik and her colleagues will present their paper at the SIGGRAPH conference in August.

Making impossible objects possible

Impossible objects can’t be fully replicated in 3D. Their constituent parts often look plausible, but these parts don’t glue together properly when assembled in 3D. But what can be computationally imitated, as the CSAIL researchers found out, is the process of how we perceive these shapes.

Take the Penrose Triangle, for instance. The object as a whole is physically impossible because the depths don’t “add up,” but we can recognize real-world 3D shapes (like its three L-shaped corners) within it. These smaller regions can be realized in 3D — a property called “local consistency” — but when we try to assemble them together, they don’t form a globally consistent shape.

The Meschers approach models’ locally consistent regions without forcing them to be globally consistent, piecing together an Escher-esque structure. Behind the scenes, Meschers represents impossible objects as if we know their x and y coordinates in the image, as well as differences in z coordinates (depth) between neighboring pixels; the tool uses these differences in depth to reason about impossible objects indirectly.

The many uses of Meschers

In addition to rendering impossible objects, Meschers can subdivide their structures into smaller shapes for more precise geometry calculations and smoothing operations. This process enabled the researchers to reduce visual imperfections of impossible shapes, such as a red heart outline they thinned out.

The researchers also tested their tool on an “impossibagel,” where a bagel is shaded in a physically impossible way. Meschers helped Dodik and her colleagues simulate heat diffusion and calculate geodesic distances between different points of the model.

“Imagine you’re an ant traversing this bagel, and you want to know how long it’ll take you to get across, for example,” says Dodik. “In the same way, our tool could help mathematicians analyze the underlying geometry of impossible shapes up close, much like how we study real-world ones.”

Much like a magician, the tool can create optical illusions out of otherwise practical objects, making it easier for computer graphics artists to create impossible objects. It can also use “inverse rendering” tools to convert drawings and images of impossible objects into high-dimensional designs. 

“Meschers demonstrates how computer graphics tools don’t have to be constrained by the rules of physical reality,” says senior author Justin Solomon, associate professor of electrical engineering and computer science and leader of the CSAIL Geometric Data Processing Group. “Incredibly, artists using Meschers can reason about shapes that we will never find in the real world.”

Meschers can also aid computer graphics artists with tweaking the shading of their creations, while still preserving an optical illusion. This versatility would allow creatives to change the lighting of their art to depict a wider variety of scenes (like a sunrise or sunset) — as Meschers demonstrated by relighting a model of a dog on a skateboard.

Despite its versatility, Meschers is just the start for Dodik and her colleagues. The team is considering designing an interface to make the tool easier to use while building more elaborate scenes. They’re also working with perception scientists to see how the computer graphics tool can be used more broadly.

Dodik and Solomon wrote the paper with CSAIL affiliates Isabella Yu ’24, SM ’25; PhD student Kartik Chandra SM ’23; MIT professors Jonathan Ragan-Kelley and Joshua Tenenbaum; and MIT Assistant Professor Vincent Sitzmann. 

Their work was supported, in part, by the MIT Presidential Fellowship, the Mathworks Fellowship, the Hertz Foundation, the U.S. National Science Foundation, the Schmidt Sciences AI2050 fellowship, MIT Quest for Intelligence, the U.S. Army Research Office, U.S. Air Force Office of Scientific Research, SystemsThatLearn@CSAIL initiative, Google, the MIT–IBM Watson AI Laboratory, from the Toyota–CSAIL Joint Research Center, Adobe Systems, the Singapore Defence Science and Technology Agency, and the U.S. Intelligence Advanced Research Projects Activity.

In the push to shrink and enhance technologies that control light, MIT researchers have unveiled a new platform that pushes the limits of modern optics through nanophotonics, the manipulation of light on the nanoscale, or billionths of a meter.

The result is a class of ultracompact optical devices that are not only smaller and more efficient than existing technologies, but also dynamically tunable, or switchable, from one optical mode to another. Until now, this has been an elusive combination in nanophotonics.

The work is reported in the July 8 issue of Nature Photonics.

“This work marks a significant step toward a future in which nanophotonic devices are not only compact and efficient, but also reprogrammable and adaptive, capable of dynamically responding to external inputs. The  marriage of emerging quantum materials and established nanophotonics architectures will surely bring advances to both fields,” says Riccardo Comin, MIT’s Class of 1947 Career Development Associate Professor of Physics and leader of the work. Comin is also affiliated with MIT’s Materials Research Laboratory and Research Laboratory of Electronics (RLE).

Comin’s colleagues on the work are Ahmet Kemal Demir, an MIT graduate student in physics; Luca Nessi, a former MIT postdoc who is now a postdoc at Politecnico di Milano; Sachin Vaidya, a postdoc in RLE; Connor A. Occhialini PhD ’24, who is now a postdoc at Columbia University; and Marin Soljačić, the Cecil and Ida Green Professor of Physics at MIT.

Demir and Nessi are co-first authors of the Nature Photonics paper.

Toward new nanophotonic materials

Nanophotonics has traditionally relied on materials like silicon, silicon nitride, or titanium dioxide. These are the building blocks of devices that guide and confine light using structures such as waveguides, resonators, and photonic crystals. The latter are periodic arrangements of materials that control how light propagates, much like how a semiconductor crystal affects electron motion.

While highly effective, these materials are constrained by two major limitations. The first involves their refractive indices. These are a measure of how strongly a material interacts with light; the higher the refractive index, the more the material “grabs” or interacts with the light, bending it more sharply and slowing it down more. The refractive indices of silicon and other traditional nanophotonic materials are often modest, which limits how tightly light can be confined and how small optical devices can be made.

A second major limitation of traditional nanophotonic materials: once a structure is fabricated, its optical behavior is essentially fixed. There is usually no way to significantly reconfigure how it responds to light without physically altering it. “Tunability is essential for many next-gen photonics applications, enabling adaptive imaging, precision sensing, reconfigurable light sources, and trainable optical neural networks,” says Vaidya.

Introducing chromium sulfide bromide

These are the longstanding challenges that chromium sulfide bromide (CrSBr) is poised to solve. CrSBr is a layered quantum material with a rare combination of magnetic order and strong optical response. Central to its unique optical properties are excitons: quasiparticles formed when a material absorbs light and an electron is excited, leaving behind a positively charged “hole.” The electron and hole remain bound together by electrostatic attraction, forming a sort of neutral particle that can strongly interact with light.

In CrSBr, excitons dominate the optical response and are highly sensitive to magnetic fields, which means they can be manipulated using external controls.

Because of these excitons, CrSBr exhibits an exceptionally large refractive index that allows researchers to sculpt the material to fabricate optical structures like photonic crystals that are up to an order of magnitude thinner than those made from traditional materials. “We can make optical structures as thin as 6 nanometers, or just seven layers of atoms stacked on top of each other,” says Demir.

And crucially, by applying a modest magnetic field, the MIT researchers were able to continuously and reversibly switch the optical mode. In other words, they demonstrated the ability to dynamically change how light flows through the nanostructure, all without any moving parts or changes in temperature. “This degree of control is enabled by a giant, magnetically induced shift in the refractive index, far beyond what is typically achievable in established photonic materials,” says Demir.

In fact, the interaction between light and excitons in CrSBr is so strong that it leads to the formation of polaritons, hybrid light-matter particles that inherit properties from both components. These polaritons enable new forms of photonic behavior, such as enhanced nonlinearities and new regimes of quantum light transport. And unlike conventional systems that require external optical cavities to reach this regime, CrSBr supports polaritons intrinsically.

While this demonstration uses standalone CrSBr flakes, the material can also be integrated into existing photonic platforms, such as integrated photonic circuits. This makes CrSBr immediately relevant to real-world applications, where it can serve as a tunable layer or component in otherwise passive devices.

The MIT results were achieved at very cold temperatures of up to 132 kelvins (-222 degrees Fahrenheit). Although this is below room temperature, there are compelling use cases, such as quantum simulation, nonlinear optics, and reconfigurable polaritonic platforms, where the unparalleled tunability of CrSBr could justify operation in cryogenic environments.

In other words, says Demir, “CrSBr is so unique with respect to other common materials that even going down to cryogenic temperatures will be worth the trouble, hopefully.”

That said, the team is also exploring related materials with higher magnetic ordering temperatures to enable similar functionality at more accessible conditions.

This work was supported by the U.S. Department of Energy, the U.S. Army Research Office, and a MathWorks Science Fellowship. The work was performed in part at MIT.nano.

A fundamental problem for governments is getting citizens to comply with their laws and policies. They can’t monitor everyone and catch all the rule-breakers. “It’s a logistical impossibility,” says Lily L. Tsai, MIT’s Ford Professor of Political Science and the director and founder of the MIT Governance Lab.

Instead, governments need citizens to choose to follow the rules of their own accord. “As a government, you have to rely on them to voluntarily comply with the laws, policies, and regulations that are put into place,” Tsai says.

One particularly important thing governments need citizens to do is pay their taxes. In a paper in the October issue of the journal World Development, Tsai and her co-authors, including Minh Trinh ’22, a graduate of the Department of Political Science, look at different factors that might affect compliance with property tax laws in China. They found that study participants in an in-person tax-paying experiment were more likely to pay their taxes if government officials were monitoring and punishing corruption.

“When people think that government authorities are motivated by the public good, have moral character, and have integrity, then the requests that those authorities make of citizens are more likely to seem legitimate, and so they’re more likely to pay their taxes,” Tsai says.

In China, only two cities, Chongqing and Shanghai, collect property taxes. Officials have been concerned that citizens might resist property taxes because homeownership is the main source of urban household wealth in China. Private homeownership accounts for 64 percent of household wealth in China, compared to only 29 percent in the United States.

Tsai and her co-authors wanted to test how governments might make people more willing to pay their property taxes. Researchers have theorized that citizens are more likely to comply with tax laws when they feel like they’re getting something in return from the government. The government can be responsive to citizens’ demands for public services, for example. Or the government can punish officials who are corrupt or perform poorly.

In the first part of the study, a survey of Chinese citizens, respondents expressed preferences for different hypothetical property tax policies. The results suggested that participants wanted the government to be responsive to their needs and to hold officials accountable. People preferred a policy that allowed for citizen input on the use of tax revenue over one that did not, and a policy that allowed for the sanctioning of corrupt officials garnered more support than a policy that did not.

Survey participants also preferred a lighter penalty for not paying their taxes over a harsher penalty, and they supported a tax exemption for first apartments. Interestingly to the researchers, policies that allowed for government responsiveness and accountability received roughly the same support as these policies with economic benefits. “This is evidence to show that we should really pay attention to non-economic factors, because they can have similar magnitudes of impact on tax-paying behavior,” Tsai says.

For the second stage of the study, researchers recruited people for a lab experiment in Shanghai (one of the two cities that collects property taxes). Participants played a game on an iPad in which they chose repeatedly whether or not to pay property taxes. At the end of the game, they received an amount of real money that varied depending on how they and other participants played the game.

Participants were then randomly split into different groups. In one group, participants were given an opportunity to voice their preference for how their property tax revenue was used. Some were told the government incorporated their feedback, while others were told their preferences were not considered — in other words, participants learned whether or not the government was responsive to their needs. In another group, participants learned that a corrupt official had stolen money from property tax revenue. Some were told that the official had been caught and punished, while others were told the official got away with stealing.

The researchers measured whether game players’ willingness to pay property taxes changed after receiving this new information. They found that while the willingness of players who learned the government was responsive to their needs did not change significantly, players who learned the government punished corrupt officials paid their property taxes more frequently.

“It was kind of amazing to see that people care a lot about whether or not higher-level authorities are making sure that tax dollars are not being wasted through corruption,” Tsai says. She argues in her 2021 book, “When People Want Punishment: Retributive Justice and the Puzzle of Authoritarian Popularity,” that when authorities are willing to punish their own officials, it may signal to people that leaders have moral integrity and share the values of ordinary people, making them appear more legitimate.

While the researchers expected to see government responsiveness affect tax payment as well, Tsai says it’s not totally surprising that for people living in places without direct channels for citizen input, the opportunity to participate in the decision-making process in a lab setting might not resonate as strongly.

The findings don’t mean that government responsiveness isn’t important. But they suggest that even when there aren’t opportunities for citizens to make their voices heard, there are other ways for governments to appear legitimate and get people to comply with rules voluntarily.

As the strength of democratic institutions declines globally, scholars wonder whether perceptions of governments’ legitimacy will decline at the same time. “These findings suggest that maybe that’s not necessarily the case,” Tsai says.

Imagine a ball bouncing down a flight of stairs. Now think about a cascade of water flowing down those same stairs. The ball and the water behave very differently, and it turns out that your brain has different regions for processing visual information about each type of physical matter.

In a new study, MIT neuroscientists have identified parts of the brain’s visual cortex that respond preferentially when you look at “things” — that is, rigid or deformable objects like a bouncing ball. Other brain regions are more activated when looking at “stuff” — liquids or granular substances such as sand.

This distinction, which has never been seen in the brain before, may help the brain plan how to interact with different kinds of physical materials, the researchers say.

“When you’re looking at some fluid or gooey stuff, you engage with it in different way than you do with a rigid object. With a rigid object, you might pick it up or grasp it, whereas with fluid or gooey stuff, you probably are going to have to use a tool to deal with it,” says Nancy Kanwisher, the Walter A. Rosenblith Professor of Cognitive Neuroscience; a member of the McGovern Institute for Brain Research and MIT’s Center for Brains, Minds, and Machines; and the senior author of the study.

MIT postdoc Vivian Paulun, who is joining the faculty of the University of Wisconsin at Madison this fall, is the lead author of the paper, which appears today in the journal Current Biology. RT Pramod, an MIT postdoc, and Josh Tenenbaum, an MIT professor of brain and cognitive sciences, are also authors of the study.

Stuff vs. things

Decades of brain imaging studies, including early work by Kanwisher, have revealed regions in the brain’s ventral visual pathway that are involved in recognizing the shapes of 3D objects, including an area called the lateral occipital complex (LOC). A region in the brain’s dorsal visual pathway, known as the frontoparietal physics network (FPN), analyzes the physical properties of materials, such as mass or stability.

Although scientists have learned a great deal about how these pathways respond to different features of objects, the vast majority of these studies have been done with solid objects, or “things.”

“Nobody has asked how we perceive what we call ‘stuff’ — that is, liquids or sand, honey, water, all sorts of gooey things. And so we decided to study that,” Paulun says.

These gooey materials behave very differently from solids. They flow rather than bounce, and interacting with them usually requires containers and tools such as spoons. The researchers wondered if these physical features might require the brain to devote specialized regions to interpreting them.

To explore how the brain processes these materials, Paulun used a software program designed for visual effects artists to create more than 100 video clips showing different types of things or stuff interacting with the physical environment. In these videos, the materials could be seen sloshing or tumbling inside a transparent box, being dropped onto another object, or bouncing or flowing down a set of stairs.

The researchers used functional magnetic resonance imaging (fMRI) to scan the visual cortex of people as they watched the videos. They found that both the LOC and the FPN respond to “things” and “stuff,” but that each pathway has distinctive subregions that respond more strongly to one or the other.

“Both the ventral and the dorsal visual pathway seem to have this subdivision, with one part responding more strongly to ‘things,’ and the other responding more strongly to ‘stuff,’” Paulun says. “We haven’t seen this before because nobody has asked that before.”

Roland Fleming, a professor of experimental psychology at Justus Liebig University of Geissen, described the findings as a “major breakthrough in the scientific understanding of how our brains represent the physical properties of our surrounding world.”

“We’ve known the distinction exists for a long time psychologically, but this is the first time that it’s been really mapped onto separate cortical structures in the brain. Now we can investigate the different computations that the distinct brain regions use to process and represent objects and materials,” says Fleming, who was not involved in the study.

Physical interactions

The findings suggest that the brain may have different ways of representing these two categories of material, similar to the artificial physics engines that are used to create video game graphics. These engines usually represent a 3D object as a mesh, while fluids are represented as sets of particles that can be rearranged.

“The interesting hypothesis that we can draw from this is that maybe the brain, similar to artificial game engines, has separate computations for representing and simulating ‘stuff’ and ‘things.’ And that would be something to test in the future,” Paulun says.

The researchers also hypothesize that these regions may have developed to help the brain understand important distinctions that allow it to plan how to interact with the physical world. To further explore this possibility, the researchers plan to study whether the areas involved in processing rigid objects are also active when a brain circuit involved in planning to grasp objects is active.

They also hope to look at whether any of the areas within the FPN correlate with the processing of more specific features of materials, such as the viscosity of liquids or the bounciness of objects. And in the LOC, they plan to study how the brain represents changes in the shape of fluids and deformable substances.

The research was funded by the German Research Foundation, the U.S. National Institutes of Health, and a U.S. National Science Foundation grant to the Center for Brains, Minds, and Machines.

Salt creeping, a phenomenon that occurs in both natural and industrial processes, describes the collection and migration of salt crystals from evaporating solutions onto surfaces. Once they start collecting, the crystals climb, spreading away from the solution. This creeping behavior, according to researchers, can cause damage or be harnessed for good, depending on the context. New research published June 30 in the journal Langmuir is the first to show salt creeping at a single-crystal scale and beneath a liquid’s meniscus.

“The work not only explains how salt creeping begins, but why it begins and when it does,” says Joseph Phelim Mooney, a postdoc in the MIT Device Research Laboratory and one of the authors of the new study. “We hope this level of insight helps others, whether they’re tackling water scarcity, preserving ancient murals, or designing longer-lasting infrastructure.”

The work is the first to directly visualize how salt crystals grow and interact with surfaces underneath a liquid meniscus, something that’s been theorized for decades but never actually imaged or confirmed at this level, and it offers fundamental insights that could impact a wide range of fields — from mineral extraction and desalination to anti-fouling coatings, membrane design for separation science, and even art conservation, where salt damage is a major threat to heritage materials.

In civil engineering applications, for example, the research can help explain why and when salt crystals start growing across surfaces like concrete, stone, or building materials. “These crystals can exert pressure and cause cracking or flaking, reducing the long-term durability of structures,” says Mooney. “By pinpointing the moment when salt begins to creep, engineers can better design protective coatings or drainage systems to prevent this form of degradation.”

For a field like art conservation, where salt can be devastating to murals, frescoes, and ancient artifacts, often forming beneath the surface before visible damage appears, the work can help identify the exact conditions that cause salt to start moving and spreading, allowing conservators to act earlier and more precisely to protect heritage objects.

The work began during Mooney’s Marie Curie Fellowship at MIT. “I was focused on improving desalination systems and quickly ran into [salt buildup as] a major roadblock,” he says. “[Salt] was everywhere, coating surfaces, clogging flow paths, and undermining the efficiency of our designs. I realized we didn’t fully understand how or why salt starts creeping across surfaces in the first place.”

That experience led Mooney to team up with colleagues to dig into the fundamentals of salt crystallization at the air–liquid–solid interface. “We wanted to zoom in, to really see the moment salt begins to move, so we turned to in situ X-ray microscopy,” he says. “What we found gave us a whole new way to think about surface fouling, material degradation, and controlled crystallization.”

The new research may, in fact, allow better control of a crystallization processes required to remove salt from water in zero-liquid discharge systems. It can also be used to explain how and when scaling happens on equipment surfaces, and may support emerging climate technologies that depend on smart control of evaporation and crystallization.

The work also supports mineral and salt extraction applications, where salt creeping can be both a bottleneck and an opportunity. In these applications, Mooney says, “by understanding the precise physics of salt formation at surfaces, operators can optimize crystal growth, improving recovery rates and reducing material losses.”

Mooney’s co-authors on the paper include fellow MIT Device Lab researchers Omer Refet Caylan, Bachir El Fil (now an associate professor at Georgia Tech), and Lenan Zhang (now an associate professor at Cornell University); Jeff Punch and Vanessa Egan of the University of Limerick; and Jintong Gao of Cornell.

The research was conducted using in situ X-ray microscopy. Mooney says the team’s big realization moment occurred when they were able to observe a single salt crystal pinning itself to the surface, which kicked off a cascading chain reaction of growth.

“People had speculated about this, but we captured it on X-ray for the first time. It felt like watching the microscopic moment where everything tips, the ignition points of a self-propagating process,” says Mooney. “Even more surprising was what followed: The salt crystal didn’t just grow passively to fill the available space. It pierced through the liquid-air interface and reshaped the meniscus itself, setting up the perfect conditions for the next crystal. That subtle, recursive mechanism had never been visually documented before — and seeing it play out in real time completely changed how we thought about salt crystallization.”

The paper, “In Situ X-ray Microscopy Unraveling the Onset of Salt Creeping at a Single-Crystal Level,” is available now in the journal Langmuir. Research was conducted in MIT.nano. 

If you rotate an image of a molecular structure, a human can tell the rotated image is still the same molecule, but a machine-learning model might think it is a new data point. In computer science parlance, the molecule is “symmetric,” meaning the fundamental structure of that molecule remains the same if it undergoes certain transformations, like rotation.

If a drug discovery model doesn’t understand symmetry, it could make inaccurate predictions about molecular properties. But despite some empirical successes, it’s been unclear whether there is a computationally efficient method to train a good model that is guaranteed to respect symmetry.

A new study by MIT researchers answers this question, and shows the first method for machine learning with symmetry that is provably efficient in terms of both the amount of computation and data needed.

These results clarify a foundational question, and they could aid researchers in the development of more powerful machine-learning models that are designed to handle symmetry. Such models would be useful in a variety of applications, from discovering new materials to identifying astronomical anomalies to unraveling complex climate patterns.

“These symmetries are important because they are some sort of information that nature is telling us about the data, and we should take it into account in our machine-learning models. We’ve now shown that it is possible to do machine-learning with symmetric data in an efficient way,” says Behrooz Tahmasebi, an MIT graduate student and co-lead author of this study.

He is joined on the paper by co-lead author and MIT graduate student Ashkan Soleymani; Stefanie Jegelka, an associate professor of electrical engineering and computer science (EECS) and a member of the Institute for Data, Systems, and Society (IDSS) and the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Patrick Jaillet, the Dugald C. Jackson Professor of Electrical Engineering and Computer Science and a principal investigator in the Laboratory for Information and Decision Systems (LIDS). The research was recently presented at the International Conference on Machine Learning.

Studying symmetry

Symmetric data appear in many domains, especially the natural sciences and physics. A model that recognizes symmetries is able to identify an object, like a car, no matter where that object is placed in an image, for example.

Unless a machine-learning model is designed to handle symmetry, it could be less accurate and prone to failure when faced with new symmetric data in real-world situations. On the flip side, models that take advantage of symmetry could be faster and require fewer data for training.

But training a model to process symmetric data is no easy task.

One common approach is called data augmentation, where researchers transform each symmetric data point into multiple data points to help the model generalize better to new data. For instance, one could rotate a molecular structure many times to produce new training data, but if researchers want the model to be guaranteed to respect symmetry, this can be computationally prohibitive.

An alternative approach is to encode symmetry into the model’s architecture. A well-known example of this is a graph neural network (GNN), which inherently handles symmetric data because of how it is designed.

“Graph neural networks are fast and efficient, and they take care of symmetry quite well, but nobody really knows what these models are learning or why they work. Understanding GNNs is a main motivation of our work, so we started with a theoretical evaluation of what happens when data are symmetric,” Tahmasebi says.

They explored the statistical-computational tradeoff in machine learning with symmetric data. This tradeoff means methods that require fewer data can be more computationally expensive, so researchers need to find the right balance.

Building on this theoretical evaluation, the researchers designed an efficient algorithm for machine learning with symmetric data.

Mathematical combinations

To do this, they borrowed ideas from algebra to shrink and simplify the problem. Then, they reformulated the problem using ideas from geometry that effectively capture symmetry.

Finally, they combined the algebra and the geometry into an optimization problem that can be solved efficiently, resulting in their new algorithm.

“Most of the theory and applications were focusing on either algebra or geometry. Here we just combined them,” Tahmasebi says.

The algorithm requires fewer data samples for training than classical approaches, which would improve a model’s accuracy and ability to adapt to new applications.

By proving that scientists can develop efficient algorithms for machine learning with symmetry, and demonstrating how it can be done, these results could lead to the development of new neural network architectures that could be more accurate and less resource-intensive than current models.

Scientists could also use this analysis as a starting point to examine the inner workings of GNNs, and how their operations differ from the algorithm the MIT researchers developed.

“Once we know that better, we can design more interpretable, more robust, and more efficient neural network architectures,” adds Soleymani.

This research is funded, in part, by the National Research Foundation of Singapore, DSO National Laboratories of Singapore, the U.S. Office of Naval Research, the U.S. National Science Foundation, and an Alexander von Humboldt Professorship.

Researchers from MIT and elsewhere have designed a novel transmitter chip that significantly improves the energy efficiency of wireless communications, which could boost the range and battery life of a connected device.

Their approach employs a unique modulation scheme to encode digital data into a wireless signal, which reduces the amount of error in the transmission and leads to more reliable communications.

The compact, flexible system could be incorporated into existing internet-of-things devices to provide immediate gains, while also meeting the more stringent efficiency requirements of future 6G technologies.

The versatility of the chip could make it well-suited for a range of applications that require careful management of energy for communications, such as industrial sensors that continuously monitor factory conditions and smart appliances that provide real-time notifications.

“By thinking outside the box, we created a more efficient, intelligent circuit for next-generation devices that is also even better than the state-of-the-art for legacy architectures. This is just one example of how adopting a modular approach to allow for adaptability can drive innovation at every level,” says Muriel Médard, the School of Science NEC Professor of Software Science and Engineering, a professor in the MIT Department of Electrical Engineering and Computer Science (EECS), and co-author of a paper on the new transmitter.

Médard’s co-authors include Timur Zirtiloglu, the lead author and a graduate student at Boston University; Arman Tan, a graduate student at BU; Basak Ozaydin, an MIT graduate student in EECS; Ken Duffy, a professor at Northeastern University; and Rabia Tugce Yazicigil, associate professor of electrical and computer engineering at BU. The research was recently presented at the IEEE Radio Frequency Circuits Symposium.

Optimizing transmissions

In wireless devices, a transmitter converts digital data into an electromagnetic signal that is sent over the airwaves to a receiver. The transmitter does this by mapping digital bits to symbols that represent the amplitude and phase of the electromagnetic signal, which is a process called modulation.

Traditional systems transmit signals that are evenly spaced by creating a uniform pattern of symbols, which helps avoid interference. But this uniform structure lacks adaptability and can be inefficient, since wireless channel conditions are dynamic and often change rapidly.

As an alternative, optimal modulation schemes follow a non-uniform pattern that can adapt to changing channel conditions, maximizing the amount of data transmitted while minimizing energy usage.

But while optimal modulation can be more energy efficient, it is also more susceptible to errors, especially in crowded wireless environments. When the signals aren’t uniform in length, it can be harder for the receiver to distinguish between symbols and noise that squeezed into the transmission.

To overcome this problem, the MIT transmitter adds a small amount of padding, in the form of extra bits between symbols, so that every transmission is the same length.

This helps the receiver identify the beginning and end of each transmission, preventing misinterpretation of the message. However, the device enjoys the energy efficiency gains of using a non-uniform, optimal modulation scheme.

This approach works because of a technique the researchers previously developed known as GRAND, which is a universal decoding algorithm that crack any code by guessing the noise that affected the transmission.

Here, they employ a GRAND-inspired algorithm to adjust the length of the received transmission by guessing the extra bits that have been added. In this way, the receiver can effectively reconstruct the original message.

“Now, thanks to GRAND, we can have a transmitter that is capable of doing these more efficient transmissions with non-uniform constellations of data, and we can see the gains,” Médard says.

A flexible circuit

The new chip, which has a compact architecture that allows the researchers to integrate additional efficiency-boosting methods, enabled transmissions with only about one-quarter the amount of signal error of methods that use optimal modulation.

Surprisingly, the device also achieved significantly lower error rates than transmitters that use traditional modulation.

“The traditional approach has become so ingrained that it was challenging to not get lured back to the status quo, especially since we were changing things that we often take for granted and concepts we’ve been teaching for decades,” Médard says.

This innovative architecture could be used to improve the energy efficiency and reliability of current wireless communication devices, while also offering the flexibility to be incorporated into future devices that employ optimal modulation.

Next, the researchers want to adapt their approach to leverage additional techniques that could boost efficiency and reduce the error rates in wireless transmissions.

“This optimal modulation transmitter radio frequency integrated circuit is a game-changing innovation over the traditional RF signal modulation. It’s set to play a major role for the next generation of wireless connectivity such as 6G and Wi-Fi,” says Rocco Tam, NXP Fellow for Wireless Connectivity SoC Research and Development at NXP Semiconductors, who was not involved with this research.

This work is supported, in part, by the U.S. Defense Advanced Research Projects Agency (DARPA), the National Science Foundation (NSF), and the Texas Analog Center for Excellence. 

A lot of attention has been paid to how climate change can drive biodiversity loss. Now, MIT researchers have shown the reverse is also true: Reductions in biodiversity can jeopardize one of Earth’s most powerful levers for mitigating climate change.

In a paper published in PNAS, the researchers showed that following deforestation, naturally-regrowing tropical forests, with healthy populations of seed-dispersing animals, can absorb up to four times more carbon than similar forests with fewer seed-dispersing animals.

Because tropical forests are currently Earth’s largest land-based carbon sink, the findings improve our understanding of a potent tool to fight climate change.

“The results underscore the importance of animals in maintaining healthy, carbon-rich tropical forests,” says Evan Fricke, a research scientist in the MIT Department of Civil and Environmental Engineering and the lead author of the new study. “When seed-dispersing animals decline, we risk weakening the climate-mitigating power of tropical forests.”

Fricke’s co-authors on the paper include César Terrer, the Tianfu Career Development Associate Professor at MIT; Charles Harvey, an MIT professor of civil and environmental engineering; and Susan Cook-Patton of The Nature Conservancy.

The study combines a wide array of data on animal biodiversity, movement, and seed dispersal across thousands of animal species, along with carbon accumulation data from thousands of tropical forest sites.

The researchers say the results are the clearest evidence yet that seed-dispersing animals play an important role in forests’ ability to absorb carbon, and that the findings underscore the need to address biodiversity loss and climate change as connected parts of a delicate ecosystem rather as separate problems in isolation.

“It’s been clear that climate change threatens biodiversity, and now this study shows how biodiversity losses can exacerbate climate change,” Fricke says. “Understanding that two-way street helps us understand the connections between these challenges, and how we can address them. These are challenges we need to tackle in tandem, and the contribution of animals to tropical forest carbon shows that there are win-wins possible when supporting biodiversity and fighting climate change at the same time.”

Putting the pieces together

The next time you see a video of a monkey or bird enjoying a piece of fruit, consider that the animals are actually playing an important role in their ecosystems. Research has shown that by digesting the seeds and defecating somewhere else, animals can help with the germination, growth, and long-term survival of the plant.

Fricke has been studying animals that disperse seeds for nearly 15 years. His previous research has shown that without animal seed dispersal, trees have lower survival rates and a harder time keeping up with environmental changes.

“We’re now thinking more about the roles that animals might play in affecting the climate through seed dispersal,” Fricke says. “We know that in tropical forests, where more than three-quarters of trees rely on animals for seed dispersal, the decline of seed dispersal could affect not just the biodiversity of forests, but how they bounce back from deforestation. We also know that all around the world, animal populations are declining.”

Regrowing forests is an often-cited way to mitigate the effects of climate change, but the influence of biodiversity on forests’ ability to absorb carbon has not been fully quantified, especially at larger scales.

For their study, the researchers combined data from thousands of separate studies and used new tools for quantifying disparate but interconnected ecological processes. After analyzing data from more than 17,000 vegetation plots, the researchers decided to focus on tropical regions, looking at data on where seed-dispersing animals live, how many seeds each animal disperses, and how they affect germination.

The researchers then incorporated data showing how human activity impacts different seed-dispersing animals’ presence and movement. They found, for example, that animals move less when they consume seeds in areas with a bigger human footprint.

Combining all that data, the researchers created an index of seed-dispersal disruption that revealed a link between human activities and declines in animal seed dispersal. They then analyzed the relationship between that index and records of carbon accumulation in naturally regrowing tropical forests over time, controlling for factors like drought conditions, the prevalence of fires, and the presence of grazing livestock.

“It was a big task to bring data from thousands of field studies together into a map of the disruption of seed dispersal,” Fricke says. “But it lets us go beyond just asking what animals are there to actually quantifying the ecological roles those animals are playing and understanding how human pressures affect them.”

The researchers acknowledged that the quality of animal biodiversity data could be improved and introduces uncertainty into their findings. They also note that other processes, such as pollination, seed predation, and competition influence seed dispersal and can constrain forest regrowth. Still, the findings were in line with recent estimates.

“What’s particularly new about this study is we’re actually getting the numbers around these effects,” Fricke says. “Finding that seed dispersal disruption explains a fourfold difference in carbon absorption across the thousands of tropical regrowth sites included in the study points to seed dispersers as a major lever on tropical forest carbon.”

Quantifying lost carbon

In forests identified as potential regrowth sites, the researchers found seed-dispersal declines were linked to reductions in carbon absorption each year averaging 1.8 metric tons per hectare, equal to a reduction in regrowth of 57 percent.

The researchers say the results show natural regrowth projects will be more impactful in landscapes where seed-dispersing animals have been less disrupted, including areas that were recently deforested, are near high-integrity forests, or have higher tree cover.

“In the discussion around planting trees versus allowing trees to regrow naturally, regrowth is basically free, whereas planting trees costs money, and it also leads to less diverse forests,” Terrer says. “With these results, now we can understand where natural regrowth can happen effectively because there are animals planting the seeds for free, and we also can identify areas where, because animals are affected, natural regrowth is not going to happen, and therefore planting trees actively is necessary.”

To support seed-dispersing animals, the researchers encourage interventions that protect or improve their habitats and that reduce pressures on species, ranging from wildlife corridors to restrictions on wildlife trade. Restoring the ecological roles of seed dispersers is also possible by reintroducing seed-dispersing species where they’ve been lost or planting certain trees that attract those animals.

The findings could also make modeling the climate impact of naturally regrowing forests more accurate.

“Overlooking the impact of seed-dispersal disruption may overestimate natural regrowth potential in many areas and underestimate it in others,” the authors write.

The researchers believe the findings open up new avenues of inquiry for the field.

“Forests provide a huge climate subsidy by sequestering about a third of all human carbon emissions,” Terrer says. “Tropical forests are by far the most important carbon sink globally, but in the last few decades, their ability to sequester carbon has been declining. We will next explore how much of that decline is due to an increase in extreme droughts or fires versus declines in animal seed dispersal.”

Overall, the researchers hope the study helps improves our understanding of the planet’s complex ecological processes.

“When we lose our animals, we’re losing the ecological infrastructure that keeps our tropical forests healthy and resilient,” Fricke says.

The research was supported by the MIT Climate and Sustainability Consortium, the Government of Portugal, and the Bezos Earth Fund.

Scientists often seek new materials derived from polymers. Rather than starting a polymer search from scratch, they save time and money by blending existing polymers to achieve desired properties.

But identifying the best blend is a thorny problem. Not only is there a practically limitless number of potential combinations, but polymers interact in complex ways, so the properties of a new blend are challenging to predict.

To accelerate the discovery of new materials, MIT researchers developed a fully autonomous experimental platform that can efficiently identify optimal polymer blends.

The closed-loop workflow uses a powerful algorithm to explore a wide range of potential polymer blends, feeding a selection of combinations to a robotic system that mixes chemicals and tests each blend.

Based on the results, the algorithm decides which experiments to conduct next, continuing the process until the new polymer meets the user’s goals.

During experiments, the system autonomously identified hundreds of blends that outperformed their constituent polymers. Interestingly, the researchers found that the best-performing blends did not necessarily use the best individual components.

“I found that to be good confirmation of the value of using an optimization algorithm that considers the full design space at the same time,” says Connor Coley, the Class of 1957 Career Development Assistant Professor with shared appointments in the MIT departments of Chemical Engineering and Electrical Engineering and Computer Science and the MIT Schwarzman College of Computing, and senior author of a paper on this new approach. “If you consider the full formulation space, you can potentially find new or better properties. Using a different approach, you could easily overlook the underperforming components that happen to be the important parts of the best blend.”

This workflow could someday facilitate the discovery of polymer blend materials that lead to advancements like improved battery electrolytes, more cost-effective solar panels, or tailored nanoparticles for safer drug delivery.

Coley is joined on the paper by lead author Guangqi Wu, a former MIT postdoc who is now a Marie Skłodowska-Curie Postdoctoral Fellow at Oxford University; Tianyi Jin, an MIT graduate student; and Alfredo Alexander-Katz, the Michael and Sonja Koerner Professor in the MIT Department of Materials Science and Engineering. The work appears today in Matter.

Building better blends

When scientists design new polymer blends, they are faced with a nearly endless number of possible polymers to start with. Once they select a few to mix, they still must choose the composition of each polymer and the concentration of polymers in the blend.

“Having that large of a design space necessitates algorithmic solutions and higher-throughput workflows because you simply couldn’t test all the combinations using brute force,” Coley adds.

While researchers have studied autonomous workflows for single polymers, less work has focused on polymer blends because of the dramatically larger design space.

In this study, the MIT researchers sought new random heteropolymer blends, made by mixing two or more polymers with different structural features. These versatile polymers have shown particularly promising relevance to high-temperature enzymatic catalysis, a process that increases the rate of chemical reactions.

Their closed-loop workflow begins with an algorithm that, based on the user’s desired properties, autonomously identifies a handful of promising polymer blends.

The researchers originally tried a machine-learning model to predict the performance of new blends, but it was difficult to make accurate predictions across the astronomically large space of possibilities. Instead, they utilized a genetic algorithm, which uses biologically inspired operations like selection and mutation to find an optimal solution.

Their system encodes the composition of a polymer blend into what is effectively a digital chromosome, which the genetic algorithm iteratively improves to identify the most promising combinations.

“This algorithm is not new, but we had to modify the algorithm to fit into our system. For instance, we had to limit the number of polymers that could be in one material to make discovery more efficient,” Wu adds.

In addition, because the search space is so large, they tuned the algorithm to balance its choice of exploration (searching for random polymers) versus exploitation (optimizing the best polymers from the last experiment).

The algorithm sends 96 polymer blends at a time to the autonomous robotic platform, which mixes the chemicals and measures the properties of each.

The experiments were focused on improving the thermal stability of enzymes by optimizing the retained enzymatic activity (REA), a measure of how stable an enzyme is after mixing with the polymer blends and being exposed to high temperatures.

These results are sent back to the algorithm, which uses them to generate a new set of polymers until the system finds the optimal blend.

Accelerating discovery

Building the robotic system involved numerous challenges, such as developing a technique to evenly heat polymers and optimizing the speed at which the pipette tip moves up and down.

“In autonomous discovery platforms, we emphasize algorithmic innovations, but there are many detailed and subtle aspects of the procedure you have to validate before you can trust the information coming out of it,” Coley says.

When tested, the optimal blends their system identified often outperformed the polymers that formed them. The best overall blend performed 18 percent better than any of its individual components, achieving an REA of 73 percent.

“This indicates that, instead of developing new polymers, we could sometimes blend existing polymers to design new materials that perform even better than individual polymers do,” Wu says.

Moreover, their autonomous platform can generate and test 700 new polymer blends per day and only requires human intervention for refilling and replacing chemicals.

While this research focused on polymers for protein stabilization, their platform could be modified for other uses, like the development or new plastics or battery electrolytes.

In addition to exploring additional polymer properties, the researchers want to use experimental data to improve the efficiency of their algorithm and develop new algorithms to streamline the operations of the autonomous liquid handler.

“Technologically, there are urgent needs to enhance thermal stability of proteins and enzymes. The results demonstrated here are quite impressive. Being a platform technology and given the rapid advancement in machine learning and AI for material science, one can envision the possibility for this team to further enhance random heteropolymer performances or to optimize design based on end needs and usages,” says Ting Xu, a professor at the University of California at Berkeley, who was not involved with this work.

This work is funded, in part, by the U.S. Department of Energy, the National Science Foundation, and the Class of 1947 Career Development Chair.

MIT physicists have performed an idealized version of one of the most famous experiments in quantum physics. Their findings demonstrate, with atomic-level precision, the dual yet evasive nature of light. They also happen to confirm that Albert Einstein was wrong about this particular quantum scenario.

The experiment in question is the double-slit experiment, which was first performed in 1801 by the British scholar Thomas Young to show how light behaves as a wave. Today, with the formulation of quantum mechanics, the double-slit experiment is now known for its surprisingly simple demonstration of a head-scratching reality: that light exists as both a particle and a wave. Stranger still, this duality cannot be simultaneously observed. Seeing light in the form of particles instantly obscures its wave-like nature, and vice versa.

The original experiment involved shining a beam of light through two parallel slits in a screen and observing the pattern that formed on a second, faraway screen. One might expect to see two overlapping spots of light, which would imply that light exists as particles, a.k.a. photons, like paintballs that follow a direct path. But instead, the light produces alternating bright and dark stripes on the screen, in an interference pattern similar to what happens when two ripples in a pond meet. This suggests light behaves as a wave. Even weirder, when one tries to measure which slit the light is traveling through, the light suddenly behaves as particles and the interference pattern disappears.

The double-slit experiment is taught today in most high school physics classes as a simple way to illustrate the fundamental principle of quantum mechanics: that all physical objects, including light, are simultaneously particles and waves.

Nearly a century ago, the experiment was at the center of a friendly debate between physicists Albert Einstein and Niels Bohr. In 1927, Einstein argued that a photon particle should pass through just one of the two slits and in the process generate a slight force on that slit, like a bird rustling a leaf as it flies by. He proposed that one could detect such a force while also observing an interference pattern, thereby catching light’s particle and wave nature at the same time. In response, Bohr applied the quantum mechanical uncertainty principle and showed that the detection of the photon’s path would wash out the interference pattern.

Scientists have since carried out multiple versions of the double-slit experiment, and they have all, to various degrees, confirmed the validity of the quantum theory formulated by Bohr. Now, MIT physicists have performed the most “idealized” version of the double-slit experiment to date. Their version strips down the experiment to its quantum essentials. They used individual atoms as slits, and used weak beams of light so that each atom scattered at most one photon. By preparing the atoms in different quantum states, they were able to modify what information the atoms obtained about the path of the photons. The researchers thus confirmed the predictions of quantum theory: The more information was obtained about the path (i.e. the particle nature) of light, the lower the visibility of the interference pattern was. 

They demonstrated what Einstein got wrong. Whenever an atom is “rustled” by a passing photon, the wave interference is diminished.

“Einstein and Bohr would have never thought that this is possible, to perform such an experiment with single atoms and single photons,” says Wolfgang Ketterle, the John D. MacArthur Professor of Physics and leader of the MIT team. “What we have done is an idealized Gedanken experiment.”

Their results appear in the journal Physical Review Letters. Ketterle’s MIT co-authors include first author Vitaly Fedoseev, Hanzhen Lin, Yu-Kun Lu, Yoo Kyung Lee, and Jiahao Lyu, who all are affiliated with MIT’s Department of Physics, the Research Laboratory of Electronics, and the MIT-Harvard Center for Ultracold Atoms.

Cold confinement

Ketterle’s group at MIT experiments with atoms and molecules that they super-cool to temperatures just above absolute zero and arrange in configurations that they confine with laser light. Within these ultracold, carefully tuned clouds, exotic phenomena that only occur at the quantum, single-atom scale can emerge.

In a recent experiment, the team was investigating a seemingly unrelated question, studying how light scattering can reveal the properties of materials built from ultracold atoms.

“We realized we can quantify the degree to which this scattering process is like a particle or a wave, and we quickly realized we can apply this new method to realize this famous experiment in a very idealized way,” Fedoseev says.

In their new study, the team worked with more than 10,000 atoms, which they cooled to microkelvin temperatures. They used an array of laser beams to arrange the frozen atoms into an evenly spaced, crystal-like lattice configuration. In this arrangement, each atom is far enough away from any other atom that each can effectively be considered a single, isolated and identical atom. And 10,000 such atoms can produce a signal that is more easily detected, compared to a single atom or two.

The group reasoned that with this arrangement, they might shine a weak beam of light through the atoms and observe how a single photon scatters off two adjacent atoms, as a wave or a particle. This would be similar to how, in the original double-slit experiment, light passes through two slits.

“What we have done can be regarded as a new variant to the double-slit experiment,” Ketterle says. “These single atoms are like the smallest slits you could possibly build.”

Tuning fuzz

Working at the level of single photons required repeating the experiment many times and using an ultrasensitive detector to record the pattern of light scattered off the atoms. From the intensity of the detected light, the researchers could directly infer whether the light behaved as a particle or a wave.

They were particularly interested in the situation where half the photons they sent in behaved as waves, and half behaved as particles. They achieved this by using a method to tune the probability that a photon will appear as a wave versus a particle, by adjusting an atom’s “fuzziness,” or the certainty of its location. In their experiment, each of the 10,000 atoms is held in place by laser light that can be adjusted to tighten or loosen the light’s hold. The more loosely an atom is held, the fuzzier, or more “spatially extensive,” it appears. The fuzzier atom rustles more easily and records the path of the photon. Therefore, in tuning up an atom’s fuzziness, researchers can increase the probability that a photon will exhibit particle-like behavior. Their observations were in full agreement with the theoretical description.

Springs away

In their experiment, the group tested Einstein’s idea about how to detect the path of the photon. Conceptually, if each slit were cut into an extremely thin sheet of paper that was suspended in the air by a spring, a photon passing through one slit should shake the corresponding spring by a certain degree that would be a signal of the photon’s particle nature. In previous realizations of the double slit experiment, physicists have incorporated such a spring-like ingredient, and the spring played a major role in describing the photon’s dual nature.

But Ketterle and his colleagues were able to perform the experiment without the proverbial springs. The team’s cloud of atoms is initially held in place by laser light, similar to Einstein’s conception of a slit suspended by a spring. The researchers reasoned that if they were to do away with their “spring,” and observe exactly the same phenomenon, then it would show that the spring has no effect on a photon’s wave/particle duality.

This, too, was what they found. Over multiple runs, they turned off the spring-like laser holding the atoms in place and then quickly took a measurement in a millionth of a second,  before the atoms became more fuzzy and eventually fell down due to gravity. In this tiny amount of time, the atoms were effectively floating in free space. In this spring-free scenario, the team observed the same phenomenon: A photon’s wave and particle nature could not be observed simultaneously.

“In many descriptions, the springs play a major role. But we show, no, the springs do not matter here; what matters is only the fuzziness of the atoms,” Fedoseev says. “Therefore,  one has to use a more profound description, which uses quantum correlations between photons and atoms.”

The researchers note that the year 2025 has been declared by the United Nations as the International Year of Quantum Science and Technology, celebrating the formulation of quantum mechanics 100 years ago. The discussion between Bohr and Einstein about the double-slit experiment took place only two years later.

“It’s a wonderful coincidence that we could help clarify this historic controversy in the same year we celebrate quantum physics,” says co-author Lee.

This work was supported, in part, by the National Science Foundation, the U.S. Department of Defense, and the Gordon and Betty Moore Foundation.

In an office at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), a soft robotic hand carefully curls its fingers to grasp a small object. The intriguing part isn’t the mechanical design or embedded sensors — in fact, the hand contains none. Instead, the entire system relies on a single camera that watches the robot’s movements and uses that visual data to control it.

This capability comes from a new system CSAIL scientists developed, offering a different perspective on robotic control. Rather than using hand-designed models or complex sensor arrays, it allows robots to learn how their bodies respond to control commands, solely through vision. The approach, called Neural Jacobian Fields (NJF), gives robots a kind of bodily self-awareness. An open-access paper about the work was published in Nature on June 25.

“This work points to a shift from programming robots to teaching robots,” says Sizhe Lester Li, MIT PhD student in electrical engineering and computer science, CSAIL affiliate, and lead researcher on the work. “Today, many robotics tasks require extensive engineering and coding. In the future, we envision showing a robot what to do, and letting it learn how to achieve the goal autonomously.”

The motivation stems from a simple but powerful reframing: The main barrier to affordable, flexible robotics isn't hardware — it’s control of capability, which could be achieved in multiple ways. Traditional robots are built to be rigid and sensor-rich, making it easier to construct a digital twin, a precise mathematical replica used for control. But when a robot is soft, deformable, or irregularly shaped, those assumptions fall apart. Rather than forcing robots to match our models, NJF flips the script — giving robots the ability to learn their own internal model from observation.

Look and learn

This decoupling of modeling and hardware design could significantly expand the design space for robotics. In soft and bio-inspired robots, designers often embed sensors or reinforce parts of the structure just to make modeling feasible. NJF lifts that constraint. The system doesn’t need onboard sensors or design tweaks to make control possible. Designers are freer to explore unconventional, unconstrained morphologies without worrying about whether they’ll be able to model or control them later.

“Think about how you learn to control your fingers: you wiggle, you observe, you adapt,” says Li. “That’s what our system does. It experiments with random actions and figures out which controls move which parts of the robot.”

The system has proven robust across a range of robot types. The team tested NJF on a pneumatic soft robotic hand capable of pinching and grasping, a rigid Allegro hand, a 3D-printed robotic arm, and even a rotating platform with no embedded sensors. In every case, the system learned both the robot’s shape and how it responded to control signals, just from vision and random motion.

The researchers see potential far beyond the lab. Robots equipped with NJF could one day perform agricultural tasks with centimeter-level localization accuracy, operate on construction sites without elaborate sensor arrays, or navigate dynamic environments where traditional methods break down.

At the core of NJF is a neural network that captures two intertwined aspects of a robot’s embodiment: its three-dimensional geometry and its sensitivity to control inputs. The system builds on neural radiance fields (NeRF), a technique that reconstructs 3D scenes from images by mapping spatial coordinates to color and density values. NJF extends this approach by learning not only the robot’s shape, but also a Jacobian field, a function that predicts how any point on the robot’s body moves in response to motor commands.

To train the model, the robot performs random motions while multiple cameras record the outcomes. No human supervision or prior knowledge of the robot’s structure is required — the system simply infers the relationship between control signals and motion by watching.

Once training is complete, the robot only needs a single monocular camera for real-time closed-loop control, running at about 12 Hertz. This allows it to continuously observe itself, plan, and act responsively. That speed makes NJF more viable than many physics-based simulators for soft robots, which are often too computationally intensive for real-time use.

In early simulations, even simple 2D fingers and sliders were able to learn this mapping using just a few examples. By modeling how specific points deform or shift in response to action, NJF builds a dense map of controllability. That internal model allows it to generalize motion across the robot’s body, even when the data are noisy or incomplete.

“What’s really interesting is that the system figures out on its own which motors control which parts of the robot,” says Li. “This isn’t programmed — it emerges naturally through learning, much like a person discovering the buttons on a new device.”

The future is soft

For decades, robotics has favored rigid, easily modeled machines — like the industrial arms found in factories — because their properties simplify control. But the field has been moving toward soft, bio-inspired robots that can adapt to the real world more fluidly. The trade-off? These robots are harder to model.

“Robotics today often feels out of reach because of costly sensors and complex programming. Our goal with Neural Jacobian Fields is to lower the barrier, making robotics affordable, adaptable, and accessible to more people. Vision is a resilient, reliable sensor,” says senior author and MIT Assistant Professor Vincent Sitzmann, who leads the Scene Representation group. “It opens the door to robots that can operate in messy, unstructured environments, from farms to construction sites, without expensive infrastructure.”

“Vision alone can provide the cues needed for localization and control — eliminating the need for GPS, external tracking systems, or complex onboard sensors. This opens the door to robust, adaptive behavior in unstructured environments, from drones navigating indoors or underground without maps to mobile manipulators working in cluttered homes or warehouses, and even legged robots traversing uneven terrain,” says co-author Daniela Rus, MIT professor of electrical engineering and computer science and director of CSAIL. “By learning from visual feedback, these systems develop internal models of their own motion and dynamics, enabling flexible, self-supervised operation where traditional localization methods would fail.”

While training NJF currently requires multiple cameras and must be redone for each robot, the researchers are already imagining a more accessible version. In the future, hobbyists could record a robot’s random movements with their phone, much like you’d take a video of a rental car before driving off, and use that footage to create a control model, with no prior knowledge or special equipment required.

The system doesn’t yet generalize across different robots, and it lacks force or tactile sensing, limiting its effectiveness on contact-rich tasks. But the team is exploring new ways to address these limitations: improving generalization, handling occlusions, and extending the model’s ability to reason over longer spatial and temporal horizons.

“Just as humans develop an intuitive understanding of how their bodies move and respond to commands, NJF gives robots that kind of embodied self-awareness through vision alone,” says Li. “This understanding is a foundation for flexible manipulation and control in real-world environments. Our work, essentially, reflects a broader trend in robotics: moving away from manually programming detailed models toward teaching robots through observation and interaction.”

This paper brought together the computer vision and self-supervised learning work from the Sitzmann lab and the expertise in soft robots from the Rus lab. Li, Sitzmann, and Rus co-authored the paper with CSAIL affiliates Annan Zhang SM ’22, a PhD student in electrical engineering and computer science (EECS); Boyuan Chen, a PhD student in EECS; Hanna Matusik, an undergraduate researcher in mechanical engineering; and Chao Liu, a postdoc in the Senseable City Lab at MIT. 

The research was supported by the Solomon Buchsbaum Research Fund through MIT’s Research Support Committee, an MIT Presidential Fellowship, the National Science Foundation, and the Gwangju Institute of Science and Technology.

City life is often described as “fast-paced.” A new study suggests that’s more true than ever.

The research, co-authored by MIT scholars, shows that the average walking speed of pedestrians in three northeastern U.S. cities increased 15 percent from 1980 to 2010. The number of people lingering in public spaces declined by 14 percent in that time as well.

The researchers used machine-learning tools to assess 1980s-era video footage captured by renowned urbanist William Whyte, in Boston, New York, and Philadelphia. They compared the old material with newer videos from the same locations.

“Something has changed over the past 40 years,” says MIT professor of the practice Carlo Ratti, a co-author of the new study. “How fast we walk, how people meet in public space — what we’re seeing here is that public spaces are working in somewhat different ways, more as a thoroughfare and less a space of encounter.”

The paper, “Exploring the social life of urban spaces through AI,” is published this week in the Proceedings of the National Academy of Sciences. The co-authors are Arianna Salazar-Miranda MCP ’16, PhD ’23, an assistant professor at Yale University’s School of the Environment; Zhuanguan Fan of the University of Hong Kong; Michael Baick; Keith N. Hampton, a professor at Michigan State University; Fabio Duarte, associate director of the Senseable City Lab; Becky P.Y. Loo of the University of Hong Kong; Edward Glaeser, the Fred and Eleanor Glimp Professor of Economics at Harvard University; and Ratti, who is also director of MIT’s Senseable City Lab.

The results could help inform urban planning, as designers seek to create new public areas or modify existing ones.

“Public space is such an important element of civic life, and today partly because it counteracts the polarization of digital space,” says Salazar-Miranda. “The more we can keep improving public space, the more we can make our cities suited for convening.”

Meet you at the Met

Whyte was a prominent social thinker whose famous 1956 book, “The Organization Man,” probing the apparent culture of corporate conformity in the U.S., became a touchstone of its decade.

However, Whyte spent the latter decades of his career focused on urbanism. The footage he filmed, from 1978 through 1980, was archived by a Brooklyn-based nonprofit organization called the Project for Public Spaces and later digitized by Hampton and his students.

Whyte chose to make his recording at four spots in the three cities combined: Boston’s Downtown Crossing area; New York City’s Bryant Park; the steps of the Metropolitan Museum of Art in New York, a famous gathering point and people-watching spot; and Philadelphia’s Chestnut Street.

In 2010, a group led by Hampton then shot new footage at those locations, at the same times of day Whyte had, to compare and contrast current-day dynamics with those of Whyte’s time. To conduct the study, the co-authors used computer vision and AI models to summarize and quantify the activity in the videos.

The researchers have found that some things have not changed greatly. The percentage of people walking alone barely moved, from 67 percent in 1980 to 68 percent in 2010. On the other hand, the percentage of individuals entering these public spaces who became part of a group declined a bit. In 1980, 5.5 percent of the people approaching these spots met up with a group; in 2010, that was down to 2 percent.

“Perhaps there’s a more transactional nature to public space today,” Ratti says.

Fewer outdoor groups: Anomie or Starbucks?

If people’s behavioral patterns have altered since 1980, it’s natural to ask why. Certainly some of the visible changes seem consistent with the pervasive use of cellphones; people organize their social lives by phone now, and perhaps zip around more quickly from place to place as a result.

“When you look at the footage from William Whyte, the people in public spaces were looking at each other more,” Ratti says. “It was a place you could start a conversation or run into a friend. You couldn’t do things online then. Today, behavior is more predicated on texting first, to meet in public space.”

As the scholars note, if groups of people hang out together slightly less often in public spaces, there could be still another reason for that: Starbucks and its competitors. As the paper states, outdoor group socializing may be less common due to “the proliferation of coffee shops and other indoor venues. Instead of lingering on sidewalks, people may have moved their social interactions into air-conditioned, more comfortable private spaces.”

Certainly coffeeshops were far less common in big cities in 1980, and the big chain coffeeshops did not exist.

On the other hand, public-space behavior might have been evolving all this time regardless of Starbucks and the like. The researchers say the new study offers a proof-of-concept for its method and has encouraged them to conduct additional work. Ratti, Duarte, and other researchers from MIT’s Senseable City Lab have turned their attention to an extensive survey of European public spaces in an attempt to shed more light on the interaction between people and the public forum.

“We are collecting footage from 40 squares in Europe,” Duarte says. “The question is: How can we learn at a larger scale? This is in part what we’re doing.” 

One of the shared, fundamental goals of most chemistry researchers is the need to predict a molecule’s properties, such as its boiling or melting point. Once researchers can pinpoint that prediction, they’re able to move forward with their work yielding discoveries that lead to medicines, materials, and more. Historically, however, the traditional methods of unveiling these predictions are associated with a significant cost — expending time and wear and tear on equipment, in addition to funds.

Enter a branch of artificial intelligence known as machine learning (ML). ML has lessened the burden of molecule property prediction to a degree, but the advanced tools that most effectively expedite the process — by learning from existing data to make rapid predictions for new molecules — require the user to have a significant level of programming expertise. This creates an accessibility barrier for many chemists, who may not have the significant computational proficiency required to navigate the prediction pipeline. 

To alleviate this challenge, researchers in the McGuire Research Group at MIT have created ChemXploreML, a user-friendly desktop app that helps chemists make these critical predictions without requiring advanced programming skills. Freely available, easy to download, and functional on mainstream platforms, this app is also built to operate entirely offline, which helps keep research data proprietary. The exciting new technology is outlined in an article published recently in the Journal of Chemical Information and Modeling.

One specific hurdle in chemical machine learning is translating molecular structures into a numerical language that computers can understand. ChemXploreML automates this complex process with powerful, built-in "molecular embedders" that transform chemical structures into informative numerical vectors. Next, the software implements state-of-the-art algorithms to identify patterns and accurately predict molecular properties like boiling and melting points, all through an intuitive, interactive graphical interface. 

"The goal of ChemXploreML is to democratize the use of machine learning in the chemical sciences,” says Aravindh Nivas Marimuthu, a postdoc in the McGuire Group and lead author of the article. “By creating an intuitive, powerful, and offline-capable desktop application, we are putting state-of-the-art predictive modeling directly into the hands of chemists, regardless of their programming background. This work not only accelerates the search for new drugs and materials by making the screening process faster and cheaper, but its flexible design also opens doors for future innovations.” 

ChemXploreML is designed to to evolve over time, so as future techniques and algorithms are developed, they can be seamlessly integrated into the app, ensuring that researchers are always able to access and implement the most up-to-date methods. The application was tested on five key molecular properties of organic compounds — melting point, boiling point, vapor pressure, critical temperature, and critical pressure — and achieved high accuracy scores of up to 93 percent for the critical temperature. The researchers also demonstrated that a new, more compact method of representing molecules (VICGAE) was nearly as accurate as standard methods, such as Mol2Vec, but was up to 10 times faster.

“We envision a future where any researcher can easily customize and apply machine learning to solve unique challenges, from developing sustainable materials to exploring the complex chemistry of interstellar space,” says Marimuthu. Joining him on the paper is senior author and Class of 1943 Career Development Assistant Professor of Chemistry Brett McGuire.

The following press release was issued today by the Broad Institute of MIT and Harvard.

Although outbreaks of Ebola virus are rare, the disease is severe and often fatal, with few treatment options. Rather than targeting the virus itself, one promising therapeutic approach would be to interrupt proteins in the human host cell that the virus relies upon. However, finding those regulators of viral infection using existing methods has been difficult and is especially challenging for the most dangerous viruses like Ebola that require stringent high-containment biosafety protocols.

Now, researchers at the Broad Institute and the National Emerging Infectious Diseases Laboratories (NEIDL) at Boston University have used an image-based screening method developed at the Broad to identify human genes that, when silenced, impair the Ebola virus’s ability to infect. The method, known as optical pooled screening (OPS), enabled the scientists to test, in about 40 million CRISPR-perturbed human cells, how silencing each gene in the human genome affects virus replication.

Using machine-learning-based analyses of images of perturbed cells, they identified multiple host proteins involved in various stages of Ebola infection that when suppressed crippled the ability of the virus to replicate. Those viral regulators could represent avenues to one day intervene therapeutically and reduce the severity of disease in people already infected with the virus. The approach could be used to explore the role of various proteins during infection with other pathogens, as a way to find new drugs for hard-to-treat infections.

The study appears in Nature Microbiology.

“This study demonstrates the power of OPS to probe the dependency of dangerous viruses like Ebola on host factors at all stages of the viral life cycle and explore new routes to improve human health,” said co-senior author Paul Blainey, a Broad core faculty member and professor in the Department of Biological Engineering at MIT.

Previously, members of the Blainey lab developed the optical pooled screening method as a way to combine the benefits of high-content imaging, which can show a range of detailed changes in large numbers of cells at once, with those of pooled perturbational screens, which show how genetic elements influence these changes. In this study, they partnered with the laboratory of Robert Davey at BU to apply optical pooled screening to Ebola virus.

The team used CRISPR to knock out each gene in the human genome, one at a time, in nearly 40 million human cells, and then infected each cell with Ebola virus. They next fixed those cells in place in laboratory dishes and inactivated them, so that the remaining processing could occur outside of the high-containment lab.

After taking images of the cells, they measured overall viral protein and RNA in each cell using the CellProfiler image analysis software, and to get even more information from the images, they turned to AI. With help from team members in the Eric and Wendy Schmidt Center at the Broad, led by study co-author and Broad core faculty member Caroline Uhler, they used a deep learning model to automatically determine the stage of Ebola infection for each single cell. The model was able to make subtle distinctions between stages of infection in a high-throughput way that wasn’t possible using prior methods.

“The work represents the deepest dive yet into how Ebola virus rewires the cell to cause disease, and the first real glimpse into the timing of that reprogramming,” said co-senior author Robert Davey, director of the National Emerging Infectious Diseases Laboratories at Boston University, and professor of microbiology at BU Chobanian and Avedisian School of Medicine. “AI gave us an unprecedented ability to do this at scale.”

By sequencing parts of the CRISPR guide RNA in all 40 million cells individually, the researchers determined which human gene had been silenced in each cell, indicating which host proteins (and potential viral regulators) were targeted. The analysis revealed hundreds of host proteins that, when silenced, altered overall infection level, including many required for viral entry into the cell.

Knocking out other genes enhanced the amount of virus within inclusion bodies, structures that form in the human cell to act as viral factories, and prevented the infection from progressing further. Some of these human genes, such as UQCRB, pointed to a previously unrecognized role for mitochondria in the Ebola virus infection process that could possibly be exploited therapeutically. Indeed, treating cells with a small molecule inhibitor of UQCRB reduced Ebola infection with no impact on the cell’s own health.

Other genes, when silenced, altered the balance between viral RNA and protein. For example, perturbing a gene called STRAP resulted in increased viral RNA relative to protein. The researchers are currently doing further studies in the lab to better understand the role of STRAP and other proteins in Ebola infection and whether they could be targeted therapeutically.

In a series of secondary screens, the scientists examined some of the highlighted genes’ roles in infection with related filoviruses. Silencing some of these genes interrupted replication of Sudan and Marburg viruses, which have high fatality rates and no approved treatments, so it’s possible a single treatment could be effective against multiple related viruses.

The study’s approach could also be used to examine other pathogens and emerging infectious diseases and look for new ways to treat them.

“With our method, we can measure many features at once and uncover new clues about the interplay between virus and host, in a way that’s not possible through other screening approaches,” said co-first author Rebecca Carlson, a former graduate researcher in the labs of Blainey and Nir Hacohen at the Broad and who co-led the work along with co-first author J.J. Patten at Boston University.

This work was funded in part by the Broad Institute, the National Human Genome Research Institute, the Burroughs Wellcome Fund, the Fannie and John Hertz Foundation, the National Science Foundation, the George F. Carrier Postdoctoral Fellowship, the Eric and Wendy Schmidt Center at the Broad Institute, the National Institutes of Health, and the Office of Naval Research.

Astronomers at MIT, Columbia University, and elsewhere have used NASA’s James Webb Space Telescope (JWST) to peer through the dust of nearby galaxies and into the aftermath of a black hole’s stellar feast.

In a study appearing today in Astrophysical Journal Letters, the researchers report that for the first time, JWST has observed several tidal disruption events — instances when a galaxy’s central black hole draws in a nearby star and whips up tidal forces that tear the star to shreds, giving off an enormous burst of energy in the process.

Scientists have observed about 100 tidal disruption events (TDEs) since the 1990s, mostly as X-ray or optical light that flashes across relatively dust-free galaxies. But as MIT researchers recently reported, there may be many more star-shredding events in the universe that are “hiding” in dustier, gas-veiled galaxies.

In their previous work, the team found that most of the X-ray and optical light that a TDE gives off can be obscured by a galaxy’s dust, and therefore can go unseen by traditional X-ray and optical telescopes. But that same burst of light can heat up the surrounding dust and generate a new signal, in the form of infrared light.

Now, the same researchers have used JWST — the world’s most powerful infrared detector — to study signals from four dusty galaxies where they suspect tidal disruption events have occurred. Within the dust, JWST detected clear fingerprints of black hole accretion, a process by which material, such as stellar debris, circles and eventually falls into a black hole. The telescope also detected patterns that are strikingly different from the dust that surrounds active galaxies, where the central black hole is constantly pulling in surrounding material.

Together, the observations confirm that a tidal disruption event did indeed occur in each of the four galaxies. What’s more, the researchers conclude that the four events were products of not active black holes but rather dormant ones, which experienced little to no activity until a star happened to pass by.

The new results highlight JWST’s potential to study in detail otherwise hidden tidal disruption events. They are also helping scientists to reveal key differences in the environments around active versus dormant black holes.

“These are the first JWST observations of tidal disruption events, and they look nothing like what we’ve ever seen before,” says lead author Megan Masterson, a graduate student in MIT’s Kavli Institute for Astrophysics and Space Research. “We’ve learned these are indeed powered by black hole accretion, and they don’t look like environments around normal active black holes. The fact that we’re now able to study what that dormant black hole environment actually looks like is an exciting aspect.”

The study’s MIT authors include Christos Panagiotou, Erin Kara, Anna-Christina Eilers, along with Kishalay De of Columbia University and collaborators from multiple other institutions.

Seeing the light

The new study expands on the team’s previous work using another infrared detector — NASA’s Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE) mission. Using an algorithm developed by co-author Kishalay De of Columbia University, the team searched through a decade’s worth of data from the telescope, looking for infrared “transients,” or short peaks of infrared activity from otherwise quiet galaxies that could be signals of a black hole briefly waking up and feasting on a passing star. That search unearthed about a dozen signals that the group determined were likely produced by a tidal disruption event.

“With that study, we found these 12 sources that look just like TDEs,” Masterson says. “We made a lot of arguments about how the signals were very energetic, and the galaxies didn’t look like they were active before, so the signals must have been from a sudden TDE. But except for these little pieces, there was no direct evidence.”

With the much more sensitive capabilities of JWST, the researchers hoped to discern key “spectral lines,” or infrared light at specific wavelengths, that would be clear fingerprints of conditions associated with a tidal disruption event.

“With NEOWISE, it’s as if our eyes could only see red light or blue light, whereas with JWST, we’re seeing the full rainbow,” Masterson says.

A Bonafide signal

In their new work, the group looked specifically for a peak in infrared, that could only be produced by black hole accretion — a process by which material is drawn toward a black hole in a circulating disk of gas. This disk produces an enormous amount of radiation that is so intense that it can kick out electrons from individual atoms. In particular, such accretion processes can blast several electrons out from atoms of neon, and the resulting ion can transition, releasing infrared radiation at a very specific wavelength that JWST can detect. 

“There’s nothing else in the universe that can excite this gas to these energies, except for black hole accretion,” Masterson says.

The researchers searched for this smoking-gun signal in four of the 12 TDE candidates they previously identified. The four signals include: the closest tidal disruption event detected to date, located in a galaxy some 130 million light years away; a TDE that also exhibits a burst of X-ray light; a signal that may have been produced by gas circulating at incredibly high speeds around a central black hole; and a signal that also included an optical flash, which scientists had previously suspected to be a supernova, or the collapse of a dying star, rather than tidal disruption event.

“These four signals were as close as we could get to a sure thing,” Masterson says. “But the JWST data helped us say definitively these are bonafide TDEs.”

When the team pointed JWST toward the galaxies of each of the four signals, in a program designed by De, they observed that the telltale spectral lines showed up in all four sources. These measurements confirmed that black hole accretion occurred in all four galaxies. But the question remained: Was this accretion a temporary feature, triggered by a tidal disruption and a black hole that briefly woke up to feast on a passing star? Or was this accretion a more permanent trait of “active” black holes that are always on? In the case of the latter, it would be less likely that a tidal disruption event had occurred.

To differentiate between the two possibilities, the team used the JWST data to detect another wavelength of infrared light, which indicates the presence of silicates, or dust in the galaxy. They then mapped this dust in each of the four galaxies and compared the patterns to those of active galaxies, which are known to harbor clumpy, donut-shaped dust clouds around the central black hole. Masterson observed that all four sources showed very different patterns compared to typical active galaxies, suggesting that the black hole at the center of each of the galaxies is not normally active, but dormant. If an accretion disk formed around such a black hole, the researchers conclude that it must have been a result of a tidal disruption event.

“Together, these observations say the only thing these flares could be are TDEs,” Masterson says.

She and her collaborators plan to uncover many more previously hidden tidal disruption events, with NEOWISE, JWST, and other infrared telescopes. With enough detections, they say TDEs can serve as effective probes of black hole properties. For instance, how much of a star is shredded, and how fast its debris is accreted and consumed, can reveal fundamental properties of a black hole, such as how massive it is and how fast it spins.

“The actual process of a black hole gobbling down all that stellar material takes a long time,” Masterson says. “It’s not an instantaneous process. And hopefully we can start to probe how long that process takes and what that environment looks like. No one knows because we just started discovering and studying these events.”

This research was supported, in part, by NASA.

A new theory-guided framework could help scientists probe the properties of new semiconductors for next-generation microelectronic devices, or discover materials that boost the performance of quantum computers.

Research to develop new or better materials typically involves investigating properties that can be reliably measured with existing lab equipment, but this represents just a fraction of the properties that scientists could potentially probe in principle. Some properties remain effectively “invisible” because they are too difficult to capture directly with existing methods.

Take electron-phonon interaction — this property plays a critical role in a material’s electrical, thermal, optical, and superconducting properties, but directly capturing it using existing techniques is notoriously challenging.

Now, MIT researchers have proposed a theoretically justified approach that could turn this challenge into an opportunity. Their method reinterprets neutron scattering, an often-overlooked interference effect as a potential direct probe of electron-phonon coupling strength.

The procedure creates two interaction effects in the material. The researchers show that, by deliberately designing their experiment to leverage the interference between the two interactions, they can capture the strength of a material’s electron-phonon interaction.

The researchers’ theory-informed methodology could be used to shape the design of future experiments, opening the door to measuring new quantities that were previously out of reach.

“Rather than discovering new spectroscopy techniques by pure accident, we can use theory to justify and inform the design of our experiments and our physical equipment,” says Mingda Li, the Class of 1947 Career Development Professor and an associate professor of nuclear science and engineering, and senior author of a paper on this experimental method.

Li is joined on the paper by co-lead authors Chuliang Fu, an MIT postdoc; Phum Siriviboon and Artittaya Boonkird, both MIT graduate students; as well as others at MIT, the National Institute of Standards and Technology, the University of California at Riverside, Michigan State University, and Oak Ridge National Laboratory. The research appears this week in Materials Today Physics.

Investigating interference

Neutron scattering is a powerful measurement technique that involves aiming a beam of neutrons at a material and studying how the neutrons are scattered after they strike it. The method is ideal for measuring a material’s atomic structure and magnetic properties.

When neutrons collide with the material sample, they interact with it through two different mechanisms, creating a nuclear interaction and a magnetic interaction. These interactions can interfere with each other.

“The scientific community has known about this interference effect for a long time, but researchers tend to view it as a complication that can obscure measurement signals. So it hasn’t received much focused attention,” Fu says.

The team and their collaborators took a conceptual “leap of faith” and decided to explore this oft-overlooked interference effect more deeply.

They flipped the traditional materials research approach on its head by starting with a multifaceted theoretical analysis. They explored what happens inside a material when the nuclear interaction and magnetic interaction interfere with each other.

Their analysis revealed that this interference pattern is directly proportional to the strength of the material’s electron-phonon interaction.

“This makes the interference effect a probe we can use to detect this interaction,” explains Siriviboon.

Electron-phonon interactions play a role in a wide range of material properties. They affect how heat flows through a material, impact a material’s ability to absorb and emit light, and can even lead to superconductivity.

But the complexity of these interactions makes them hard to directly measure using existing experimental techniques. Instead, researchers often rely on less precise, indirect methods to capture electron-phonon interactions.

However, leveraging this interference effect enables direct measurement of the electron-phonon interaction, a major advantage over other approaches.

“Being able to directly measure the electron-phonon interaction opens the door to many new possibilities,” says Boonkird.

Rethinking materials research

Based on their theoretical insights, the researchers designed an experimental setup to demonstrate their approach.

Since the available equipment wasn’t powerful enough for this type of neutron scattering experiment, they were only able to capture a weak electron-phonon interaction signal — but the results were clear enough to support their theory.

“These results justify the need for a new facility where the equipment might be 100 to 1,000 times more powerful, enabling scientists to clearly resolve the signal and measure the interaction,” adds Landry.

With improved neutron scattering facilities, like those proposed for the upcoming Second Target Station at Oak Ridge National Laboratory, this experimental method could be an effective technique for measuring many crucial material properties.

For instance, by helping scientists identify and harness better semiconductors, this approach could enable more energy-efficient appliances, faster wireless communication devices, and more reliable medical equipment like pacemakers and MRI scanners.   

Ultimately, the team sees this work as a broader message about the need to rethink the materials research process.

“Using theoretical insights to design experimental setups in advance can help us redefine the properties we can measure,” Fu says.

To that end, the team and their collaborators are currently exploring other types of interactions they could leverage to investigate additional material properties.

“This is a very interesting paper,” says Jon Taylor, director of the neutron scattering division at Oak Ridge National Laboratory, who was not involved with this research. “It would be interesting to have a neutron scattering method that is directly sensitive to charge lattice interactions or more generally electronic effects that were not just magnetic moments. It seems that such an effect is expectedly rather small, so facilities like STS could really help develop that fundamental understanding of the interaction and also leverage such effects routinely for research.”

This work is funded, in part, by the U.S. Department of Energy and the National Science Foundation.

Inspired by a hitchhiking fish that uses a specialized suction organ to latch onto sharks and other marine animals, researchers from MIT and other institutions have designed a mechanical adhesive device that can attach to soft surfaces underwater or in extreme conditions, and remain there for days or weeks.

This device, the researchers showed, can adhere to the lining of the GI tract, whose mucosal layer makes it very difficult to attach any kind of sensor or drug-delivery capsule. Using their new adhesive system, the researchers showed that they could achieve automatic self-adhesion, without motors, to deliver HIV antiviral drugs or RNA to the GI tract, and they could also deploy it as a sensor for gastroesophageal reflux disease (GERD). The device can also be attached to a swimming fish to monitor aquatic environments.

The design is based on the research team’s extensive studies of the remora’s sucker-like disc. These discs have several unique properties that allow them to adhere tightly to a variety of hosts, including sharks, marlins, and rays. However, how remoras maintain adhesion to soft, dynamically shifting surfaces remains largely unknown.

Understanding the fundamental physics and mechanics of how this part of the fish sticks to another organism helped us to establish the underpinnings of how to engineer a synthetic adhesive system,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, an associate member of the Broad Institute of MIT and Harvard, and the senior author of the study.

MIT research scientist Ziliang (Troy) Kang is the lead author of the study, which appears today in Nature. The research team also includes authors from Brigham and Women’s Hospital, the Broad Institute, and Boston College.

Inspired by nature

Most protein and RNA drugs can’t be taken orally because they will be broken down before they can be absorbed into the GI tract. To overcome that, Traverso’s lab is working on ingestible devices that can be swallowed and then gradually release their payload over days, weeks, or even longer.

One major obstacle is that the digestive tract is lined with a slippery mucosal membrane that is constantly regenerating and is difficult for any device to stick to. Furthermore, any device that manages to attach to this lining is likely to be dislodged by food or liquids moving through the tract.

To find a solution to these challenges, the MIT team looked to the remora, also known as the sucker fish, which clings to its hosts for free transportation and access to food scraps. To explore how the remora attaches itself to dynamic, soft surfaces so strongly, Traverso’s teamed up with Christopher Kenaley, an associate professor of biology at Boston College who studies remoras and other fish.

Their studies revealed that the remora’s ability to stick to its host depends on a few different features. First, the large suction disc creates adhesion through pressure-based suction, just like a plunger. Additionally, each disc is divided into individual small adhesive compartments by rows of plates called lamellae wrapped in soft tissue. These compartments can independently create additional suction on nonhomogeneous soft surfaces.

There are nine species of remora, and in each one, these rows of lamellae are aligned a little bit differently — some are exclusively parallel, while others form patterns with rows tilted at different angles. These differences, the researchers found, could be the key to elucidating each species’ evolutionary adaptation to its host.

Remora albescens, a unique species that exhibits mucoadhesion in the oral cavity of rays, inspired the team to develop devices with enhanced adhesion to soft surfaces with its unparallel, highly tilted lamellae orientation. Other remora species, which attach to high-speed swimmers such as marlins and swordfish, tend to have highly parallel orientations, which help the hitchhikers slide without losing adhesion as they are rapidly dragged through the water. Still other species, which have a mix of parallel and angled rows, can attach to a variety of hosts.

Tiny spines that protrude from the lamellae help to achieve additional adhesion by interlocking with the host tissue. These spines, also called spinules, are several hundred microns long and grasp onto the tissue with minimal invasiveness.

“If the compartment suction is subjected to a shear force, the friction enabled by the mechanical interlocking of the spinules can help to maintain the suction,” Kang says.

Watery environments

By mimicking these anatomical features, the MIT team was able to create a device with similarly strong adhesion for a variety of applications in underwater environments.

The researchers used silicone rubber and temperature-responsive smart materials to create their adhesive device, which they call MUSAS (for “mechanical underwater soft adhesion system”). The fully passive, disc-shaped device contains rows of lamellae similar to those of the remora, and can self-adhere to the mucosal lining, leveraging GI contractions. The researchers found that for their purposes, a pattern of tilted rows was the most effective.

Within the lamellae are tiny microneedle-like structures that mimic the spinules seen in the remora. These tiny spines are made of a shape memory alloy that is activated when exposed to body temperatures, allowing the spines to interlock with each other and grasp onto the tissue surface.

The researchers showed that this device could attach to a variety of soft surfaces, even in wet or highly acidic conditions, including pig stomach tissue, nitrile gloves, and a tilapia swimming in a fish tank. Then, they tested the device for several different applications, including aquatic environmental monitoring. After adding a temperature sensor to the device, the researchers showed that they could attach the device to a fish and accurately measure water temperature as the fish swam at high speed.

To demonstrate medical applications, the researchers incorporated an impedance sensor into the device and showed that it could adhere to the esophagus in an animal model, which allowed them to monitor reflux of gastric fluid. This could offer an alternative to current sensors for GERD, which are delivered by a tube placed through the nose or mouth and pinned to the lower part of the esophagus.

They also showed that the device could be used for sustained release of two different types of therapeutics, in animal tests. First, they showed that they could integrate an HIV drug called cabotegravir into the materials that make up the device (polycaprolactone and silicone). Once adhered to the lining of the stomach, the drug gradually diffused out of the device, over a period of one week.

Cabotegravir is one of the drugs used for HIV PrEP — pre-exposure prophylaxis — as well as treatment of HIV. These treatments are usually given either as a daily pill or an injection administered every one to two months.

The researchers also created a version of the device that could be used for delivery of larger molecules such as RNA. For this kind of delivery, the researchers incorporated RNA into the microneedles of the lamellae, which could then inject them into the lining of the stomach. Using RNA encoding the gene for luciferase, a protein that emits light, the researchers showed that they could successfully deliver the gene to cells of the cheek or the esophagus.

The researchers now plan to adapt the device for delivering other types of drugs, as well as vaccines. Another possible application is using the devices for electrical stimulation, which Traverso’s lab has previously shown can activate hormones that regulate appetite.

The research was funded, in part, by the Gates Foundation, MIT’s Department of Mechanical Engineering, Brigham and Women’s Hospital, and the Advanced Research Projects Agency for Health.

AI image generation — which relies on neural networks to create new images from a variety of inputs, including text prompts — is projected to become a billion-dollar industry by the end of this decade. Even with today’s technology, if you wanted to make a fanciful picture of, say, a friend planting a flag on Mars or heedlessly flying into a black hole, it could take less than a second. However, before they can perform tasks like that, image generators are commonly trained on massive datasets containing millions of images that are often paired with associated text. Training these generative models can be an arduous chore that takes weeks or months, consuming vast computational resources in the process.

But what if it were possible to generate images through AI methods without using a generator at all? That real possibility, along with other intriguing ideas, was described in a research paper presented at the International Conference on Machine Learning (ICML 2025), which was held in Vancouver, British Columbia, earlier this summer. The paper, describing novel techniques for manipulating and generating images, was written by Lukas Lao Beyer, a graduate student researcher in MIT’s Laboratory for Information and Decision Systems (LIDS); Tianhong Li, a postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL); Xinlei Chen of Facebook AI Research; Sertac Karaman, an MIT professor of aeronautics and astronautics and the director of LIDS; and Kaiming He, an MIT associate professor of electrical engineering and computer science.

This group effort had its origins in a class project for a graduate seminar on deep generative models that Lao Beyer took last fall. In conversations during the semester, it became apparent to both Lao Beyer and He, who taught the seminar, that this research had real potential, which went far beyond the confines of a typical homework assignment. Other collaborators were soon brought into the endeavor.

The starting point for Lao Beyer’s inquiry was a June 2024 paper, written by researchers from the Technical University of Munich and the Chinese company ByteDance, which introduced a new way of representing visual information called a one-dimensional tokenizer. With this device, which is also a kind of neural network, a 256x256-pixel image can be translated into a sequence of just 32 numbers, called tokens. “I wanted to understand how such a high level of compression could be achieved, and what the tokens themselves actually represented,” says Lao Beyer.

The previous generation of tokenizers would typically break up the same image into an array of 16x16 tokens — with each token encapsulating information, in highly condensed form, that corresponds to a specific portion of the original image. The new 1D tokenizers can encode an image more efficiently, using far fewer tokens overall, and these tokens are able to capture information about the entire image, not just a single quadrant. Each of these tokens, moreover, is a 12-digit number consisting of 1s and 0s, allowing for 2¹² (or about 4,000) possibilities altogether. “It’s like a vocabulary of 4,000 words that makes up an abstract, hidden language spoken by the computer,” He explains. “It’s not like a human language, but we can still try to find out what it means.”

That’s exactly what Lao Beyer had initially set out to explore — work that provided the seed for the ICML 2025 paper. The approach he took was pretty straightforward. If you want to find out what a particular token does, Lao Beyer says, “you can just take it out, swap in some random value, and see if there is a recognizable change in the output.” Replacing one token, he found, changes the image quality, turning a low-resolution image into a high-resolution image or vice versa. Another token affected the blurriness in the background, while another still influenced the brightness. He also found a token that’s related to the “pose,” meaning that, in the image of a robin, for instance, the bird’s head might shift from right to left.

“This was a never-before-seen result, as no one had observed visually identifiable changes from manipulating tokens,” Lao Beyer says. The finding raised the possibility of a new approach to editing images. And the MIT group has shown, in fact, how this process can be streamlined and automated, so that tokens don’t have to be modified by hand, one at a time.

He and his colleagues achieved an even more consequential result involving image generation. A system capable of generating images normally requires a tokenizer, which compresses and encodes visual data, along with a generator that can combine and arrange these compact representations in order to create novel images. The MIT researchers found a way to create images without using a generator at all. Their new approach makes use of a 1D tokenizer and a so-called detokenizer (also known as a decoder), which can reconstruct an image from a string of tokens. However, with guidance provided by an off-the-shelf neural network called CLIP — which cannot generate images on its own, but can measure how well a given image matches a certain text prompt — the team was able to convert an image of a red panda, for example, into a tiger. In addition, they could create images of a tiger, or any other desired form, starting completely from scratch — from a situation in which all the tokens are initially assigned random values (and then iteratively tweaked so that the reconstructed image increasingly matches the desired text prompt).

The group demonstrated that with this same setup — relying on a tokenizer and detokenizer, but no generator — they could also do “inpainting,” which means filling in parts of images that had somehow been blotted out. Avoiding the use of a generator for certain tasks could lead to a significant reduction in computational costs because generators, as mentioned, normally require extensive training.

What might seem odd about this team’s contributions, He explains, “is that we didn’t invent anything new. We didn’t invent a 1D tokenizer, and we didn’t invent the CLIP model, either. But we did discover that new capabilities can arise when you put all these pieces together.”

“This work redefines the role of tokenizers,” comments Saining Xie, a computer scientist at New York University. “It shows that image tokenizers — tools usually used just to compress images — can actually do a lot more. The fact that a simple (but highly compressed) 1D tokenizer can handle tasks like inpainting or text-guided editing, without needing to train a full-blown generative model, is pretty surprising.”

Zhuang Liu of Princeton University agrees, saying that the work of the MIT group “shows that we can generate and manipulate the images in a way that is much easier than we previously thought. Basically, it demonstrates that image generation can be a byproduct of a very effective image compressor, potentially reducing the cost of generating images several-fold.”

There could be many applications outside the field of computer vision, Karaman suggests. “For instance, we could consider tokenizing the actions of robots or self-driving cars in the same way, which may rapidly broaden the impact of this work.”

Lao Beyer is thinking along similar lines, noting that the extreme amount of compression afforded by 1D tokenizers allows you to do “some amazing things,” which could be applied to other fields. For example, in the area of self-driving cars, which is one of his research interests, the tokens could represent, instead of images, the different routes that a vehicle might take.

Xie is also intrigued by the applications that may come from these innovative ideas. “There are some really cool use cases this could unlock,” he says. 

Let’s say you’re reading a story, or playing a game of chess. You may not have noticed, but each step of the way, your mind kept track of how the situation (or “state of the world”) was changing. You can imagine this as a sort of sequence of events list, which we use to update our prediction of what will happen next.

Language models like ChatGPT also track changes inside their own “mind” when finishing off a block of code or anticipating what you’ll write next. They typically make educated guesses using transformers — internal architectures that help the models understand sequential data — but the systems are sometimes incorrect because of flawed thinking patterns. Identifying and tweaking these underlying mechanisms helps language models become more reliable prognosticators, especially with more dynamic tasks like forecasting weather and financial markets.

But do these AI systems process developing situations like we do? A new paper from researchers in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Department of Electrical Engineering and Computer Science shows that the models instead use clever mathematical shortcuts between each progressive step in a sequence, eventually making reasonable predictions. The team made this observation by going under the hood of language models, evaluating how closely they could keep track of objects that change position rapidly. Their findings show that engineers can control when language models use particular workarounds as a way to improve the systems’ predictive capabilities.

Shell games

The researchers analyzed the inner workings of these models using a clever experiment reminiscent of a classic concentration game. Ever had to guess the final location of an object after it’s placed under a cup and shuffled with identical containers? The team used a similar test, where the model guessed the final arrangement of particular digits (also called a permutation). The models were given a starting sequence, such as “42135,” and instructions about when and where to move each digit, like moving the “4” to the third position and onward, without knowing the final result.

In these experiments, transformer-based models gradually learned to predict the correct final arrangements. Instead of shuffling the digits based on the instructions they were given, though, the systems aggregated information between successive states (or individual steps within the sequence) and calculated the final permutation.

One go-to pattern the team observed, called the “Associative Algorithm,” essentially organizes nearby steps into groups and then calculates a final guess. You can think of this process as being structured like a tree, where the initial numerical arrangement is the “root.” As you move up the tree, adjacent steps are grouped into different branches and multiplied together. At the top of the tree is the final combination of numbers, computed by multiplying each resulting sequence on the branches together.

The other way language models guessed the final permutation was through a crafty mechanism called the “Parity-Associative Algorithm,” which essentially whittles down options before grouping them. It determines whether the final arrangement is the result of an even or odd number of rearrangements of individual digits. Then, the mechanism groups adjacent sequences from different steps before multiplying them, just like the Associative Algorithm.

“These behaviors tell us that transformers perform simulation by associative scan. Instead of following state changes step-by-step, the models organize them into hierarchies,” says MIT PhD student and CSAIL affiliate Belinda Li SM ’23, a lead author on the paper. “How do we encourage transformers to learn better state tracking? Instead of imposing that these systems form inferences about data in a human-like, sequential way, perhaps we should cater to the approaches they naturally use when tracking state changes.”

“One avenue of research has been to expand test-time computing along the depth dimension, rather than the token dimension — by increasing the number of transformer layers rather than the number of chain-of-thought tokens during test-time reasoning,” adds Li. “Our work suggests that this approach would allow transformers to build deeper reasoning trees.”

Through the looking glass

Li and her co-authors observed how the Associative and Parity-Associative algorithms worked using tools that allowed them to peer inside the “mind” of language models. 

They first used a method called “probing,” which shows what information flows through an AI system. Imagine you could look into a model’s brain to see its thoughts at a specific moment — in a similar way, the technique maps out the system’s mid-experiment predictions about the final arrangement of digits.

A tool called “activation patching” was then used to show where the language model processes changes to a situation. It involves meddling with some of the system’s “ideas,” injecting incorrect information into certain parts of the network while keeping other parts constant, and seeing how the system will adjust its predictions.

These tools revealed when the algorithms would make errors and when the systems “figured out” how to correctly guess the final permutations. They observed that the Associative Algorithm learned faster than the Parity-Associative Algorithm, while also performing better on longer sequences. Li attributes the latter’s difficulties with more elaborate instructions to an over-reliance on heuristics (or rules that allow us to compute a reasonable solution fast) to predict permutations.

“We’ve found that when language models use a heuristic early on in training, they’ll start to build these tricks into their mechanisms,” says Li. “However, those models tend to generalize worse than ones that don’t rely on heuristics. We found that certain pre-training objectives can deter or encourage these patterns, so in the future, we may look to design techniques that discourage models from picking up bad habits.”

The researchers note that their experiments were done on small-scale language models fine-tuned on synthetic data, but found the model size had little effect on the results. This suggests that fine-tuning larger language models, like GPT 4.1, would likely yield similar results. The team plans to examine their hypotheses more closely by testing language models of different sizes that haven’t been fine-tuned, evaluating their performance on dynamic real-world tasks such as tracking code and following how stories evolve.

Harvard University postdoc Keyon Vafa, who was not involved in the paper, says that the researchers’ findings could create opportunities to advance language models. “Many uses of large language models rely on tracking state: anything from providing recipes to writing code to keeping track of details in a conversation,” he says. “This paper makes significant progress in understanding how language models perform these tasks. This progress provides us with interesting insights into what language models are doing and offers promising new strategies for improving them.”

Li wrote the paper with MIT undergraduate student Zifan “Carl” Guo and senior author Jacob Andreas, who is an MIT associate professor of electrical engineering and computer science and CSAIL principal investigator. Their research was supported, in part, by Open Philanthropy, the MIT Quest for Intelligence, the National Science Foundation, the Clare Boothe Luce Program for Women in STEM, and a Sloan Research Fellowship.

The researchers presented their research at the International Conference on Machine Learning (ICML) this week.

Doing your taxes can feel like a very complicated task. Even so, it might be less intricate than trying to make sense of what people think about taxes.

Several years ago, MIT political scientist Andrea Campbell undertook an expansive research project to understand public opinion about taxation. Her efforts have now reached fruition, in a new book uncovering many complexities about attitudes toward taxes. Those complexities include a central tension: In the U.S., most people say they support the principle of progressive taxation — in which higher earners pay higher shares of their income. Yet people also say they prefer specific forms of taxes that are regressive, hitting lower- and middle-income earners relatively harder.

For instance, state sales taxes are considered regressive, since people who make less money spend a larger percentage of their incomes, meaning sales taxes eat up a larger proportion of their earnings. But a substantial portion of the public still finds them to be fair, partly because the wealthy cannot wriggle out of them.

“At an abstract or conceptual level, people say they like progressive tax systems more than flat or regressive tax systems,” Campbell says. “But when you look at public attitudes toward specific taxes, people’s views flip upside down. People say federal and state income taxes are unfair, but they say sales taxes, which are very regressive, are fair. Their attitudes on individual taxes are the opposite of what their overall commitments are.”

Now Campbell analyzes these issues in detail in her book, “Taxation and Resentment,” just published by Princeton University Press. Campbell is the Arthur and Ruth Sloan Professor of Political Science at MIT and a former head of MIT’s  Department of Political Science.

Filling out the record

Campbell originally planned “Taxation and Resentment” as a strictly historically-oriented look at the subject. But the absence of any one book compiling public-opinion data in this area was striking. So, she assembled data going back to the end of World War II, and even designed and ran a couple of her own public research surveys, which help undergird the book’s numbers.

“Political scientists write a lot about public attitudes toward spending in the United States, but not so much about attitudes toward taxes,” Campbell says. “The public-opinion record is very thin.”

The complexities of U.S. public opinion on taxes are plainly linked to the presence of numerous forms of taxes, including federal and state income taxes, sales taxes, payroll taxes, estate taxes, and capital gains taxes. The best-known, of course, is the federal income tax, whose quirks and loopholes seem to irk citizens.

“That really seizes people’s imaginations,” Campbell says. “Keeping the focus on federal income tax has been a clever strategy among those who want to cut it. People think it’s unfair because they look at all the tax breaks the rich get and think, ‘I don’t have access to those.’ Those breaks increase complexity, undermine people’s knowledge, heighten their anger, and of course are in there because they help rich people pay less. So, there ends up being a cycle.”

That same sense of unfairness does not translate to all other forms of taxation, however. Large majorities of people have supported lowering the estate tax, for example, even though the threshold at which the federal estate tax kicks in — $13.5 million — applies to very few families.

Then too, the public seems to perceive sales taxes as being fair because of the simplicity and lack of loopholes — an understandable view, but one that ignores the way that state sales taxes, as opposed to state income taxes, place a bigger burden on middle-class and lower-income workers.

“A regressive tax like a sales tax is more difficult to comprehend,” Campbell says. “We all pay the same rate, so it seems like a flat tax, but as your income goes up, the bite of that tax goes down. And that’s just very difficult for people to understand.”

Overall, as Campbell details, income levels do not have huge predictive value when it comes to tax attitudes. Party affiliation also has less impact than many people might suspect — Democrats and Republicans differ on taxes, though not as much, in some ways, as political independents, who often have the most anti-tax views of all.

Meanwhile, Campbell finds, white Americans with heightened concerns about redistribution of public goods among varying demographic groups are more opposed to taxes than those who do not share those redistribution concerns. And Black and Hispanic Americans, who may wind up on the short end of regressive policies, also express significantly anti-tax perspectives, albeit while expressing more support for the state functions funded by taxation.

“There are so many factors and components of public opinion around taxes,” Campbell says. “Many political and demographic groups have their own reasons for disliking the status quo.”

How much does public opinion matter?

The research in “Taxation and Resentment” will be of high value to many kinds of scholars. However, as Campbell notes, political scientists do not have consensus about how much public opinion influences policy. Some experts contend that donors and lobbyists essentially determine policy while the larger public is ignored. But Campbell does not agree that public sentiment amounts to nothing. Consider, she says, the vigorous and successful public campaign to lower the estate tax in the first decade of the 2000s.

“If public opinion doesn’t matter, then why were there these PR campaigns to try to convince people the estate tax was bad for small businesses, farmers, and other groups?” Campbell asks. “Clearly it’s because public opinion does matter. It’s far easier to get these policies implemented if the public is on your side than if the public is in opposition. Public opinion is not the only factor in policymaking, but it’s a contributing factor.”

To be sure, even in the formation of public opinion, there are complexities and nuance, as Campbell notes in the book. A system of progressive taxation means the people taxed at the highest rate are the most motivated to oppose the system — and may heavily influence public opinion, in a top-down manner.

Scholars in the field have praised “Taxation and Resentment.” Martin Gilens, chair of the Department of Public Policy at the University of California at Los Angeles, has called it an “important and very welcome addition to the literature on public attitudes about public policies … with rich and often unexpected findings.” Vanessa Williamson, a senior fellow at the Brookings Institution, has said the book is “essential reading for anyone who wants to understand what Americans actually think about taxes. The scope of the data Campbell brings to bear on this question is unparalleled, and the depth of her analysis of public opinion across time and demography is a monumental achievement.”

For her part, Campbell says she hopes people in a variety of groups will read the book — including policymakers, scholars in multiple fields, and students. Certainly, she thinks, after studying the issue, more people could stand to know more about taxes.

“The tax system is complex,” Campbell says, “and people don’t always understand their own stakes. There is often a fog surrounding taxes.”

The MIT Health and Life Sciences Collaborative (MIT HEALS) has announced the establishment of the Fairbairn Menstruation Science Fund, supporting a bold, high-impact initiative designed to revolutionize women’s health research.

Established through a gift from Emily and Malcolm Fairbairn, the fund will advance groundbreaking research on the function of the human uterus and its impact on sex-based differences in human immunology that contribute to gynecological disorders such as endometriosis, as well as other chronic systemic inflammatory diseases that disproportionately affect women, such as Lyme disease and lupus. The Fairbairns, based in the San Francisco Bay Area, have committed $10 million, with a call to action for an additional $10 million in matching funds.

“I’m deeply grateful to Emily and Malcolm Fairbairn for their visionary support of menstruation science at MIT. For too long, this area of research has lacked broad scientific investment and visibility, despite its profound impact on the health and lives of over half the population,” says Anantha P. Chandrakasan, MIT provost who was chief innovation and strategy officer and dean of engineering at the time of the gift, and Vannevar Bush Professor of Electrical Engineering and Computer Science.

Chandrakasan adds: “Thanks to groundbreaking work from researchers like Professor Linda Griffith and her team at the MIT Center for Gynepathology Research (CGR), we have an opportunity to advance our understanding and address critical challenges in menstruation science.”

Griffith, the School of Engineering Professor of Teaching Innovation in the departments of Biological Engineering and Mechanical Engineering, and director of CGR, says the Fairbairn Fund will permit the illumination of “the enormous sex-based differences in human immunity” and advance next-generation drug-discovery technologies.

One main thrust of the new initiative will further the development of “organs on chips,” living models of patients. Using living cells or tissues, such devices allow researchers to replicate and experiment with interactions that can occur in the body. Griffith and an interdisciplinary team of researchers have engineered a powerful microfluidic platform that supports chips that foster growth of tissues complete with blood vessels and circulating immune cells. The technology was developed for building endometriosis lesions from individual patients with known clinical characteristics. The chip allows the researchers to do preclinical testing of drugs on the human patient-derived endometriosis model rather than on laboratory animals, which often do not menstruate naturally and whose immune systems function differently than that of humans.

The Fairbairn Fund will build the infrastructure for a “living patient avatar” facility to develop such physiomimetic models for all kinds of health conditions.

“We acknowledge that there are some big-picture phenomenological questions that one can study in animals, but human immunology is so very different,” Griffith says. “Pharma and biotech realize that we need living models of patients and the computational models of carefully curated patient data if we are to move into greater success in clinical trials.”

The computational models of patient data that Griffith refers to are a key element in choosing how to design the patient avatars and determine which therapeutics to test on them. For instance, by using systems biology analysis of inflammation in patient abdominal fluid, Griffith and her collaborators identified an intracellular enzyme called jun kinase (JNK). They are now working with a biotech company to test specific inhibitors of JNK in their model. Griffith has also collaborated with Michal “Mikki” Tal, a principal scientist in MIT’s Department of Biological Engineering, on investigating a possible link between prior infection, such as by the Lyme-causing bacterium Borrelia, and a number of chronic inflammatory diseases in women. Automating assays of patient samples for higher throughput could systematically speed the generation of hypotheses guiding the development of patient model experimentation.

“This fund is catalytic,” Griffith says. “Industry and government, along with other foundations, will invest if the foundational infrastructure exists. They want to employ the technologies, but it is hard to get them developed to the point they are proven to be useful. This gets us through that difficult part of the journey.”

The fund will also support public engagement efforts to reduce stigma around menstruation and neglect of such conditions as abnormal uterine bleeding and debilitating anemia, endometriosis, and polycystic ovary syndrome — and in general bring greater attention to women’s health research. Endometriosis, for instance, in which tissue that resembles the uterine lining starts growing outside the uterus and causes painful inflammation, affects one in 10 women. It often goes undiagnosed for years, and can require repeated surgeries to remove its lesions. Meanwhile, little is known about what causes it, how to prevent it, or what could effectively stop it.

Women’s health research could further advance in many areas of medicine beyond conditions that disproportionately affect females. Griffith points out that the uterus, which sheds and regenerates its lining every month, demonstrates “scarless healing” that could warrant investigation. Also, deepened study of the uterus could shed light on immune tolerance for transplants, given that in a successful pregnancy an implanted fetus is not rejected, despite containing foreign material from the biological father.

For Emily Fairbairn, the fund is a critical step toward major advances in an often-overlooked area of medicine.

“My mission is to support intellectually honest, open-minded scientists who embrace risk, treat failure as feedback, and remain committed to discovery over dogma. This fund is a direct extension of that philosophy. It’s designed to fuel research into the biological realities of diseases that remain poorly understood, frequently dismissed, or disproportionately misdiagnosed in women,” Fairbairn says. “I’ve chosen to make this gift to MIT because Linda Griffith exemplifies the rare combination of scientific integrity and bold innovation — qualities essential for tackling the most neglected challenges in medicine.”

Fairbairn also refers to Griffith collaborator Michal Tal as being “deeply inspiring.”

“Her work embodies what’s possible when scientific excellence meets institutional courage. It is this spirit — bold, rigorous, and fearless — that inspired this gift and fuels our hope for the future of women’s health,” she says.

Fairbairn, who has suffered from both Lyme disease and endometriosis that required multiple surgeries, originally directed her philanthropy, including previous gifts to MIT, toward the study of Lyme disease and associated infections.

“My own experience with both Lyme and endometriosis deepened my conviction that science must better account for how female physiology, genetics, and psychology differ from men’s,” she says. “MIT stands out for treating women’s health not as a niche, but as a frontier. The Institute’s willingness to bridge immunology, neurobiology, bioengineering, and data science — alongside its development of cutting-edge platforms like human chips — offers a rare and necessary seriousness of purpose.”

For her part, Griffith refers to Fairbairn as “a citizen scientist who inspires us daily.”

“Her tireless advocacy for patients, especially women, who are dismissed and gas-lit, is priceless,” Griffith adds. “Emily has made me a better scientist, in service of humanity.”

As countries across the world experience a resurgence in nuclear energy projects, the questions of where and how to dispose of nuclear waste remain as politically fraught as ever. The United States, for instance, has indefinitely stalled its only long-term underground nuclear waste repository. Scientists are using both modeling and experimental methods to study the effects of underground nuclear waste disposal and ultimately, they hope, build public trust in the decision-making process.

New research from scientists at MIT, Lawrence Berkeley National Lab, and the University of Orléans makes progress in that direction. The study shows that simulations of underground nuclear waste interactions, generated by new, high-performance-computing software, aligned well with experimental results from a research facility in Switzerland.

The study, which was co-authored by MIT PhD student Dauren Sarsenbayev and Assistant Professor Haruko Wainwright, along with Christophe Tournassat and Carl Steefel, appears in the journal PNAS.

“These powerful new computational tools, coupled with real-world experiments like those at the Mont Terri research site in Switzerland, help us understand how radionuclides will migrate in coupled underground systems,” says Sarsenbayev, who is first author of the new study.

The authors hope the research will improve confidence among policymakers and the public in the long-term safety of underground nuclear waste disposal.

“This research — coupling both computation and experiments — is important to improve our confidence in waste disposal safety assessments,” says Wainwright. “With nuclear energy re-emerging as a key source for tackling climate change and ensuring energy security, it is critical to validate disposal pathways.”

Comparing simulations with experiments

Disposing of nuclear waste in deep underground geological formations is currently considered the safest long-term solution for managing high-level radioactive waste. As such, much effort has been put into studying the migration behaviors of radionuclides from nuclear waste within various natural and engineered geological materials.

Since its founding in 1996, the Mont Terri research site in northern Switzerland has served as an important test bed for an international consortium of researchers interested in studying materials like Opalinus clay — a thick, water-tight claystone abundant in the tunneled areas of the mountain.

“It is widely regarded as one of the most valuable real-world experiment sites because it provides us with decades of datasets around the interactions of cement and clay, and those are the key materials proposed to be used by countries across the world for engineered barrier systems and geological repositories for nuclear waste,” explains Sarsenbayev.

For their study, Sarsenbayev and Wainwright collaborated with co-authors Tournassat and Steefel, who have developed high-performance computing software to improve modeling of interactions between the nuclear waste and both engineered and natural materials.

To date, several challenges have limited scientists’ understanding of how nuclear waste reacts with cement-clay barriers. For one thing, the barriers are made up of irregularly mixed materials deep underground. Additionally, the existing class of models commonly used to simulate radionuclide interactions with cement-clay do not take into account electrostatic effects associated with the negatively charged clay minerals in the barriers.

Tournassat and Steefel’s new software accounts for electrostatic effects, making it the only one that can simulate those interactions in three-dimensional space. The software, called CrunchODiTi, was developed from established software known as CrunchFlow and was most recently updated this year. It is designed to be run on many high-performance computers at once in parallel.

For the study, the researchers looked at a 13-year-old experiment, with an initial focus on cement-clay rock interactions. Within the last several years, a mix of both negatively and positively charged ions were added to the borehole located near the center of the cement emplaced in the formation. The researchers focused on a 1-centimeter-thick zone between the radionuclides and cement-clay referred to as the “skin.” They compared their experimental results to the software simulation, finding the two datasets aligned.

“The results are quite significant because previously, these models wouldn’t fit field data very well,” Sarsenbayev says. “It’s interesting how fine-scale phenomena at the ‘skin’ between cement and clay, the physical and chemical properties of which changes over time, could be used to reconcile the experimental and simulation data.” 

The experimental results showed the model successfully accounted for electrostatic effects associated with the clay-rich formation and the interaction between materials in Mont Terri over time.

“This is all driven by decades of work to understand what happens at these interfaces,” Sarsenbayev says. “It’s been hypothesized that there is mineral precipitation and porosity clogging at this interface, and our results strongly suggest that.”

“This application requires millions of degrees of freedom because these multibarrier systems require high resolution and a lot of computational power,” Sarsenbayev says. “This software is really ideal for the Mont Terri experiment.”

Assessing waste disposal plans

The new model could now replace older models that have been used to conduct safety and performance assessments of underground geological repositories.

“If the U.S. eventually decides to dispose nuclear waste in a geological repository, then these models could dictate the most appropriate materials to use,” Sarsenbayev says. “For instance, right now clay is considered an appropriate storage material, but salt formations are another potential medium that could be used. These models allow us to see the fate of radionuclides over millennia. We can use them to understand interactions at timespans that vary from months to years to many millions of years.”

Sarsenbayev says the model is reasonably accessible to other researchers and that future efforts may focus on the use of machine learning to develop less computationally expensive surrogate models.

Further data from the experiment will be available later this month. The team plans to compare those data to additional simulations.

“Our collaborators will basically get this block of cement and clay, and they’ll be able to run experiments to determine the exact thickness of the skin along with all of the minerals and processes present at this interface,” Sarsenbayev says. “It’s a huge project and it takes time, but we wanted to share initial data and this software as soon as we could.”

For now, the researchers hope their study leads to a long-term solution for storing nuclear waste that policymakers and the public can support.

“This is an interdisciplinary study that includes real world experiments showing we’re able to predict radionuclides’ fate in the subsurface,” Sarsenbayev says. “The motto of MIT’s Department of Nuclear Science and Engineering is ‘Science. Systems. Society.’ I think this merges all three domains.”

Teaching a robot new skills used to require coding expertise. But a new generation of robots could potentially learn from just about anyone.

Engineers are designing robotic helpers that can “learn from demonstration.” This more natural training strategy enables a person to lead a robot through a task, typically in one of three ways: via remote control, such as operating a joystick to remotely maneuver a robot; by physically moving the robot through the motions; or by performing the task themselves while the robot watches and mimics.

Learning-by-doing robots usually train in just one of these three demonstration approaches. But MIT engineers have now developed a three-in-one training interface that allows a robot to learn a task through any of the three training methods. The interface is in the form of a handheld, sensor-equipped tool that can attach to many common collaborative robotic arms. A person can use the attachment to teach a robot to carry out a task by remotely controlling the robot, physically manipulating it, or demonstrating the task themselves — whichever style they prefer or best suits the task at hand.

The MIT team tested the new tool, which they call a “versatile demonstration interface,” on a standard collaborative robotic arm. Volunteers with manufacturing expertise used the interface to perform two manual tasks that are commonly carried out on factory floors.

The researchers say the new interface offers increased training flexibility that could expand the type of users and “teachers” who interact with robots. It may also enable robots to learn a wider set of skills. For instance, a person could remotely train a robot to handle toxic substances, while further down the production line another person could physically move the robot through the motions of boxing up a product, and at the end of the line, someone else could use the attachment to draw a company logo as the robot watches and learns to do the same.

“We are trying to create highly intelligent and skilled teammates that can effectively work with humans to get complex work done,” says Mike Hagenow, a postdoc at MIT in the Department of Aeronautics and Astronautics. “We believe flexible demonstration tools can help far beyond the manufacturing floor, in other domains where we hope to see increased robot adoption, such as home or caregiving settings.”

Hagenow will present a paper detailing the new interface, at the IEEE Intelligent Robots and Systems (IROS) conference in October. The paper’s MIT co-authors are Dimosthenis Kontogiorgos, a postdoc at the MIT Computer Science and Artificial Intelligence Lab (CSAIL); Yanwei Wang PhD ’25, who recently earned a doctorate in electrical engineering and computer science; and Julie Shah, MIT professor and head of the Department of Aeronautics and Astronautics.

Training together

Shah’s group at MIT designs robots that can work alongside humans in the workplace, in hospitals, and at home. A main focus of her research is developing systems that enable people to teach robots new tasks or skills “on the job,” as it were. Such systems would, for instance, help a factory floor worker quickly and naturally adjust a robot’s maneuvers to improve its task in the moment, rather than pausing to reprogram the robot’s software from scratch — a skill that a worker may not necessarily have.

The team’s new work builds on an emerging strategy in robot learning called “learning from demonstration,” or LfD, in which robots are designed to be trained in more natural, intuitive ways. In looking through the LfD literature, Hagenow and Shah found LfD training methods developed so far fall generally into the three main categories of teleoperation, kinesthetic training, and natural teaching.

One training method may work better than the other two for a particular person or task. Shah and Hagenow wondered whether they could design a tool that combines all three methods to enable a robot to learn more tasks from more people.

“If we could bring together these three different ways someone might want to interact with a robot, it may bring benefits for different tasks and different people,” Hagenow says.

Tasks at hand

With that goal in mind, the team engineered a new versatile demonstration interface (VDI). The interface is a handheld attachment that can fit onto the arm of a typical collaborative robotic arm. The attachment is equipped with a camera and markers that track the tool’s position and movements over time, along with force sensors to measure the amount of pressure applied during a given task.

When the interface is attached to a robot, the entire robot can be controlled remotely, and the interface’s camera records the robot’s movements, which the robot can use as training data to learn the task on its own. Similarly, a person can physically move the robot through a task, with the interface attached. The VDI can also be detached and physically held by a person to perform the desired task. The camera records the VDI’s motions, which the robot can also use to mimic the task when the VBI is reattached.

To test the attachment’s usability, the team brought the interface, along with a collaborative robotic arm, to a local innovation center where manufacturing experts learn about and test technology that can improve factory-floor processes. The researchers set up an experiment where they asked volunteers at the center to use the robot and all three of the interface’s training methods to complete two common manufacturing tasks: press-fitting and molding. In press-fitting, the user trained the robot to press and fit pegs into holes, similar to many fastening tasks. For molding, a volunteer trained the robot to push and roll a rubbery, dough-like substance evenly around the surface of a center rod, similar to some thermomolding tasks.

For each of the two tasks, the volunteers were asked to use each of the three training methods, first teleoperating the robot using a joystick, then kinesthetically manipulating the robot, and finally, detaching the robot’s attachment and using it to “naturally” perform the task as the robot recorded the attachment’s force and movements.

The researchers found the volunteers generally preferred the natural method over teleoperation and kinesthetic training. The users, who were all experts in manufacturing, did offer scenarios in which each method might have advantages over the others. Teleoperation, for instance, may be preferable in training a robot to handle hazardous or toxic substances. Kinesthetic training could help workers adjust the positioning of a robot that is tasked with moving heavy packages. And natural teaching could be beneficial in demonstrating tasks that involve delicate and precise maneuvers.

“We imagine using our demonstration interface in flexible manufacturing environments where one robot might assist across a range of tasks that benefit from specific types of demonstrations,” says Hagenow, who plans to refine the attachment’s design based on user feedback and will use the new design to test robot learning. “We view this study as demonstrating how greater flexibility in collaborative robots can be achieved through interfaces that expand the ways that end-users interact with robots during teaching.”

This work was supported, in part, by the MIT Postdoctoral Fellowship Program for Engineering Excellence and the Wallenberg Foundation Postdoctoral Research Fellowship.

Large language models (LLMs) excel at using textual reasoning to understand the context of a document and provide a logical answer about its contents. But these same LLMs often struggle to correctly answer even the simplest math problems.

Textual reasoning is usually a less-than-ideal way to deliberate over computational or algorithmic tasks. While some LLMs can generate code like Python to handle symbolic queries, the models don’t always know when to use code, or what kind of code would work best.

LLMs, it seems, may need a coach to steer them toward the best technique.

Enter CodeSteer, a smart assistant developed by MIT researchers that guides an LLM to switch between code and text generation until it correctly answers a query.

CodeSteer, itself a smaller LLM, automatically generates a series of prompts to iteratively steer a larger LLM. It reviews the model’s current and previous answers after each round and provides guidance for how it can fix or refine that solution until it deems the answer is correct.

The researchers found that augmenting a larger LLM with CodeSteer boosted its accuracy on symbolic tasks, like multiplying numbers, playing Sudoku, and stacking blocks, by more than 30 percent. It also enabled less sophisticated models to outperform more advanced models with enhanced reasoning skills.

This advance could improve the problem-solving capabilities of LLMs for complex tasks that are especially difficult to solve with textual reasoning alone, such as generating paths for robots in uncertain environments or scheduling shipments in an international supply chain.

“There is a race to develop better and better models that are capable of doing everything, but we’ve taken a complementary approach. Researchers have spent years developing effective technologies and tools to tackle problems in many domains. We want to enable LLMs to select the right tools and methods, and make use of others’ expertise to enhance their own capabilities,” says Chuchu Fan, an associate professor of aeronautics and astronautics (AeroAstro) and principal investigator in the MIT Laboratory for Information and Decision Systems (LIDS).

Fan, the senior author of the study, is joined on a paper about the work by LIDS graduate student Yongchao Chen; AeroAstro graduate student Yilun Hao; University of Illinois at Urbana-Champaign graduate student Yueying Liu; and MIT-IBM Watson AI Lab Research Scientist Yang Zhang. The research will be presented at the International Conference on Machine Learning.

An LLM “trainer”  

Ask an LLM which number is bigger, 9.11 or 9.9, and it will often give the wrong answer by using textual reasoning. But ask it to use code to answer the same question, and it can generate and execute a Python script to compare the two numbers, easily solving the problem.

Initially trained to understand and predict human language, LLMs are more likely to answer queries using text, even when code would be more effective. And while they have learned to generate code through fine-tuning, these models often generate an incorrect or less efficient version of the code.

Rather than trying to retrain a powerful LLM like GPT-4 or Claude to improve these capabilities, the MIT researchers fine-tune a smaller, lightweight LLM to guide a larger model between text and code. Fine-tuning a smaller model doesn’t change the larger LLM, so there is no risk it would undermine the larger model’s other abilities.

“We were also inspired by humans. In sports, a trainer may not be better than the star athlete on the team, but the trainer can still give helpful suggestions to guide the athlete. This steering method works for LLMs, too,” Chen says.

This trainer, CodeSteer, works in conjunction with the larger LLM. It first reviews a query and determines whether text or code is suitable for this problem, and which sort of code would be best.

Then it generates a prompt for the larger LLM, telling it to use a coding method or textual reasoning to answer the query. The larger model follows this prompt to answer the query and sends the result back to CodeSteer, which reviews it.

If the answer is not correct, CodeSteer will continue prompting the LLM to try different things that might fix the problem, such as incorporating a search algorithm or constraint into its Python code, until the answer is correct.

“We found that oftentimes, the larger LLM will try to be lazy and use a shorter, less efficient code that will not carry the correct symbolic calculation. We’ve designed CodeSteer to avoid this phenomenon,” Chen says.

A symbolic checker evaluates the code’s complexity and sends a signal to CodeSteer if it is too simple or inefficient. The researchers also incorporate a self-answer checker into CodeSteer, which prompts the LLM to generate code that calculates the answer to verify it is correct.

Tackling complex tasks

As the researchers designed CodeSteer, they couldn’t find suitable symbolic datasets to fine-tune and test the model, since many existing benchmarks don’t point out whether a certain query could be best solved with text or code.

So, they gathered a corpus of 37 complex symbolic tasks, including spatial reasoning, mathematics, order reasoning, and optimization, and built their own dataset, called SymBench. They implemented a fine-tuning approach that leverages SymBench to maximize the performance of CodeSteer.

In their experiments, CodeSteer outperformed all nine baseline methods they evaluated and boosted average accuracy from 53.3 percent to 86.4 percent. It maintains similar performance even on unseen tasks, and on a variety of LLMs.

In addition, a general-purpose model augmented with CodeSteer can achieve higher accuracy than state-of-the-art models designed to focus on complex reasoning and planning, while requiring much less computation.

“Our method uses an LLM’s own capabilities. By augmenting an LLM with the ability to smartly use coding, we can take a model that is already very strong and improve its performance even more,” Chen says.

In the future, the researchers want to streamline CodeSteer to speed up its iterative prompting process. In addition, they are studying how to effectively fine-tune a unified model with the ability to switch between textual reasoning and code generation, rather than relying on a separate assistant.

“The authors present an elegant solution to the critical challenge of tool utilization in LLMs. This simple yet impactful method enables state-of-the-art LLMs to achieve significant performance improvements without requiring direct fine-tuning,” says Jinsung Yoon, a staff research scientist at Google Cloud AI, who was not involved with this work. “This research represents a substantial contribution that promises to significantly enhance the application of LLMs to a diverse range of tasks with which they currently struggle.”

“Their success in training a smaller, specialized model to strategically guide larger, advanced models is particularly impactful,” adds Chi Wang, a senior staff scientist at Google DeepMind who was not involved with this work. “This intelligent collaboration among diverse AI ‘agents’ paves the way for more robust and versatile applications in complex real-world scenarios.”

This research is supported, in part, by the U.S. Office of Naval Research and the MIT-IBM Watson AI Lab.

Imagine a future where artificial intelligence quietly shoulders the drudgery of software development: refactoring tangled code, migrating legacy systems, and hunting down race conditions, so that human engineers can devote themselves to architecture, design, and the genuinely novel problems still beyond a machine’s reach. Recent advances appear to have nudged that future tantalizingly close, but a new paper by researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and several collaborating institutions argues that this potential future reality demands a hard look at present-day challenges. 

Titled “Challenges and Paths Towards AI for Software Engineering,” the work maps the many software-engineering tasks beyond code generation, identifies current bottlenecks, and highlights research directions to overcome them, aiming to let humans focus on high-level design while routine work is automated. 

“Everyone is talking about how we don’t need programmers anymore, and there’s all this automation now available,” says Armando Solar‑Lezama, MIT professor of electrical engineering and computer science, CSAIL principal investigator, and senior author of the study. “On the one hand, the field has made tremendous progress. We have tools that are way more powerful than any we’ve seen before. But there’s also a long way to go toward really getting the full promise of automation that we would expect.”

Solar-Lezama argues that popular narratives often shrink software engineering to “the undergrad programming part: someone hands you a spec for a little function and you implement it, or solving LeetCode-style programming interviews.” Real practice is far broader. It includes everyday refactors that polish design, plus sweeping migrations that move millions of lines from COBOL to Java and reshape entire businesses. It requires nonstop testing and analysis — fuzzing, property-based testing, and other methods — to catch concurrency bugs, or patch zero-day flaws. And it involves the maintenance grind: documenting decade-old code, summarizing change histories for new teammates, and reviewing pull requests for style, performance, and security.

Industry-scale code optimization — think re-tuning GPU kernels or the relentless, multi-layered refinements behind Chrome’s V8 engine — remains stubbornly hard to evaluate. Today’s headline metrics were designed for short, self-contained problems, and while multiple-choice tests still dominate natural-language research, they were never the norm in AI-for-code. The field’s de facto yardstick, SWE-Bench, simply asks a model to patch a GitHub issue: useful, but still akin to the “undergrad programming exercise” paradigm. It touches only a few hundred lines of code, risks data leakage from public repositories, and ignores other real-world contexts — AI-assisted refactors, human–AI pair programming, or performance-critical rewrites that span millions of lines. Until benchmarks expand to capture those higher-stakes scenarios, measuring progress — and thus accelerating it — will remain an open challenge.

If measurement is one obstacle, human‑machine communication is another. First author Alex  Gu, an MIT graduate student in electrical engineering and computer science, sees today’s interaction as “a thin line of communication.” When he asks a system to generate code, he often receives a large, unstructured file and even a set of unit tests, yet those tests tend to be superficial. This gap extends to the AI’s ability to effectively use the wider suite of software engineering tools, from debuggers to static analyzers, that humans rely on for precise control and deeper understanding. “I don’t really have much control over what the model writes,” he says. “Without a channel for the AI to expose its own confidence — ‘this part’s correct … this part, maybe double‑check’ — developers risk blindly trusting hallucinated logic that compiles, but collapses in production. Another critical aspect is having the AI know when to defer to the user for clarification.” 

Scale compounds these difficulties. Current AI models struggle profoundly with large code bases, often spanning millions of lines. Foundation models learn from public GitHub, but “every company’s code base is kind of different and unique,” Gu says, making proprietary coding conventions and specification requirements fundamentally out of distribution. The result is code that looks plausible yet calls non‑existent functions, violates internal style rules, or fails continuous‑integration pipelines. This often leads to AI-generated code that “hallucinates,” meaning it creates content that looks plausible but doesn’t align with the specific internal conventions, helper functions, or architectural patterns of a given company. 

Models will also often retrieve incorrectly, because it retrieves code with a similar name (syntax) rather than functionality and logic, which is what a model might need to know how to write the function. “Standard retrieval techniques are very easily fooled by pieces of code that are doing the same thing but look different,” says Solar‑Lezama. 

The authors mention that since there is no silver bullet to these issues, they’re calling instead for community‑scale efforts: richer, having data that captures the process of developers writing code (for example, which code developers keep versus throw away, how code gets refactored over time, etc.), shared evaluation suites that measure progress on refactor quality, bug‑fix longevity, and migration correctness; and transparent tooling that lets models expose uncertainty and invite human steering rather than passive acceptance. Gu frames the agenda as a “call to action” for larger open‑source collaborations that no single lab could muster alone. Solar‑Lezama imagines incremental advances—“research results taking bites out of each one of these challenges separately”—that feed back into commercial tools and gradually move AI from autocomplete sidekick toward genuine engineering partner.

“Why does any of this matter? Software already underpins finance, transportation, health care, and the minutiae of daily life, and the human effort required to build and maintain it safely is becoming a bottleneck. An AI that can shoulder the grunt work — and do so without introducing hidden failures — would free developers to focus on creativity, strategy, and ethics” says Gu. “But that future depends on acknowledging that code completion is the easy part; the hard part is everything else. Our goal isn’t to replace programmers. It’s to amplify them. When AI can tackle the tedious and the terrifying, human engineers can finally spend their time on what only humans can do.”

“With so many new works emerging in AI for coding, and the community often chasing the latest trends, it can be hard to step back and reflect on which problems are most important to tackle,” says Baptiste Rozière, an AI scientist at Mistral AI, who wasn’t involved in the paper. “I enjoyed reading this paper because it offers a clear overview of the key tasks and challenges in AI for software engineering. It also outlines promising directions for future research in the field.”

Gu and Solar-Lezama wrote the paper with University of California at Berkeley Professor Koushik Sen and PhD students Naman Jain and Manish Shetty, Cornell University Assistant Professor Kevin Ellis and PhD student Wen-Ding Li, Stanford University Assistant Professor Diyi Yang and PhD student Yijia Shao, and incoming Johns Hopkins University assistant professor Ziyang Li. Their work was supported, in part, by the National Science Foundation (NSF), SKY Lab industrial sponsors and affiliates, Intel Corp. through an NSF grant, and the Office of Naval Research.

The researchers are presenting their work at the International Conference on Machine Learning (ICML). 

MIT researchers have developed a new theoretical framework for studying the mechanisms of treatment interactions. Their approach allows scientists to efficiently estimate how combinations of treatments will affect a group of units, such as cells, enabling a researcher to perform fewer costly experiments while gathering more accurate data.

As an example, to study how interconnected genes affect cancer cell growth, a biologist might need to use a combination of treatments to target multiple genes at once. But because there could be billions of potential combinations for each round of the experiment, choosing a subset of combinations to test might bias the data their experiment generates. 

In contrast, the new framework considers the scenario where the user can efficiently design an unbiased experiment by assigning all treatments in parallel, and can control the outcome by adjusting the rate of each treatment.

The MIT researchers theoretically proved a near-optimal strategy in this framework and performed a series of simulations to test it in a multiround experiment. Their method minimized the error rate in each instance.

This technique could someday help scientists better understand disease mechanisms and develop new medicines to treat cancer or genetic disorders.

“We’ve introduced a concept people can think more about as they study the optimal way to select combinatorial treatments at each round of an experiment. Our hope is this can someday be used to solve biologically relevant questions,” says graduate student Jiaqi Zhang, an Eric and Wendy Schmidt Center Fellow and co-lead author of a paper on this experimental design framework.

She is joined on the paper by co-lead author Divya Shyamal, an MIT undergraduate; and senior author Caroline Uhler, the Andrew and Erna Viterbi Professor of Engineering in EECS and the MIT Institute for Data, Systems, and Society (IDSS), who is also director of the Eric and Wendy Schmidt Center and a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS). The research was recently presented at the International Conference on Machine Learning.

Simultaneous treatments

Treatments can interact with each other in complex ways. For instance, a scientist trying to determine whether a certain gene contributes to a particular disease symptom may have to target several genes simultaneously to study the effects.

To do this, scientists use what are known as combinatorial perturbations, where they apply multiple treatments at once to the same group of cells.

“Combinatorial perturbations will give you a high-level network of how different genes interact, which provides an understanding of how a cell functions,” Zhang explains.

Since genetic experiments are costly and time-consuming, the scientist aims to select the best subset of treatment combinations to test, which is a steep challenge due to the huge number of possibilities.

Picking a suboptimal subset can generate biased results by focusing only on combinations the user selected in advance.

The MIT researchers approached this problem differently by looking at a probabilistic framework. Instead of focusing on a selected subset, each unit randomly takes up combinations of treatments based on user-specified dosage levels for each treatment.

The user sets dosage levels based on the goal of their experiment — perhaps this scientist wants to study the effects of four different drugs on cell growth. The probabilistic approach generates less biased data because it does not restrict the experiment to a predetermined subset of treatments.

The dosage levels are like probabilities, and each cell receives a random combination of treatments. If the user sets a high dosage, it is more likely most of the cells will take up that treatment. A smaller subset of cells will take up that treatment if the dosage is low.

“From there, the question is how do we design the dosages so that we can estimate the outcomes as accurately as possible? This is where our theory comes in,” Shyamal adds.

Their theoretical framework shows the best way to design these dosages so one can learn the most about the characteristic or trait they are studying.

After each round of the experiment, the user collects the results and feeds those back into the experimental framework. It will output the ideal dosage strategy for the next round, and so on, actively adapting the strategy over multiple rounds.

Optimizing dosages, minimizing error

The researchers proved their theoretical approach generates optimal dosages, even when the dosage levels are affected by a limited supply of treatments or when noise in the experimental outcomes varies at each round.

In simulations, this new approach had the lowest error rate when comparing estimated and actual outcomes of multiround experiments, outperforming two baseline methods.

In the future, the researchers want to enhance their experimental framework to consider interference between units and the fact that certain treatments can lead to selection bias. They would also like to apply this technique in a real experimental setting.

“This is a new approach to a very interesting problem that is hard to solve. Now, with this new framework in hand, we can think more about the best way to design experiments for many different applications,” Zhang says.

This research is funded, in part, by the Advanced Undergraduate Research Opportunities Program at MIT, Apple, the National Institutes of Health, the Office of Naval Research, the Department of Energy, the Eric and Wendy Schmidt Center at the Broad Institute, and a Simons Investigator Award.

Scientists have long known that the brain’s visual system isn’t fully hardwired from the start — it becomes refined by what babies see — but the authors of a new MIT study still weren’t prepared for the degree of rewiring they observed when they took a first-ever look at the process in mice as it happened in real-time.

As the researchers in The Picower Institute for Learning and Memory tracked hundreds of “spine” structures housing individual network connections, or “synapses,” on the dendrite branches of neurons in the visual cortex over 10 days, they saw that only 40 percent of the ones that started the process survived. Refining binocular vision (integrating input from both eyes) required numerous additions and removals of spines along the dendrites to establish an eventual set of connections.

Former graduate student Katya Tsimring led the study, published this month in Nature Communications, which the team says is the first in which scientists tracked the same connections all the way through the “critical period,” when binocular vision becomes refined.

“What Katya was able to do is to image the same dendrites on the same neurons repeatedly over 10 days in the same live mouse through a critical period of development, to ask, what happens to the synapses or spines on them?,” says senior author Mriganka Sur, the Paul and Lilah Newton Professor in the Picower Institute and MIT’s Department of Brain and Cognitive Sciences. “We were surprised by how much change there is.”

Extensive turnover

In the experiments, young mice watched as black-and-white gratings with lines of specific orientations and directions of movement drifted across their field of view. At the same time, the scientists observed both the structure and activity of the neurons’ main body (or “soma”) and of the spines along their dendrites. By tracking the structure of 793 dendritic spines on 14 neurons at roughly Day 1, Day 5 and Day 10 of the critical period, they could quantify the addition and loss of the spines, and therefore the synaptic connections they housed. And by tracking their activity at the same time, they could quantify the visual information the neurons received at each synaptic connection. For example, a spine might respond to one specific orientation or direction of grating, several orientations, or might not respond at all. Finally, by relating a spine’s structural changes across the critical period to its activity, they sought to uncover the process by which synaptic turnover refined binocular vision.

Structurally, the researchers saw that 32 percent of the spines evident on Day 1 were gone by Day 5, and that 24 percent of the spines apparent on Day 5 had been added since Day 1. The period between Day 5 and Day 10 showed similar turnover: 27 percent were eliminated, but 24 percent were added. Overall, only 40 percent of the spines seen on Day 1 were still there on Day 10.

Meanwhile, only four of the 13 neurons they were tracking that responded to visual stimuli still responded on Day 10. The scientists don’t know for sure why the other nine stopped responding, at least to the stimuli they once responded to, but it’s likely they now served a different function.

What are the rules?

Having beheld this extensive wiring and rewiring, the scientists then asked what entitled some spines to survive over the 10-day critical period.

Previous studies have shown that the first inputs to reach binocular visual cortex neurons are from the “contralateral” eye on the opposite side of the head (so in the left hemisphere, the right eye’s inputs get there first), Sur says. These inputs drive a neuron’s soma to respond to specific visual properties such as the orientation of a line — for instance, a 45-degree diagonal. By the time the critical period starts, inputs from the “ipsilateral” eye on the same side of the head begin joining the race to visual cortex neurons, enabling some to become binocular.

It’s no accident that many visual cortex neurons are tuned to lines of different directions in the field of view, Sur says.

“The world is made up of oriented line segments,” Sur notes. “They may be long line segments; they may be short line segments. But the world is not just amorphous globs with hazy boundaries. Objects in the world — trees, the ground, horizons, blades of grass, tables, chairs — are bounded by little line segments.”

Because the researchers were tracking activity at the spines, they could see how often they were active and what orientation triggered that activity. As the data accumulated, they saw that spines were more likely to endure if (a) they were more active, and (b) they responded to the same orientation as the one the soma preferred. Notably, spines that responded to both eyes were more active than spines that responded to just one, meaning binocular spines were more likely to survive than non-binocular ones.

“This observation provides compelling evidence for the ‘use it or lose it’ hypothesis,” says Tsimring. “The more active a spine was, the more likely it was to be retained during development.”

The researchers also noticed another trend. Across the 10 days, clusters emerged along the dendrites in which neighboring spines were increasingly likely to be active at the same time. Other studies have shown that by clustering together, spines are able to combine their activity to be greater than they would be in isolation.

By these rules, over the course of the critical period, neurons apparently refined their role in binocular vision by selectively retaining inputs that reinforced their budding orientation preferences, both via their volume of activity (a synaptic property called “Hebbian plasticity”) and their correlation with their neighbors (a property called “heterosynaptic plasticity”). To confirm that these rules were enough to produce the outcomes they were seeing under the microscope, they built a computer model of a neuron, and indeed the model recapitulated the same trends as what they saw in the mice.

“Both mechanisms are necessary during the critical period to drive the turnover of spines that are misaligned to the soma and to neighboring spine pairs,” the researchers wrote, “which ultimately leads to refinement of [binocular] responses such as orientation matching between the two eyes.”

In addition to Tsimring and Sur, the paper’s other authors are Kyle Jenks, Claudia Cusseddu, Greggory Heller, Jacque Pak Kan Ip, and Julijana Gjorgjieva. Funding sources for the research came from the National Institutes of Health, The Picower Institute for Learning and Memory, and the Freedom Together Foundation.

In 1954, the world’s first successful organ transplant took place at Brigham and Women’s Hospital, in the form of a kidney donated from one twin to the other. At the time, a group of doctors and scientists had correctly theorized that the recipient’s antibodies were unlikely to reject an organ from an identical twin. One Nobel Prize and a few decades later, advancements in immune-suppressing drugs increased the viability of and demand for organ transplants. Today, over 1 million organ transplants have been performed in the United States, more than any other country in the world.

The impressive scale of this achievement was made possible due to advances in organ matching systems: The first computer-based organ matching system was released in 1977. Despite continued innovation in computing, medicine, and matching technology over the years, over 100,000 people in the U.S. are currently on the national transplant waiting list and 13 people die each day waiting for an organ transplant. 

Most computational research in organ allocation is focused on the initial stages, when waitlisted patients are being prioritized for organ transplants. In a new paper presented at ACM Conference on Fairness, Accountability, and Transparency (FAccT) in Athens, Greece, researchers from MIT and Massachusetts General Hospital focused on the final, less-studied stage: organ offer acceptance, when an offer is made and the physician at the transplant center decides on behalf of the patient whether to accept or reject the offered organ. 

“I don’t think we were terribly surprised, but we were obviously disappointed,” co-first author and MIT PhD student Hammaad Adam says. Using computational models to analyze transplantation data from over 160,000 transplant candidates in the Scientific Registry of Transplant Recipients (SRTR) between 2010 and 2020, the researchers found that physicians were overall less likely to accept liver and lung offers on behalf of Black candidates, resulting in additional barriers for Black patients in the organ offer acceptance process.  

For livers, Black patients had 7 percent lower odds of offer acceptance than white patients. When it came to lungs, the disparity became even larger, with 20 percent lower odds of having an offer acceptance than white patients with similar characteristics.

The data don’t necessarily point to clinician bias as the main influence. “The bigger takeaway is that even if there are factors that justify clinical decision-making, there could be clinical conditions that we didn’t control for, that are more common for Black patients,” Adam explains. If the wait-list fails to account for certain patterns in decision-making, they could create obstacles in the process even if the process itself is “unbiased.”

The researchers also point out that high variability in offer acceptance and risk tolerances among transplant centers is a potential factor complicating the decision-making process. Their FAccT paper references a 2020 paper published in JAMA Cardiology, which concluded that wait-list candidates listed at transplant centers with lower offer acceptance rates have a higher likelihood of mortality. 

Another key finding was that an offer was more likely to be accepted if the donor and candidate were of the same race. The paper describes this trend as “concerning,” given the historical inequities in organ procurement that have limited donation from racial and ethnic minority groups. 

Previous work from Adam and his collaborators has aimed to address this gap. Last year, they compiled and released Organ Retrieval and Collection of Health Information for Donation (ORCHID), the first multi-center dataset describing the performance of organ procurement organizations (OPOs). ORCHID contains 10 years’ worth of OPO data, and is intended to facilitate research that addresses bias in organ procurement.

“Being able to do good work in this field takes time,” says Adam, who notes that the entirety of the organ offer acceptance project took years to complete. To his knowledge, only one paper to date studies the association between offer acceptance and race. 

While the bureaucratic and highly interdisciplinary nature of clinical AI projects can dissuade computer science graduate students from pursuing them, Adam committed to the project for the duration of his PhD in the lab of associate professor of electrical engineering Marzyeh Ghassemi, an affiliate of the MIT Jameel Clinic and the Institute of Medical Engineering and Sciences.

To graduate students interested in pursuing clinical AI research projects, Adam recommends that they “free [themselves] from the cycle of publishing every four months.”

“I found it freeing, to be honest — it’s OK if these collaborations take a while,” he says. “It’s hard to avoid that. I made the conscious choice a few years ago and I was happy doing that work.”

This work was supported with funding from the MIT Jameel Clinic. It was also supported, in part, by Takeda Development Center Americas Inc. (successor in interest to Millennium Pharmaceuticals Inc.), an NIH Ruth L. Kirschstein National Research Service Award, a CIFAR AI Chair at the Vector Institute, and by the National Institutes of Health.

Researchers at MIT and other institutions have identified compounds that can fight off viral infection by activating a defense pathway inside host cells. These compounds, they believe, could be used as antiviral drugs that work against not just one but any kind of virus.

The researchers identified these compounds, which activate a host cell defense system known as the integrated stress response pathway, in a screen of nearly 400,000 molecules. In tests in human cells, the researchers showed that the compounds help cells fend off infection from RSV, herpes virus, and Zika virus. They also proved effective in combating herpes infection in a mouse model.

The research team now plans to test the compounds against additional viruses, in hopes of developing them for eventual clinical trials.

“We’re very excited about this work, which allows us to harness the stress response of the host cells to arrive at a means to identify and develop broad-spectrum antivirals,” says James Collins, the Termeer Professor of Medical Engineering and Science in MIT’s Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering.

Collins and Maxwell Wilson, an associate professor of molecular biology at the University of California, Santa Barbara and chief scientific officer of Integrated Biosciences, are the senior authors of the new study, which appears in Cell. Felix Wong, a former MIT postdoc and chief executive officer of Integrated Biosciences, is the lead author of the paper. In addition to MIT, UCSB, and Integrated Biosciences, the research team also includes scientists from Illumina Ventures and Princeton University.

Boosting cell defense

In human cells, the integrated stress response pathway is turned on in response to viral infection as well as other types of stress such as starvation. During viral infection, the pathway is triggered by double-stranded RNA, a molecule produced during the replication cycle of viruses. When that RNA is detected, the cell shuts down protein synthesis, which blocks the virus from producing the proteins it needs to replicate.

Compounds that boost this pathway, the researchers believe, could be good candidates for new antiviral drugs that could combat any type of virus.

“Typically, how antivirals are developed is that you develop one antiviral for one specific virus,” Wong says. “In this case, we hypothesized that being able to modulate the host cell stress response might give us a new class of broad-spectrum antivirals — compounds that directly act on the host cells to alter something fundamental about how all viruses replicate.”

To help them identify compounds that would enhance the activity of this pathway during viral infection, the researchers invented a novel optogenetic screen. Optogenetics is a bioengineering technique that allows researchers to insert light-sensitive proteins into the genome of a cell. In this case, the researchers engineered modifications to a protein called PKR, which turns on the stress pathway, so that they could turn it on with light.

Using this technique, the researchers screened a library of nearly 400,000 commercially available and proprietary chemical compounds. Each of these compounds was applied to human cells as the cells were also exposed to blue light, which simulated viral infection by activating PKR.

By measuring the cells’ survival rates, the researchers could determine which compounds boosted activation of the pathway and amplified the cells’ ability to shut down viral reproduction. This screen yielded about 3,500 compounds with potential antiviral activity, which were evaluated further.

“If the pathway were turned on in response to viral infection, what our compounds do is they turn it on full blast,” Wong says. “Even in the presence of a small amount of virus, if the pathway is triggered, then the antiviral response is also maximized.”

Fighting infection

The researchers then selected eight of the most promising compounds and screened them for their ability to kill viruses while avoiding harmful effects in human cells. Based on these tests, the researchers chose three top candidates, which they called IBX-200, IBX-202, and IBX-204.

In cells that were infected with either Zika virus, herpes virus, or RSV, treatment with these compounds significantly reduced the amount of virus in the cells. The researchers then tested one of the compounds, IBX-200, in mice infected with herpes virus, and found that it was able to reduce the viral load and improve symptoms.

Experiments showed that these compounds appear to turn on an enzyme that is involved in detecting stress. This activates the stress response pathway and primes the cells to become more responsive to viral infection. When applied to cells that are not already infected, the compounds have no effect.

The researchers now plan to evaluate their lead candidates against a broader range of viruses. They also aim to identify additional compounds that activate the integrated stress response, as well as other cellular stress pathways with the potential to clear viral or bacterial infections.

The research was funded by the Defense Threat Reduction Agency, the National Science Foundation, the U.S. Army Research Office, and Integrated Biosciences.

When ChatGPT or Gemini give what seems to be an expert response to your burning questions, you may not realize how much information it relies on to give that reply. Like other popular generative artificial intelligence (AI) models, these chatbots rely on backbone systems called foundation models that train on billions, or even trillions, of data points.

In a similar vein, engineers are hoping to build foundation models that train a range of robots on new skills like picking up, moving, and putting down objects in places like homes and factories. The problem is that it’s difficult to collect and transfer instructional data across robotic systems. You could teach your system by teleoperating the hardware step-by-step using technology like virtual reality (VR), but that can be time-consuming. Training on videos from the internet is less instructive, since the clips don’t provide a step-by-step, specialized task walk-through for particular robots.

A simulation-driven approach called “PhysicsGen” from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Robotics and AI Institute customizes robot training data to help robots find the most efficient movements for a task. The system can multiply a few dozen VR demonstrations into nearly 3,000 simulations per machine. These high-quality instructions are then mapped to the precise configurations of mechanical companions like robotic arms and hands. 

PhysicsGen creates data that generalize to specific robots and condition via a three-step process. First, a VR headset tracks how humans manipulate objects like blocks using their hands. These interactions are mapped in a 3D physics simulator at the same time, visualizing the key points of our hands as small spheres that mirror our gestures. For example, if you flipped a toy over, you’d see 3D shapes representing different parts of your hands rotating a virtual version of that object.

The pipeline then remaps these points to a 3D model of the setup of a specific machine (like a robotic arm), moving them to the precise “joints” where a system twists and turns. Finally, PhysicsGen uses trajectory optimization — essentially simulating the most efficient motions to complete a task — so the robot knows the best ways to do things like repositioning a box.

Each simulation is a detailed training data point that walks a robot through potential ways to handle objects. When implemented into a policy (or the action plan that the robot follows), the machine has a variety of ways to approach a task, and can try out different motions if one doesn’t work.

“We’re creating robot-specific data without needing humans to re-record specialized demonstrations for each machine,” says Lujie Yang, an MIT PhD student in electrical engineering and computer science and CSAIL affiliate who is the lead author of a new paper introducing the project. “We’re scaling up the data in an autonomous and efficient way, making task instructions useful to a wider range of machines.”

Generating so many instructional trajectories for robots could eventually help engineers build a massive dataset to guide machines like robotic arms and dexterous hands. For example, the pipeline might help two robotic arms collaborate on picking up warehouse items and placing them in the right boxes for deliveries. The system may also guide two robots to work together in a household on tasks like putting away cups.

PhysicsGen’s potential also extends to converting data designed for older robots or different environments into useful instructions for new machines. “Despite being collected for a specific type of robot, we can revive these prior datasets to make them more generally useful,” adds Yang.

Addition by multiplication

PhysicsGen turned just 24 human demonstrations into thousands of simulated ones, helping both digital and real-world robots reorient objects.

Yang and her colleagues first tested their pipeline in a virtual experiment where a floating robotic hand needed to rotate a block into a target position. The digital robot executed the task at a rate of 81 percent accuracy by training on PhysicGen’s massive dataset, a 60 percent improvement from a baseline that only learned from human demonstrations.

The researchers also found that PhysicsGen could improve how virtual robotic arms collaborate to manipulate objects. Their system created extra training data that helped two pairs of robots successfully accomplish tasks as much as 30 percent more often than a purely human-taught baseline.

In an experiment with a pair of real-world robotic arms, the researchers observed similar improvements as the machines teamed up to flip a large box into its designated position. When the robots deviated from the intended trajectory or mishandled the object, they were able to recover mid-task by referencing alternative trajectories from their library of instructional data.

Senior author Russ Tedrake, who is the Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering at MIT, adds that this imitation-guided data generation technique combines the strengths of human demonstration with the power of robot motion planning algorithms.

“Even a single demonstration from a human can make the motion planning problem much easier,” says Tedrake, who is also a senior vice president of large behavior models at the Toyota Research Institute and CSAIL principal investigator. “In the future, perhaps the foundation models will be able to provide this information, and this type of data generation technique will provide a type of post-training recipe for that model.”

The future of PhysicsGen

Soon, PhysicsGen may be extended to a new frontier: diversifying the tasks a machine can execute.

“We’d like to use PhysicsGen to teach a robot to pour water when it’s only been trained to put away dishes, for example,” says Yang. “Our pipeline doesn’t just generate dynamically feasible motions for familiar tasks; it also has the potential of creating a diverse library of physical interactions that we believe can serve as building blocks for accomplishing entirely new tasks a human hasn’t demonstrated.”

Creating lots of widely applicable training data may eventually help build a foundation model for robots, though MIT researchers caution that this is a somewhat distant goal. The CSAIL-led team is investigating how PhysicsGen can harness vast, unstructured resources — like internet videos — as seeds for simulation. The goal: transform everyday visual content into rich, robot-ready data that could teach machines to perform tasks no one explicitly showed them.

Yang and her colleagues also aim to make PhysicsGen even more useful for robots with diverse shapes and configurations in the future. To make that happen, they plan to leverage datasets with demonstrations of real robots, capturing how robotic joints move instead of human ones.

The researchers also plan to incorporate reinforcement learning, where an AI system learns by trial and error, to make PhysicsGen expand its dataset beyond human-provided examples. They may augment their pipeline with advanced perception techniques to help a robot perceive and interpret their environment visually, allowing the machine to analyze and adapt to the complexities of the physical world.

For now, PhysicsGen shows how AI can help us teach different robots to manipulate objects within the same category, particularly rigid ones. The pipeline may soon help robots find the best ways to handle soft items (like fruits) and deformable ones (like clay), but those interactions aren’t easy to simulate yet.

Yang and Tedrake wrote the paper with two CSAIL colleagues: co-lead author and MIT PhD student Hyung Ju “Terry” Suh SM ’22 and MIT PhD student Bernhard Paus Græsdal. Robotics and AI Institute researchers Tong Zhao ’22, MEng ’23, Tarik Kelestemur, Jiuguang Wang, and Tao Pang PhD ’23 are also authors. Their work was supported by the Robotics and AI Institute and Amazon.

The researchers recently presented their work at the Robotics: Science and Systems conference.

In order to produce effective targeted therapies for cancer, scientists need to isolate the genetic and phenotypic characteristics of cancer cells, both within and across different tumors, because those differences impact how tumors respond to treatment.

Part of this work requires a deep understanding of the RNA or protein molecules each cancer cell expresses, where it is located in the tumor, and what it looks like under a microscope.

Traditionally, scientists have looked at one or more of these aspects separately, but now a new deep learning AI tool, CellLENS (Cell Local Environment and Neighborhood Scan), fuses all three domains together, using a combination of convolutional neural networks and graph neural networks to build a comprehensive digital profile for every single cell. This allows the system to group cells with similar biology — effectively separating even those that appear very similar in isolation, but behave differently depending on their surroundings.

The study, published recently in Nature Immunology, details the results of a collaboration between researchers from MIT, Harvard Medical School, Yale University, Stanford University, and University of Pennsylvania — an effort led by Bokai Zhu, an MIT postdoc and member of the Broad Institute of MIT and Harvard and the Ragon Institute of MGH, MIT, and Harvard.

Zhu explains the impact of this new tool: “Initially we would say, oh, I found a cell. This is called a T cell. Using the same dataset, by applying CellLENS, now I can say this is a T cell, and it is currently attacking a specific tumor boundary in a patient.

“I can use existing information to better define what a cell is, what is the subpopulation of that cell, what that cell is doing, and what is the potential functional readout of that cell. This method may be used to identify a new biomarker, which provides specific and detailed information about diseased cells, allowing for more targeted therapy development.”

This is a critical advance because current methodologies often miss critical molecular or contextual information — for example, immunotherapies may target cells that only exist at the boundary of a tumor, limiting efficacy. By using deep learning, the researchers can detect many different layers of information with CellLENS, including morphology and where the cell is spatially in a tissue.

When applied to samples from healthy tissue and several types of cancer, including lymphoma and liver cancer, CellLENS uncovered rare immune cell subtypes and revealed how their activity and location relate to disease processes — such as tumor infiltration or immune suppression.

These discoveries could help scientists better understand how the immune system interacts with tumors and pave the way for more precise cancer diagnostics and immunotherapies.

“I’m extremely excited by the potential of new AI tools, like CellLENS, to help us more holistically understand aberrant cellular behaviors within tissues,” says co-author Alex K. Shalek, the director of the Institute for Medical Engineering and Science (IMES), the J. W. Kieckhefer Professor in IMES and Chemistry, and an extramural member of the Koch Institute for Integrative Cancer Research at MIT, as well as an Institute member of the Broad Institute and a member of the Ragon Institute. “We can now measure a tremendous amount of information about individual cells and their tissue contexts with cutting-edge, multi-omic assays. Effectively leveraging that data to nominate new therapeutic leads is a critical step in developing improved interventions. When coupled with the right input data and careful downsteam validations, such tools promise to accelerate our ability to positively impact human health and wellness.”

For many years, health experts have been concerned about “food deserts,” places where residents lack good nutritional options. Now, an MIT-led study of three major global cities uses a new, granular method to examine the issue, and concludes that having fewer and less nutritional eating options nearby correlates with obesity and other health outcomes.

Rather than just mapping geographic areas, the researchers examined the dietary value of millions of food items on roughly 30,000 restaurant menus and derived a more precise assessment of the connection between neighborhoods and nutrition.

“We show that what is sold in a restaurant has a direct correlation to people’s health,” says MIT researcher Fabio Duarte, co-author of a newly published paper outlining the study’s results. “The food landscape matters.”

The open-access paper, “Data-driven nutritional assessment of urban food landscapes: insights from Boston, London, Dubai,” was published this week in Nature: Scientific Reports.

The co-authors are Michael Tufano, a PhD student at Wageningen University, in the Netherlands; Duarte, associate director of MIT’s Senseable City Lab, which uses data to study cities as dynamic systems; Martina Mazzarello, a postdoc at the Senseable City Lab; Javad Eshtiyagh, a research fellow at the Senseable City Lab; Carlo Ratti, professor of the practice and director of the Senseable City Lab; and Guido Camps, a senior researcher at Wageningen University.

Scanning the menu

To conduct the study, the researchers examined menus from Boston, Dubai, and London, in the summer of 2023, compiling a database of millions of items available through popular food-delivery platforms. The team then evaluated the food items as rated by the USDA’s FoodData Central database, an information bank with 375,000 kinds of food products listed. The study deployed two main metrics, the Meal Balance Index, and the Nutrient-Rich Foods Index.

The researchers examined about 222,000 menu items from over 2,000 restaurants in Boston, about 1.6 million menu items from roughly 9,000 restaurants in Dubai, and about 3.1 million menu items from about 18,000 restaurants in London. In Boston, about 71 percent of the items were in the USDA database; in Dubai and London, that figure was 42 percent and 56 percent, respectively.

The team then rated the nutritional value of the items appearing on menus, and correlated the food data with health-outcome data from Boston and London. In London, they found a clear correlation between neighborhood menu offerings and obesity, or the lack thereof; with a slightly less firm correlation in Boston. Areas with food options that include a lot of dietary fibers, sometimes along with fruits and vegetables, tend to have better health data.

In Dubai, the researchers did not have the same types of health data available but did observe a strong correlation between rental prices and the nutritional value of neighborhood-level food, suggesting that wealthier residents have better nourishment options.

“At the item level, when we have less nutritional food, we see more cases of obsesity,” Tufano says. “It’s true that not only do we have more fast food in poor neighborhoods, but the nutritional value is not the same.”

Re-mapping the food landscape

By conducting the study in this fashion, the scholars added a layer of analysis to past studies of food deserts. While past work has broken ground by identifying neighborhoods and areas lacking good food access, this research makes a more comprehensive assessment of what people consume. The research moves toward evaluating the complex mix of food available in any given area, which can be true even of areas with more limited options.

“We were not satisfied with this idea that if you only have fast food, it’s a food desert, but if you have a Whole Foods, it’s not,” Duarte says. “It’s not necessarily like that.”

For the Senseable City Lab researchers, the study is a new technique further enabling them to understand city dynamics and the effects of the urban environment on health. Past lab studies have often focused on issues such as urban mobility, while extending to matters such as mobility and air pollution, among other topics.

Being able to study food and health at the neighborhood level, though, is still another example of the ways that data-rich spheres of life can be studied in close detail.

“When we started working on cities and data, the data resolution was so low,” Ratti says. “Today the amount of data is so immense we see this great opportunity to look at cities and see the influence of the urban environment as a big determinant of health. We see this as one of the new frontiers of our lab. It’s amazing how we can now look at this very precisely in cities.”

MIT researchers have developed a new bionic knee that can help people with above-the-knee amputations walk faster, climb stairs, and avoid obstacles more easily than they could with a traditional prosthesis.

Unlike prostheses in which the residual limb sits within a socket, the new system is directly integrated with the user’s muscle and bone tissue. This enables greater stability and gives the user much more control over the movement of the prosthesis.

Participants in a small clinical study also reported that the limb felt more like a part of their own body, compared to people who had more traditional above-the-knee amputations.

“A prosthesis that's tissue-integrated — anchored to the bone and directly controlled by the nervous system — is not merely a lifeless, separate device, but rather a system that is carefully integrated into human physiology, offering a greater level of prosthetic embodiment. It’s not simply a tool that the human employs, but rather an integral part of self,” says Hugh Herr, a professor of media arts and sciences, co-director of the K. Lisa Yang Center for Bionics at MIT, an associate member of MIT’s McGovern Institute for Brain Research, and the senior author of the new study.

Tony Shu PhD ’24 is the lead author of the paper, which appears today in Science.

Better control

Over the past several years, Herr’s lab has been working on new prostheses that can extract neural information from muscles left behind after an amputation and use that information to help guide a prosthetic limb.

During a traditional amputation, pairs of muscles that take turns stretching and contracting are usually severed, disrupting the normal agonist-antagonist relationship of the muscles. This disruption makes it very difficult for the nervous system to sense the position of a muscle and how fast it’s contracting.

Using the new surgical approach developed by Herr and his colleagues, known as agonist-antagonist myoneuronal interface (AMI), muscle pairs are reconnected during surgery so that they still dynamically communicate with each other within the residual limb. This sensory feedback helps the wearer of the prosthesis to decide how to move the limb, and also generates electrical signals that can be used to control the prosthetic limb.

In a 2024 study, the researchers showed that people with amputations below the knee who received the AMI surgery were able to walk faster and navigate around obstacles much more naturally than people with traditional below-the-knee amputations.

In the new study, the researchers extended the approach to better serve people with amputations above the knee. They wanted to create a system that could not only read out signals from the muscles using AMI but also be integrated into the bone, offering more stability and better sensory feedback.

To achieve that, the researchers developed a procedure to insert a titanium rod into the residual femur bone at the amputation site. This implant allows for better mechanical control and load bearing than a traditional prosthesis. Additionally, the implant contains 16 wires that collect information from electrodes located on the AMI muscles inside the body, which enables more accurate transduction of the signals coming from the muscles.

This bone-integrated system, known as e-OPRA, transmits AMI signals to a new robotic controller developed specifically for this study. The controller uses this information to calculate the torque necessary to move the prosthesis the way that the user wants it to move.

“All parts work together to better get information into and out of the body and better interface mechanically with the device,” Shu says. “We’re directly loading the skeleton, which is the part of the body that’s supposed to be loaded, as opposed to using sockets, which is uncomfortable and can lead to frequent skin infections.”

In this study, two subjects received the combined AMI and e-OPRA system, known as an osseointegrated mechanoneural prosthesis (OMP). These users were compared with eight who had the AMI surgery but not the e-OPRA implant, and seven users who had neither AMI nor e-OPRA. All subjects took a turn at using an experimental powered knee prosthesis developed by the lab.

The researchers measured the participants’ ability to perform several types of tasks, including bending the knee to a specified angle, climbing stairs, and stepping over obstacles. In most of these tasks, users with the OMP system performed better than the subjects who had the AMI surgery but not the e-OPRA implant, and much better than users of traditional prostheses.

“This paper represents the fulfillment of a vision that the scientific community has had for a long time — the implementation and demonstration of a fully physiologically integrated, volitionally controlled robotic leg,” says Michael Goldfarb, a professor of mechanical engineering and director of the Center for Intelligent Mechatronics at Vanderbilt University, who was not involved in the research. “This is really difficult work, and the authors deserve tremendous credit for their efforts in realizing such a challenging goal.”

A sense of embodiment

In addition to testing gait and other movements, the researchers also asked questions designed to evaluate participants’ sense of embodiment — that is, to what extent their prosthetic limb felt like a part of their own body.

Questions included whether the patients felt as if they had two legs, if they felt as if the prosthesis was part of their body, and if they felt in control of the prosthesis. Each question was designed to evaluate the participants’ feelings of agency, ownership of device, and body representation.

The researchers found that as the study went on, the two participants with the OMP showed much greater increases in their feelings of agency and ownership than the other subjects.

“Another reason this paper is significant is that it looks into these embodiment questions and it shows large improvements in that sensation of embodiment,” Herr says. “No matter how sophisticated you make the AI systems of a robotic prosthesis, it’s still going to feel like a tool to the user, like an external device. But with this tissue-integrated approach, when you ask the human user what is their body, the more it’s integrated, the more they’re going to say the prosthesis is actually part of self.”

The AMI procedure is now done routinely on patients with below-the-knee amputations at Brigham and Women’s Hospital, and Herr expects it will soon become the standard for above-the-knee amputations as well. The combined OMP system will need larger clinical trials to receive FDA approval for commercial use, which Herr expects may take about five years.

The research was funded by the Yang Tan Collective and DARPA.

Marine scientists have long marveled at how animals like fish and seals swim so efficiently despite having different shapes. Their bodies are optimized for efficient, hydrodynamic aquatic navigation so they can exert minimal energy when traveling long distances.

Autonomous vehicles can drift through the ocean in a similar way, collecting data about vast underwater environments. However, the shapes of these gliding machines are less diverse than what we find in marine life — go-to designs often resemble tubes or torpedoes, since they’re fairly hydrodynamic as well. Plus, testing new builds requires lots of real-world trial-and-error.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the University of Wisconsin at Madison propose that AI could help us explore uncharted glider designs more conveniently. Their method uses machine learning to test different 3D designs in a physics simulator, then molds them into more hydrodynamic shapes. The resulting model can be fabricated via a 3D printer using significantly less energy than hand-made ones.

The MIT scientists say that this design pipeline could create new, more efficient machines that help oceanographers measure water temperature and salt levels, gather more detailed insights about currents, and monitor the impacts of climate change. The team demonstrated this potential by producing two gliders roughly the size of a boogie board: a two-winged machine resembling an airplane, and a unique, four-winged object resembling a flat fish with four fins.

Peter Yichen Chen, MIT CSAIL postdoc and co-lead researcher on the project, notes that these designs are just a few of the novel shapes his team’s approach can generate. “We’ve developed a semi-automated process that can help us test unconventional designs that would be very taxing for humans to design,” he says. “This level of shape diversity hasn’t been explored previously, so most of these designs haven’t been tested in the real world.”

But how did AI come up with these ideas in the first place? First, the researchers found 3D models of over 20 conventional sea exploration shapes, such as submarines, whales, manta rays, and sharks. Then, they enclosed these models in “deformation cages” that map out different articulation points that the researchers pulled around to create new shapes.

The CSAIL-led team built a dataset of conventional and deformed shapes before simulating how they would perform at different “angles-of-attack” — the direction a vessel will tilt as it glides through the water. For example, a swimmer may want to dive at a -30 degree angle to retrieve an item from a pool.

These diverse shapes and angles of attack were then used as inputs for a neural network that essentially anticipates how efficiently a glider shape will perform at particular angles and optimizes it as needed.

Giving gliding robots a lift

The team’s neural network simulates how a particular glider would react to underwater physics, aiming to capture how it moves forward and the force that drags against it. The goal: find the best lift-to-drag ratio, representing how much the glider is being held up compared to how much it’s being held back. The higher the ratio, the more efficiently the vehicle travels; the lower it is, the more the glider will slow down during its voyage.

Lift-to-drag ratios are key for flying planes: At takeoff, you want to maximize lift to ensure it can glide well against wind currents, and when landing, you need sufficient force to drag it to a full stop.

Niklas Hagemann, an MIT graduate student in architecture and CSAIL affiliate, notes that this ratio is just as useful if you want a similar gliding motion in the ocean.

“Our pipeline modifies glider shapes to find the best lift-to-drag ratio, optimizing its performance underwater,” says Hagemann, who is also a co-lead author on a paper that was presented at the International Conference on Robotics and Automation in June. “You can then export the top-performing designs so they can be 3D-printed.”

Going for a quick glide

While their AI pipeline seemed realistic, the researchers needed to ensure its predictions about glider performance were accurate by experimenting in more lifelike environments.

They first fabricated their two-wing design as a scaled-down vehicle resembling a paper airplane. This glider was taken to MIT’s Wright Brothers Wind Tunnel, an indoor space with fans that simulate wind flow. Placed at different angles, the glider’s predicted lift-to-drag ratio was only about 5 percent higher on average than the ones recorded in the wind experiments — a small difference between simulation and reality.

A digital evaluation involving a visual, more complex physics simulator also supported the notion that the AI pipeline made fairly accurate predictions about how the gliders would move. It visualized how these machines would descend in 3D.

To truly evaluate these gliders in the real world, though, the team needed to see how their devices would fare underwater. They printed two designs that performed the best at specific points-of-attack for this test: a jet-like device at 9 degrees and the four-wing vehicle at 30 degrees.

Both shapes were fabricated in a 3D printer as hollow shells with small holes that flood when fully submerged. This lightweight design makes the vehicle easier to handle outside of the water and requires less material to be fabricated. The researchers placed a tube-like device inside these shell coverings, which housed a range of hardware, including a pump to change the glider’s buoyancy, a mass shifter (a device that controls the machine’s angle-of-attack), and electronic components.

Each design outperformed a handmade torpedo-shaped glider by moving more efficiently across a pool. With higher lift-to-drag ratios than their counterpart, both AI-driven machines exerted less energy, similar to the effortless ways marine animals navigate the oceans.

As much as the project is an encouraging step forward for glider design, the researchers are looking to narrow the gap between simulation and real-world performance. They are also hoping to develop machines that can react to sudden changes in currents, making the gliders more adaptable to seas and oceans.

Chen adds that the team is looking to explore new types of shapes, particularly thinner glider designs. They intend to make their framework faster, perhaps bolstering it with new features that enable more customization, maneuverability, or even the creation of miniature vehicles.

Chen and Hagemann co-led research on this project with OpenAI researcher Pingchuan Ma SM ’23, PhD ’25. They authored the paper with Wei Wang, a University of Wisconsin at Madison assistant professor and recent CSAIL postdoc; John Romanishin ’12, SM ’18, PhD ’23; and two MIT professors and CSAIL members: lab director Daniela Rus and senior author Wojciech Matusik. Their work was supported, in part, by a Defense Advanced Research Projects Agency (DARPA) grant and the MIT-GIST Program.

Common sense tells us to run from molten lava flowing from active volcanoes. But MIT professors J. Jih, Cristina Parreño Alonso, and Skylar Tibbits — faculty in the Department of Architecture at the School of Architecture and Planning — have their bags packed to head to southwest Iceland in anticipation of an imminent volcanic eruption. The Nordic island nation is currently experiencing a period of intense seismic activity; seven volcanic eruptions have taken place in its southern peninsula in under a year.

Earlier this year, the faculty built and placed a series of lightweight, easily deployable steel structures close to the volcano, where a few of the recent eruptions have taken place; several more structures are on trucks waiting to be delivered to sites where fissures open and lava oozes out. Cameras are in place to record what happens when the lava meets and hits these structures to help understand the lava flows.

This new research explores what type of shapes and materials can be used to interact with lava and successfully divert it from heading in the direction of habitats or critical infrastructure that lie in its path. Their work is supported by a Professor Amar. G. Bose Research Grant.

“We’re trying to imagine new ways of conceptualizing infrastructure when it relates to lava and volcanic eruptions,” says Jih, an associate professor of the practice. “Lovely for us as designers, physical prototyping is the only way you can test some of these ideas out.” 

Currently, the Icelandic Department of Civic Protection and Emergency Management and an engineering group, EFLA, are diverting the lava with massive berms (approximately 44 to 54 yards in length and 9 yards in height) made from earth and stone.

Berms protecting the town of Grindavik, a power plant, and the popular Blue Lagoon geothermal spa have met with mixed results. In November 2024, a volcano erupted for the seventh time in less than a year, forcing the evacuation of town residents and the Blue Lagoon’s guests and employees. The latter’s parking lot was consumed by lava.

Sigurdur Thorsteinsson, chief brand, design, and innovation officer of the Blue Lagoon, as well as a designer and a partner in Design Group Italia, was on site for this eruption and several others.

“Some magma went into the city of Grindavik and three or four houses were destroyed,” says Thorsteinsson. “One of our employees watched her house go under magma on television, which was an emotional moment.”

While staff at the Blue Lagoon have become very efficient at evacuating guests, says Thorsteinsson, each eruption forces the tourist destination to close and townspeople to evacuate, disrupting lives and livelihoods.

“You cannot really stop the magma,” says Thorsteinsson, who is working with the MIT faculty on this research project. “It’s too powerful.”

Tibbits, associate professor of design research and founder and co-director of the Self-Assembly Lab, agrees. His research explores how to guide or work with the forces of nature.

Last year, Tibbits and Jih were in Iceland on another research project when erupting volcanoes interrupted their work. The two started thinking about how the lava could be redirected.

“The question is: Can we find more strategic interventions in the field that could work with the lava, rather than fight it?” says Tibbits.

To investigate what kinds of materials would withstand this type of interaction, they invited Parreño Alonso, a senior lecturer in the Department of Architecture, to join them.

“Cristina, being the department authority on magma, was an obvious and important partner for us,” says Jih with a smile.

Parreño Alonso has been working with volcanic rock for years and taught a series of design studios exploring volcanic rock as an architectural material. She also has proposed designing structures to engage directly with lava flows and recently has been examining volcanic rock in a molten state and melting basalt in MIT’s foundry with Michael Tarkanian, a senior lecturer in MIT’s Department of Materials Science and Engineering, and Metals Lab director. For this project, she is exploring the potential of molten rock as a substitute for concrete, a widely used material because of its pliability.

“It’s exciting how this idea of working with volcanoes was taking shape in parallel, from different angles, within the same department,” says Parreño Alonso. “I love how these parallel interests have led to such a beautiful collaboration.”

She also sees other opportunities by collaborating with these forces of nature.

“We are interested in the potential of generating something out of the interaction with the lava,” she says. “Could it be a landscape that becomes a park? There are many possibilities.”

The steel structures were first tested at MIT’s Metals Lab with Tarkanian and then built onsite in Iceland. The team wanted to make the structures lightweight so they could be quickly set up in the field, but strong enough so they wouldn’t be easily destroyed. Various designs were created; this iteration of the design has V-shaped structures that can guide the lava to flow around them, or they can be reconfigured as ramps or tunnels.

“There is a road that has been hit by many of the recent eruptions and must keep being rebuilt,” says Tibbits. “We created two ramps that could in the future serve as tunnels, allowing the lava to flow over the road and create a type of lava cave where the cars could drive under the cooled lava.”

Tibbits says they see the structures in the field now as an initial intervention. After documenting and studying how they interact with the lava, the architects will develop new iterations of what they believe will eventually become critical infrastructure for locations around the world with active volcanoes.

“If we can show and prove what kinds of shapes and structures and what kinds of materials can divert magma flows, I think it’s incredibly valuable research,” says Thorsteinsson.

Thorsteinsson lives in Italy half of the year and says the volcanoes there — Mount Etna in Sicily and Mount Vesuvius in the Gulf of Naples — pose a greater danger than those in Iceland because of the densely populated neighborhoods nearby. Volcanoes in Hawaii and Japan are in similarly populated areas.

“Whatever information you can learn about diverting magma flows to other directions and what kinds of structures are needed — it would be priceless,” he says.

For people with Type 1 diabetes, developing hypoglycemia, or low blood sugar, is an ever-present threat. When glucose levels become extremely low, it creates a life-threatening situation for which the standard treatment of care is injecting a hormone called glucagon.

As an emergency backup, for cases where patients may not realize that their blood sugar is dropping to dangerous levels, MIT engineers have designed an implantable reservoir that can remain under the skin and be triggered to release glucagon when blood sugar levels get too low.

This approach could also help in cases where hypoglycemia occurs during sleep, or for diabetic children who are unable to administer injections on their own.

“This is a small, emergency-event device that can be placed under the skin, where it is ready to act if the patient’s blood sugar drops too low,” says Daniel Anderson, a professor in MIT’s Department of Chemical Engineering, a member of MIT’s Koch Institute for Integrative Cancer Research and Institute for Medical Engineering and Science (IMES), and the senior author of the study. “Our goal was to build a device that is always ready to protect patients from low blood sugar. We think this can also help relieve the fear of hypoglycemia that many patients, and their parents, suffer from.”

The researchers showed that this device could also be used to deliver emergency doses of epinephrine, a drug that is used to treat heart attacks and can also prevent severe allergic reactions, including anaphylactic shock.

Siddharth Krishnan, a former MIT research scientist who is now an assistant professor of electrical engineering at Stanford University, is the lead author of the study, which appears today in Nature Biomedical Engineering.

Emergency response

Most patients with type 1 diabetes use daily insulin injections to help their body absorb sugar and prevent their blood sugar levels from getting too high. However, if their blood sugar levels get too low, they develop hypoglycemia, which can lead to confusion and seizures, and may be fatal if it goes untreated.

To combat hypoglycemia, some patients carry preloaded syringes of glucagon, a hormone that stimulates the liver to release glucose into the bloodstream. However, it isn’t always easy for people, especially children, to know when they are becoming hypoglycemic.

“Some patients can sense when they’re getting low blood sugar, and go eat something or give themselves glucagon,” Anderson says. “But some are unaware that they’re hypoglycemic, and they can just slip into confusion and coma. This is also a problem when patients sleep, as they are reliant on glucose sensor alarms to wake them when sugar drops dangerously low.”

To make it easier to counteract hypoglycemia, the MIT team set out to design an emergency device that could be triggered either by the person using it, or automatically by a sensor.

The device, which is about the size of a quarter, contains a small drug reservoir made of a 3D-printed polymer. The reservoir is sealed with a special material known as a shape-memory alloy, which can be programmed to change its shape when heated. In this case, the researcher used a nickel-titanium alloy that is programmed to curl from a flat slab into a U-shape when heated to 40 degrees Celsius.

Like many other protein or peptide drugs, glucagon tends to break down quickly, so the liquid form can’t be stored long-term in the body. Instead, the MIT team created a powdered version of the drug, which remains stable for much longer and stays in the reservoir until released.

Each device can carry either one or four doses of glucagon, and it also includes an antenna tuned to respond to a specific frequency in the radiofrequency range. That allows it to be remotely triggered to turn on a small electrical current, which is used to heat the shape-memory alloy. When the temperature reaches the 40-degree threshold, the slab bends into a U shape, releasing the contents of the reservoir.

Because the device can receive wireless signals, it could also be designed so that drug release is triggered by a glucose monitor when the wearer’s blood sugar drops below a certain level.

“One of the key features of this type of digital drug delivery system is that you can have it talk to sensors,” Krishnan says. “In this case, the continuous glucose-monitoring technology that a lot of patients use is something that would be easy for these types of devices to interface with.”

Reversing hypoglycemia

After implanting the device in diabetic mice, the researchers used it to trigger glucagon release as the animals’ blood sugar levels were dropping. Within less than 10 minutes of activating the drug release, blood sugar levels began to level off, allowing them to remain within the normal range and avert hypoglycemia.

The researchers also tested the device with a powdered version of epinephrine. They found that within 10 minutes of drug release, epinephrine levels in the bloodstream became elevated and heart rate increased.

In this study, the researchers kept the devices implanted for up to four weeks, but they now plan to see if they can extend that time up to at least a year.

“The idea is you would have enough doses that can provide this therapeutic rescue event over a significant period of time. We don’t know exactly what that is — maybe a year, maybe a few years, and we’re currently working on establishing what the optimal lifetime is. But then after that, it would need to be replaced,” Krishnan says.

Typically, when a medical device is implanted in the body, scar tissue develops around the device, which can interfere with its function. However, in this study, the researchers showed that even after fibrotic tissue formed around the implant, they were able to successfully trigger the drug release.

The researchers are now planning for additional animal studies and hope to begin testing the device in clinical trials within the next three years.

“It’s really exciting to see our team accomplish this, which I hope will someday help diabetic patients and could more broadly provide a new paradigm for delivering any emergency medicine,” says Robert Langer, the David H. Koch Institute Professor at MIT and an author of the paper.

Other authors of the paper include Laura O’Keeffe, Arnab Rudra, Derin Gumustop, Nima Khatib, Claudia Liu, Jiawei Yang, Athena Wang, Matthew Bochenek, Yen-Chun Lu, Suman Bose, and Kaelan Reed.

The research was funded by the Leona M. and Harry B. Helmsley Charitable Trust, the National Institutes of Health, a JDRF postdoctoral fellowship, and the National Institute of Biomedical Imaging and Bioengineering.

This work was carried out, in part, through the use of MIT.nano’s facilities.

The first time Steve Jobs held a public demo of the Apple Macintosh, in early 1984, scripted jokes were part of the rollout. First, Jobs pulled the machine out of a bag. Then, using speech technology from Samsung, the Macintosh made a quip about rival IBM’s mainframes: “Never trust a computer you can’t lift.”

There’s a reason Jobs was doing that. For the first few decades that computing became part of cultural life, starting in the 1950s, computers seemed unfriendly, grim, and liable to work against human interests. Take the 1968 film “2001: A Space Odyssey,” in which the onboard computer, HAL, turns against the expedition’s astronauts. It’s a famous cultural touchstone. Jobs, in selling the idea of a personal computer, was using humor to ease concerns about the machines.

“Against the sense of computing as cold and numbers-driven, the fact that this computer was using voice technology to deliver jokes made it seem less forbidding, less evil,” says MIT scholar Benjamin Mangrum.

In fact, this dynamic turns up throughout modern culture, in movies, television, fiction, and the theater. We often deal with our doubts and fears about computing through humor, whether reconciling ourselves to machines or critiquing them. Now, Mangrum analyzes this phenomenon in a new book, “The Comedy of Computation: Or, How I Learned to Stop Worrying and Love Obsolescence,” published this month by Stanford University Press.

“Comedy has been a form for making this technology seem ordinary,” says Mangrum, an associate professor in MIT’s literature program. “Where in other circumstances computing might seem inhuman or impersonal, comedy allows us to incorporate it into our lives in a way that makes it make sense.”

Reversals of fortune

Mangrum’s interest in the subject was sparked partly by William Marchant’s 1955 play, “The Desk Set” — a romantic comedy later turned into a film starring Katharine Hepburn and Spencer Tracy — which queries, among other things, how office workers will co-exist alongside computers.

Perhaps against expectations, romantic comedies have turned out to be one of the most prominent contemporary forms of culture that grapple with technology and its effects on us. Mangrum, in the book, explains why: Their plot structure often involves reversals, which sometimes are extended to technology, too. Computing might seem forbidding, but it might also pull people together.

“One of the common tropes about romantic comedies is that there are characters or factors in the drama that obstruct the happy union of two people,” Mangrum observes. “And often across the arc of the drama, the obstruction or obstructive character is transformed into a partner, or collaborator, and assimilated within the happy couple’s union. That provides a template for how some cultural producers want to present the experience of computing. It begins as an obstruction and ends as a partner.”

That plot structure, Mangrum notes, dates to antiquity and was common in Shakespeare’s day. Still, as he writes in the book, there is “no timeless reality called Comedy,” as the vehicles and forms of it change over time. Beyond that, specific jokes about computing can quickly become outmoded. Steve Jobs made fun of mainframes, and the 1998 Nora Ephron comedy “You’ve Got Mail” got laughs out of dial-up modems, but those jokes might leave most people puzzled today.

“Comedy is not a fixed resource,” Mangrum says. “It’s an ever-changing toolbox.”

Continuing this evolution into the 21st century, Mangrum observes that a lot of computational comedy centers on an entire category of commentary he calls “the Great Tech-Industrial Joke.” This focuses on the gap between noble-sounding declared aspirations of technology and the sometimes-dismal outcomes it creates.

Social media, for instance, promised new worlds of connectivity and social exploration, and has benefits people enjoy — but it has also generated polarization, misinformation, and toxicity. Technology’s social effects are complex. Whole televisions shows, such as “Silicon Valley,” have dug into this terrain.

“The tech industry announces that some of its products have revolutionary or utopian aims, but the achievements of many of them fall far short of that,” Mangrum says. “It’s a funny setup for a joke. People have been claiming we’re saving the world, when actually we’re just processing emails faster. But it’s a mode of criticism aimed at big tech, since its products are more complicated.”

A complicated, messy picture

“The Comedy of Computation” digs into several other facets of modern culture and technology. The notion of personal authenticity, as Mangrum observes, is a fairly recent and modern construct in society — and it’s another sphere of life that collides with computing, since social media is full of charges of inauthenticity.

“That ethics of authenticity connects to comedy, as we make jokes about people not being authentic,” Mangrum says.

“The Comedy of Computation” has received praise from other scholars. Mark Goble, a professor of English at the University of California at Berkeley, has called it “essential for understanding the technological world in its complexity, absurdity, and vibrancy.”

For his part, Mangrum emphasizes that his book is an exploration of the full complexity of technology, culture, and society.

“There’s this really complicated, messy picture,” Mangrum says. “And comedy sometimes finds a way of experiencing and finding pleasure in that messiness, and other times it neatly wraps it up in a lesson that can make things neater than they actually are.”

Mangrum adds that the book focuses on “the combination of the threat and pleasure that’s involved across the history of the computer, in the ways it’s been assimilated and shaped society, with real advances and benefits, along with real threats, for instance to employment. I’m interested in the duality, the simultaneous and seemingly conflicting features of that experience.”

For all their impressive capabilities, large language models (LLMs) often fall short when given challenging new tasks that require complex reasoning skills.

While an accounting firm’s LLM might excel at summarizing financial reports, that same model could fail unexpectedly if tasked with predicting market trends or identifying fraudulent transactions.

To make LLMs more adaptable, MIT researchers investigated how a certain training technique can be strategically deployed to boost a model’s performance on unfamiliar, difficult problems.

They show that test-time training, a method that involves temporarily updating some of a model’s inner workings during deployment, can lead to a sixfold improvement in accuracy. The researchers developed a framework for implementing a test-time training strategy that uses examples of the new task to maximize these gains.

Their work could improve a model’s flexibility, enabling an off-the-shelf LLM to adapt to complex tasks that require planning or abstraction. This could lead to LLMs that would be more accurate in many applications that require logical deduction, from medical diagnostics to supply chain management.

“Genuine learning — what we did here with test-time training — is something these models can’t do on their own after they are shipped. They can’t gain new skills or get better at a task. But we have shown that if you push the model a little bit to do actual learning, you see that huge improvements in performance can happen,” says Ekin Akyürek PhD ’25, lead author of the study.

Akyürek is joined on the paper by graduate students Mehul Damani, Linlu Qiu, Han Guo, and Jyothish Pari; undergraduate Adam Zweiger; and senior authors Yoon Kim, an assistant professor of Electrical Engineering and Computer Science (EECS) and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Jacob Andreas, an associate professor in EECS and a member of CSAIL. The research will be presented at the International Conference on Machine Learning.

Tackling hard domains

LLM users often try to improve the performance of their model on a new task using a technique called in-context learning. They feed the model a few examples of the new task as text prompts which guide the model’s outputs.

But in-context learning doesn’t always work for problems that require logic and reasoning.

The MIT researchers investigated how test-time training can be used in conjunction with in-context learning to boost performance on these challenging tasks. Test-time training involves updating some model parameters — the internal variables it uses to make predictions — using a small amount of new data specific to the task at hand.

The researchers explored how test-time training interacts with in-context learning. They studied design choices that maximize the performance improvements one can coax out of a general-purpose LLM.

“We find that test-time training is a much stronger form of learning. While simply providing examples can modestly boost accuracy, actually updating the model with those examples can lead to significantly better performance, particularly in challenging domains,” Damani says.

In-context learning requires a small set of task examples, including problems and their solutions. The researchers use these examples to create a task-specific dataset needed for test-time training.

To expand the size of this dataset, they create new inputs by slightly changing the problems and solutions in the examples, such as by horizontally flipping some input data. They find that training the model on the outputs of this new dataset leads to the best performance.

In addition, the researchers only update a small number of model parameters using a technique called low-rank adaption, which improves the efficiency of the test-time training process.

“This is important because our method needs to be efficient if it is going to be deployed in the real world. We find that you can get huge improvements in accuracy with a very small amount of parameter training,” Akyürek says.

Developing new skills

Streamlining the process is key, since test-time training is employed on a per-instance basis, meaning a user would need to do this for each individual task. The updates to the model are only temporary, and the model reverts to its original form after making a prediction.

A model that usually takes less than a minute to answer a query might take five or 10 minutes to provide an answer with test-time training, Akyürek adds.

“We wouldn’t want to do this for all user queries, but it is useful if you have a very hard task that you want to the model to solve well. There also might be tasks that are too challenging for an LLM to solve without this method,” he says.

The researchers tested their approach on two benchmark datasets of extremely complex problems, such as IQ puzzles. It boosted accuracy as much as sixfold over techniques that use only in-context learning.

Tasks that involved structured patterns or those which used completely unfamiliar types of data showed the largest performance improvements.

“For simpler tasks, in-context learning might be OK. But updating the parameters themselves might develop a new skill in the model,” Damani says.

In the future, the researchers want to use these insights toward the development of models that continually learn.

The long-term goal is an LLM that, given a query, can automatically determine if it needs to use test-time training to update parameters or if it can solve the task using in-context learning, and then implement the best test-time training strategy without the need for human intervention.

This work is supported, in part, by the MIT-IBM Watson AI Lab and the National Science Foundation.

During photosynthesis, an enzyme called rubisco catalyzes a key reaction — the incorporation of carbon dioxide into organic compounds to create sugars. However, rubisco, which is believed to be the most abundant enzyme on Earth, is very inefficient compared to the other enzymes involved in photosynthesis.

MIT chemists have now shown that they can greatly enhance a version of rubisco found in bacteria from a low-oxygen environment. Using a process known as directed evolution, they identified mutations that could boost rubisco’s catalytic efficiency by up to 25 percent.

The researchers now plan to apply their technique to forms of rubisco that could be used in plants to help boost their rates of photosynthesis, which could potentially improve crop yields.

“This is, I think, a compelling demonstration of successful improvement of a rubisco’s enzymatic properties, holding out a lot of hope for engineering other forms of rubisco,” says Matthew Shoulders, the Class of 1942 Professor of Chemistry at MIT.

Shoulders and Robert Wilson, a research scientist in the Department of Chemistry, are the senior authors of the new study, which appears this week in the Proceedings of the National Academy of Sciences. MIT graduate student Julie McDonald is the paper’s lead author.

Evolution of efficiency

When plants or photosynthetic bacteria absorb energy from the sun, they first convert it into energy-storing molecules such as ATP. In the next phase of photosynthesis, cells use that energy to transform a molecule known as ribulose bisphosphate into glucose, which requires several additional reactions. Rubisco catalyzes the first of those reactions, known as carboxylation. During that reaction, carbon from CO₂ is added to ribulose bisphosphate.

Compared to the other enzymes involved in photosynthesis, rubisco is very slow, catalyzing only one to 10 reactions per second. Additionally, rubisco can also interact with oxygen, leading to a competing reaction that incorporates oxygen instead of carbon — a process that wastes some of the energy absorbed from sunlight.

“For protein engineers, that’s a really attractive set of problems because those traits seem like things that you could hopefully make better by making changes to the enzyme’s amino acid sequence,” McDonald says.

Previous research has led to improvement in rubisco’s stability and solubility, which resulted in small gains in enzyme efficiency. Most of those studies used directed evolution — a technique in which a naturally occurring protein is randomly mutated and then screened for the emergence of new, desirable features.

This process is usually done using error-prone PCR, a technique that first generates mutations in vitro (outside of the cell), typically introducing only one or two mutations in the target gene. In past studies on rubisco, this library of mutations was then introduced into bacteria that grow at a rate relative to rubisco activity. Limitations in error-prone PCR and in the efficiency of introducing new genes restrict the total number of mutations that can be generated and screened using this approach. Manual mutagenesis and selection steps also add more time to the process over multiple rounds of evolution.

The MIT team instead used a newer mutagenesis technique that the Shoulders Lab previously developed, called MutaT7. This technique allows the researchers to perform both mutagenesis and screening in living cells, which dramatically speeds up the process. Their technique also enables them to mutate the target gene at a higher rate.

“Our continuous directed evolution technique allows you to look at a lot more mutations in the enzyme than has been done in the past,” McDonald says.

Better rubisco

For this study, the researchers began with a version of rubisco, isolated from a family of semi-anaerobic bacteria known as Gallionellaceae, that is one of the fastest rubisco found in nature. During the directed evolution experiments, which were conducted in E. coli, the researchers kept the microbes in an environment with atmospheric levels of oxygen, creating evolutionary pressure to adapt to oxygen.

After six rounds of directed evolution, the researchers identified three different mutations that improved the rubisco’s resistance to oxygen. Each of these mutations are located near the enzyme’s active site (where it performs carboxylation or oxygenation). The researchers believe that these mutations improve the enzyme’s ability to preferentially interact with carbon dioxide over oxygen, which leads to an overall increase in carboxylation efficiency.

“The underlying question here is: Can you alter and improve the kinetic properties of rubisco to operate better in environments where you want it to operate better?” Shoulders says. “What changed through the directed evolution process was that rubisco began to like to react with oxygen less. That allows this rubisco to function well in an oxygen-rich environment, where normally it would constantly get distracted and react with oxygen, which you don’t want it to do.”

In ongoing work, the researchers are applying this approach to other forms of rubisco, including rubisco from plants. Plants are believed to lose about 30 percent of the energy from the sunlight they absorb through a process called photorespiration, which occurs when rubisco acts on oxygen instead of carbon dioxide.

“This really opens the door to a lot of exciting new research, and it’s a step beyond the types of engineering that have dominated rubisco engineering in the past,” Wilson says. “There are definite benefits to agricultural productivity that could be leveraged through a better rubisco.”

The research was funded, in part, by the National Science Foundation, the National Institutes of Health, an Abdul Latif Jameel Water and Food Systems Lab Grand Challenge grant, and a Martin Family Society Fellowship for Sustainability.

The MIT Health and Life Sciences Collaborative (MIT HEALS) is launching the Biswas Postdoctoral Fellowship Program to advance the work of outstanding early-career researchers in health and life sciences. Supported by a gift from the Biswas Family Foundation, the program aims to help apply cutting-edge research to improve health care and the lives of millions.

The program will support exceptional postdocs dedicated to innovation in human health care through a full range of pathways, such as leveraging AI in health-related research, developing low-cost diagnostics, and the convergence of life sciences with such areas as economics, business, policy, or the humanities. With initial funding of $12 million, five four-year fellowships will be awarded for each of the next four years, starting in early 2026.

“An essential goal of MIT HEALS is to find new ways and opportunities to deliver health care solutions at scale, and the Biswas Family Foundation shares our commitment to scalable innovation and broad impact. MIT is also in the talent business, and the foundation’s gift allows us to bring exceptional scholars to campus to explore some of the most pressing issues in human health and build meaningful connections across academia and industry. We look forward to welcoming the first cohort of Biswas Fellows to MIT,” says MIT president Sally Kornbluth.

“We are deeply honored to launch this world-class postdoctoral fellows program,” adds Anantha P. Chandrakasan, MIT’s chief innovation and strategy officer and head of MIT HEALS. “We fully expect to attract top candidates from around the globe to lead innovative cross-cutting projects in AI and health, cancer therapies, diagnostics, and beyond. These fellows will be selected through a rigorous process overseen by a distinguished committee, and will have the opportunity to collaborate with our faculty on the most promising and impactful ideas.”

Angela Koehler, faculty lead of MIT HEALS, professor in MIT’s Department of Biological Engineering, and associate director of the Koch Institute for Integrative Cancer Research, emphasized that the objectives of MIT HEALS align well with a stated goal of the Biswas Family Foundation: to leverage “scientific and technological advancements to revolutionize health care and make a lasting impact on global public health.”

“Health care is a team sport,” Koehler says. “MIT HEALS seeks to create connections involving investigators with diverse expertise across the Institute to tackle the most transformative problems impacting human health. Members of the MIT community are well poised to participate in teams and make an impact.”

MIT HEALS also seeks to maximize its effectiveness by expanding collaboration with medical schools and hospitals, starting with defining important problems that can be approached through research, and continuing all the way to clinical studies, Koehler says.

The Biswas Family Foundation has already demonstrated a similar strategy.

“The Biswas family has a history of enabling connections and partnerships between institutions that each bring a piece to the puzzle,” Koehler says. “This could be a dataset, an algorithm, an agent, a technology platform, or patients.”

Hope Biswas, co-founder of the Biswas Family Foundation with her husband, MIT alumnus Sanjit Biswas SM ’05, also highlighted the synergies between the foundation and MIT.

“The Biswas Family Foundation is proud to support the MIT HEALS initiative, which reimagines how scientific discovery can translate into real-world health impact. Its focus on promoting interdisciplinary collaboration to find new solutions to challenges in health care aligns closely with our mission to advance science and technology to improve health outcomes at scale,” Biswas says.

“As part of this commitment,” Biswas adds, “we are especially proud to support outstanding postdoctoral scholars focused on high-impact cross-disciplinary work in fields such as computational biology, nanoscale therapeutics, women’s health, and fundamental, curiosity-driven life sciences research. We are excited to contribute to an effort that brings together cutting-edge science and a deep commitment to translating knowledge into action.”

AI and machine-learning systems present a new universe of opportunities to investigate disease, biological mechanisms, therapeutics, and health care delivery using huge datasets.

“AI and computational systems biology can improve the accuracy of diagnostic approaches, enable the development of precision medicines, improve choices related to individualized treatment strategy, and improve operational efficiency within health care systems,” says Koehler. “Sanjit and Hope’s support of broad initiatives in AI and computational systems biology will help MIT researchers explore a variety of paths to impact human health on a large scale.”

Frontiers in health-related research are increasingly found where diverse fields converge, and Koehler provides the example of how advances in high-throughput experimentation to develop large datasets “may couple well with the development of new computation or AI tools.” She adds that the four-year funding term provided by the postdoctoral fellowship is “long enough to enable fellows to think big and take on projects at interfaces, emerging as bilingual researchers at the end of the program.”

Chandrakasan sees potential in the program for the Biswas Fellows to make revolutionary progress in health research.

“I’m incredibly grateful to the Biswas Family Foundation for their generous support in enabling transformative research at MIT,” Chandrakasan says.

Lanthanides are a class of rare earth elements that in many countries are added to fertilizer as micronutrients to stimulate plant growth. But little is known about how they are absorbed by plants or influence photosynthesis, potentially leaving their benefits untapped.

Now, researchers from MIT have shed light on how lanthanides move through and operate within plants. These insights could help farmers optimize their use to grow some of the world’s most popular crops.

Published today in the Journal of the American Chemical Society, the study shows that a single nanoscale dose of lanthanides applied to seeds can make some of the world’s most common crops more resilient to UV stress. The researchers also uncovered the chemical processes by which lanthanides interact with the chlorophyll pigments that drive photosynthesis, showing that different lanthanide elements strengthen chlorophyll by replacing the magnesium at its center.

“This is a first step to better understand how these elements work in plants, and to provide an example of how they could be better delivered to plants, compared to simply applying them in the soil,” says Associate Professor Benedetto Marelli, who conducted the research with postdoc Giorgio Rizzo. “This is the first example of a thorough study showing the effects of lanthanides on chlorophyll, and their beneficial effects to protect plants from UV stress.”

Inside plant connections

Certain lanthanides are used as contrast agents in MRI and for applications including light-emitting diodes, solar cells, and lasers. Over the last 50 years, lanthanides have become increasingly used in agriculture to enhance crop yields, with China alone applying lanthanide-based fertilizers to nearly 4 million hectares of land each year.

“Lanthanides have been considered for a long time to be biologically irrelevant, but that’s changed in agriculture, especially in China,” says Rizzo, the paper’s first author. “But we largely don’t know how lanthanides work to benefit plants — nor do we understand their uptake mechanisms from plant tissues.”

Recent studies have shown that low concentrations of lanthanides can promote plant growth, root elongation, hormone synthesis, and stress tolerance, but higher doses can cause harm to plants. Striking the right balance has been hard because of our lack of understanding around how lanthanides are absorbed by plants or how they interact with root soil.

For the study, the researchers leveraged seed coating and treatment technologies they previously developed to investigate the way the plant pigment chlorophyll interacts with lanthanides, both inside and outside of plants. Up until now, researchers haven’t been sure whether chlorophyll interacts with lanthanide ions at all.

Chlorophyll drives photosynthesis, but the pigments lose their ability to efficiently absorb light when the magnesium ion at their core is removed. The researchers discovered that lanthanides can fill that void, helping chlorophyll pigments partially recover some of their optical properties in a process known as re-greening.

“We found that lanthanides can boost several parameters of plant health,” Marelli says. “They mostly accumulate in the roots, but a small amount also makes its way to the leaves, and some of the new chlorophyll molecules made in leaves have lanthanides incorporated in their structure.”

This study also offers the first experimental evidence that lanthanides can increase plant resilience to UV stress, something the researchers say was completely unexpected.

“Chlorophylls are very sensitive pigments,” Rizzo says. “They can convert light to energy in plants, but when they are isolated from the cell structure, they rapidly hydrolyze and degrade. However, in the form with lanthanides at their center, they are pretty stable, even after extracting them from plant cells.”

The researchers, using different spectroscopic techniques, found the benefits held across a range of staple crops, including chickpea, barley, corn, and soybeans.

The findings could be used to boost crop yield and increase the resilience of some of the world’s most popular crops to extreme weather.

“As we move into an environment where extreme heat and extreme climate events are more common, and particularly where we can have prolonged periods of sun in the field, we want to provide new ways to protect our plants,” Marelli says. “There are existing agrochemicals that can be applied to leaves for protecting plants from stressors such as UV, but they can be toxic, increase microplastics, and can require multiple applications. This could be a complementary way to protect plants from UV stress.”

Identifying new applications

The researchers also found that larger lanthanide elements like lanthanum were more effective at strengthening chlorophyll pigments than smaller ones. Lanthanum is considered a low-value byproduct of rare earths mining, and can become a burden to the rare earth element (REE) supply chain due to the need to separate it from more desirable rare earths. Increasing the demand for lanthanum could diversify the economics of REEs and improve the stability of their supply chain, the scientists suggest.

“This study shows what we could do with these lower-value metals,” Marelli says. “We know lanthanides are extremely useful in electronics, magnets, and energy. In the U.S., there’s a big push to recycle them. That’s why for the plant studies, we focused on lanthanum, being the most abundant, cheapest lanthanide ion.”

Moving forward, the team plans to explore how lanthanides work with other biological molecules, including proteins in the human body.

In agriculture, the team hopes to scale up its research to include field and greenhouse studies to continue testing the results of UV resilience on different crop types and in experimental farm conditions.

“Lanthanides are already widely used in agriculture,” Rizzo says. “We hope this study provides evidence that allows more conscious use of them and also a new way to apply them through seed treatments.”

The research was supported by the MIT Climate Grand Challenge and the Office for Naval Research.

Scientists are striving to discover new semiconductor materials that could boost the efficiency of solar cells and other electronics. But the pace of innovation is bottlenecked by the speed at which researchers can manually measure important material properties.

A fully autonomous robotic system developed by MIT researchers could speed things up.

Their system utilizes a robotic probe to measure an important electrical property known as photoconductance, which is how electrically responsive a material is to the presence of light.

The researchers inject materials-science-domain knowledge from human experts into the machine-learning model that guides the robot’s decision making. This enables the robot to identify the best places to contact a material with the probe to gain the most information about its photoconductance, while a specialized planning procedure finds the fastest way to move between contact points.

During a 24-hour test, the fully autonomous robotic probe took more than 125 unique measurements per hour, with more precision and reliability than other artificial intelligence-based methods.

By dramatically increasing the speed at which scientists can characterize important properties of new semiconductor materials, this method could spur the development of solar panels that produce more electricity.

“I find this paper to be incredibly exciting because it provides a pathway for autonomous, contact-based characterization methods. Not every important property of a material can be measured in a contactless way. If you need to make contact with your sample, you want it to be fast and you want to maximize the amount of information that you gain,” says Tonio Buonassisi, professor of mechanical engineering and senior author of a paper on the autonomous system.

His co-authors include lead author Alexander (Aleks) Siemenn, a graduate student; postdocs Basita Das and Kangyu Ji; and graduate student Fang Sheng. The work appears today in Science Advances.

Making contact

Since 2018, researchers in Buonassisi’s laboratory have been working toward a fully autonomous materials discovery laboratory. They’ve recently focused on discovering new perovskites, which are a class of semiconductor materials used in photovoltaics like solar panels.

In prior work, they developed techniques to rapidly synthesize and print unique combinations of perovskite material. They also designed imaging-based methods to determine some important material properties.

But photoconductance is most accurately characterized by placing a probe onto the material, shining a light, and measuring the electrical response.

“To allow our experimental laboratory to operate as quickly and accurately as possible, we had to come up with a solution that would produce the best measurements while minimizing the time it takes to run the whole procedure,” says Siemenn.

Doing so required the integration of machine learning, robotics, and material science into one autonomous system.

To begin, the robotic system uses its onboard camera to take an image of a slide with perovskite material printed on it.

Then it uses computer vision to cut that image into segments, which are fed into a neural network model that has been specially designed to incorporate domain expertise from chemists and materials scientists.

“These robots can improve the repeatability and precision of our operations, but it is important to still have a human in the loop. If we don’t have a good way to implement the rich knowledge from these chemical experts into our robots, we are not going to be able to discover new materials,” Siemenn adds.

The model uses this domain knowledge to determine the optimal points for the probe to contact based on the shape of the sample and its material composition. These contact points are fed into a path planner that finds the most efficient way for the probe to reach all points.

The adaptability of this machine-learning approach is especially important because the printed samples have unique shapes, from circular drops to jellybean-like structures.

“It is almost like measuring snowflakes — it is difficult to get two that are identical,” Buonassisi says.

Once the path planner finds the shortest path, it sends signals to the robot’s motors, which manipulate the probe and take measurements at each contact point in rapid succession.

Key to the speed of this approach is the self-supervised nature of the neural network model. The model determines optimal contact points directly on a sample image — without the need for labeled training data.

The researchers also accelerated the system by enhancing the path planning procedure. They found that adding a small amount of noise, or randomness, to the algorithm helped it find the shortest path.

“As we progress in this age of autonomous labs, you really do need all three of these expertise — hardware building, software, and an understanding of materials science — coming together into the same team to be able to innovate quickly. And that is part of the secret sauce here,” Buonassisi says.

Rich data, rapid results

Once they had built the system from the ground up, the researchers tested each component. Their results showed that the neural network model found better contact points with less computation time than seven other AI-based methods. In addition, the path planning algorithm consistently found shorter path plans than other methods.

When they put all the pieces together to conduct a 24-hour fully autonomous experiment, the robotic system conducted more than 3,000 unique photoconductance measurements at a rate exceeding 125 per hour.

In addition, the level of detail provided by this precise measurement approach enabled the researchers to identify hotspots with higher photoconductance as well as areas of material degradation.

“Being able to gather such rich data that can be captured at such fast rates, without the need for human guidance, starts to open up doors to be able to discover and develop new high-performance semiconductors, especially for sustainability applications like solar panels,” Siemenn says.

The researchers want to continue building on this robotic system as they strive to create a fully autonomous lab for materials discovery.

This work is supported, in part, by First Solar, Eni through the MIT Energy Initiative, MathWorks, the University of Toronto’s Acceleration Consortium, the U.S. Department of Energy, and the U.S. National Science Foundation.

In 1954, the world’s first successful organ transplant took place at Brigham and Women’s Hospital, in the form of a kidney donated from one twin to the other. At the time, a group of doctors and scientists had correctly theorized that the recipient’s antibodies were unlikely to reject an organ from an identical twin. One Nobel Prize and a few decades later, advancements in immune-suppressing drugs increased the viability of and demand for organ transplants. Today, over 1 million organ transplants have been performed in the United States, more than any other country in the world.

The impressive scale of this achievement was made possible due to advances in organ matching systems: The first computer-based organ matching system was released in 1977. Despite continued innovation in computing, medicine, and matching technology over the years, over 100,000 people in the U.S. are currently on the national transplant waiting list and 13 people die each day waiting for an organ transplant. 

Most computational research in organ allocation is focused on the initial stages, when waitlisted patients are being prioritized for organ transplants. In a new paper presented at ACM Conference on Fairness, Accountability, and Transparency (FAccT) in Athens, Greece, researchers from MIT and Massachusetts General Hospital focused on the final, less-studied stage: organ offer acceptance, when an offer is made and the physician at the transplant center decides on behalf of the patient whether to accept or reject the offered organ. 

“I don’t think we were terribly surprised, but we were obviously disappointed,” co-first author and MIT PhD student Hammaad Adam says. Using computational models to analyze transplantation data from over 160,000 transplant candidates in the Scientific Registry of Transplant Recipients (SRTR) between 2010 and 2020, the researchers found that physicians were overall less likely to accept liver and lung offers on behalf of Black candidates, resulting in additional barriers for Black patients in the organ offer acceptance process.  

For livers, Black patients had 7 percent lower odds of offer acceptance than white patients. When it came to lungs, the disparity became even larger, with 20 percent lower odds of having an offer acceptance than white patients with similar characteristics.

The data don’t necessarily point to clinician bias as the main influence. “The bigger takeaway is that even if there are factors that justify clinical decision-making, there could be clinical conditions that we didn’t control for, that are more common for Black patients,” Adam explains. If the wait-list fails to account for certain patterns in decision-making, they could create obstacles in the process even if the process itself is “unbiased.”

The researchers also point out that high variability in offer acceptance and risk tolerances among transplant centers is a potential factor complicating the decision-making process. Their FAccT paper references a 2020 paper published in JAMA Cardiology, which concluded that wait-list candidates listed at transplant centers with lower offer acceptance rates have a higher likelihood of mortality. 

Another key finding was that an offer was more likely to be accepted if the donor and candidate were of the same race. The paper describes this trend as “concerning,” given the historical inequities in organ procurement that have limited donation from racial and ethnic minority groups. 

Previous work from Adam and his collaborators has aimed to address this gap. Last year, they compiled and released Organ Retrieval and Collection of Health Information for Donation (ORCHID), the first multi-center dataset describing the performance of organ procurement organizations (OPOs). ORCHID contains 10 years’ worth of OPO data, and is intended to facilitate research that addresses bias in organ procurement.

“Being able to do good work in this field takes time,” says Adam, who notes that the entirety of the organ offer acceptance project took years to complete. To his knowledge, only one paper to date studies the association between offer acceptance and race. 

While the bureaucratic and highly interdisciplinary nature of clinical AI projects can dissuade computer science graduate students from pursuing them, Adam committed to the project for the duration of his PhD in the lab of associate professor of electrical engineering Marzyeh Ghassemi, an affiliate of the MIT Jameel Clinic and the Institute of Medical Engineering and Sciences.

To graduate students interested in pursuing clinical AI research projects, Adam recommends that they “free [themselves] from the cycle of publishing every four months.”

“I found it freeing, to be honest — it’s OK if these collaborations take a while,” he says. “It’s hard to avoid that. I made the conscious choice a few years ago and I was happy doing that work.”

This work was supported with funding from the MIT Jameel Clinic. It was also supported, in part, by Takeda Development Center Americas Inc. (successor in interest to Millennium Pharmaceuticals Inc.), an NIH Ruth L. Kirschstein National Research Service Award, a CIFAR AI Chair at the Vector Institute, and by the National Institutes of Health.

Incoming information from the retina is channeled into two pathways in the brain’s visual system: one that’s responsible for processing color and fine spatial detail, and another that’s involved in spatial localization and detecting high temporal frequencies. A new study from MIT provides an account for how these two pathways may be shaped by developmental factors.

Newborns typically have poor visual acuity and poor color vision because their retinal cone cells are not well-developed at birth. This means that early in life, they are seeing blurry, color-reduced imagery. The MIT team proposes that such blurry, color-limited vision may result in some brain cells specializing in low spatial frequencies and low color tuning, corresponding to the so-called magnocellular system. Later, with improved vision, cells may tune to finer details and richer color, consistent with the other pathway, known as the parvocellular system.

To test their hypothesis, the researchers trained computational models of vision on a trajectory of input similar to what human babies receive early in life — low-quality images early on, followed by full-color, sharper images later. They found that these models developed processing units with receptive fields exhibiting some similarity to the division of magnocellular and parvocellular pathways in the human visual system. Vision models trained on only high-quality images did not develop such distinct characteristics.

“The findings potentially suggest a mechanistic account of the emergence of the parvo/magno distinction, which is one of the key organizing principles of the visual pathway in the mammalian brain,” says Pawan Sinha, an MIT professor of brain and cognitive sciences and the senior author of the study.

MIT postdocs Marin Vogelsang and Lukas Vogelsang are the lead authors of the study, which appears today in the journal Communications Biology. Sidney Diamond, an MIT research affiliate, and Gordon Pipa, a professor of neuroinformatics at the University of Osnabrueck, are also authors of the paper.

Sensory input

The idea that low-quality visual input might be beneficial for development grew out of studies of children who were born blind but later had their sight restored. An effort from Sinha’s laboratory, Project Prakash, has screened and treated thousands of children in India, where reversible forms of vision loss such as cataracts are relatively common. After their sight is restored, many of these children volunteer to participate in studies in which Sinha and his colleagues track their visual development.

In one of these studies, the researchers found that children who had cataracts removed exhibited a marked drop in object-recognition performance when the children were presented with black and white images, compared to colored ones. Those findings led the researchers to hypothesize that reduced color input characteristic of early typical development, far from being a hindrance, allows the brain to learn to recognize objects even in images that have impoverished or shifted colors.

“Denying access to rich color at the outset seems to be a powerful strategy to build in resilience to color changes and make the system more robust against color loss in images,” Sinha says.

In that study, the researchers also found that when computational models of vision were initially trained on grayscale images, followed by color images, their ability to recognize objects was more robust than that of models trained only on color images. Similarly, another study from the lab found that models performed better when they were trained first on blurry images, followed by sharper images.

To build on those findings, the MIT team wanted to explore what might be the consequences of both of those features — color and visual acuity — being limited at the outset of development. They hypothesized that these limitations might contribute to the development of the magnocellular and parvocellular pathways.

In addition to being highly attuned to color, cells in the parvocellular pathway have small receptive fields, meaning that they receive input from more compact clusters of retinal ganglion cells. This helps them to process fine detail. Cells in the magnocellular pathway pool information across larger areas, allowing them to process more global spatial information.

To test their hypothesis that developmental progressions could contribute to the magno and parvo cell selectivities, the researchers trained models on two different sets of images. One model was presented with a standard dataset of images that are used to train models to categorize objects. The other dataset was designed to roughly mimic the input that the human visual system receives from birth. This “biomimetic” data consists of low-resolution, grayscale images in the first half of the training, followed by high-resolution, colorful images in the second half.

After the models were trained, the researchers analyzed the models’ processing units — nodes within the network that bear some resemblance to the clusters of cells that process visual information in the brain. They found that the models trained on the biomimetic data developed a distinct subset of units that are jointly responsive to low-color and low-spatial-frequency inputs, similar to the magnocellular pathway. Additionally, these biomimetic models exhibited groups of more heterogenous parvocellular-like units tuned predominantly to higher spatial frequencies or richer color signals. Such distinction did not emerge in the models trained on full color, high-resolution images from the start.

“This provides some support for the idea that the ‘correlation’ we see in the biological system could be a consequence of the types of inputs that are available at the same time in normal development,” Lukas Vogelsang says.

Object recognition

The researchers also performed additional tests to reveal what strategies the differently trained models were using for object recognition tasks. In one, they asked the models to categorize images of objects where the shape and texture did not match — for example, an animal with the shape of cat but the texture of an elephant.

This is a technique several researchers in the field have employed to determine which image attributes a model is using to categorize objects: the overall shape or the fine-grained textures. The MIT team found that models trained on biomimetic input were markedly more likely to use an object’s shape to make those decisions, just as humans usually do. Moreover, when the researchers systematically removed the magnocellular-like units from the models, the models quickly lost their tendency to use shape to make categorizations.

In another set of experiments, the researchers trained the models on videos instead of images, which introduces a temporal dimension. In addition to low spatial resolution and color sensitivity, the magnocellular pathway responds to high temporal frequencies, allowing it to quickly detect changes in the position of an object. When models were trained on biomimetic video input, the units most tuned to high temporal frequencies were indeed the ones that also exhibited magnocellular-like properties in the spatial domain.

Overall, the results support the idea that low-quality sensory input early in life may contribute to the organization of sensory processing pathways of the brain, the researchers say. The findings do not rule out innate specification of the magno and parvo pathways, but provide a proof of principle that visual experience over the course of development could also play a role.

“The general theme that seems to be emerging is that the developmental progression that we go through is very carefully structured in order to give us certain kinds of perceptual proficiencies, and it may also have consequences in terms of the very organization of the brain,” Sinha says.

The research was funded by the National Institutes of Health, the Simons Center for the Social Brain, the Japan Society for the Promotion of Science, and the Yamada Science Foundation.

Being placed in foster care is a necessary intervention for some children. But many advocates worry that kids can languish in foster care too long, with harmful effects for children who are temporarily unattached from a permanent family.

A new study co-authored by an MIT economist shows that an innovative Chilean program providing legal aid to children shortens the length of foster-care stays, returning them to families faster. In the process, it improves long-term social outcomes for kids and even reduces government spending on the foster care system.

“It was amazingly successful because the program got kids out of foster care about 30 percent faster,” says Joseph Doyle, an economist at the MIT Sloan School of Management, who helped lead the research. “Because foster care is expensive, that paid for the program by itself about four times over. If you improve the case management of kids in foster care, you can improve a child’s well-being and save money.”

The paper, “Effects of Enhanced Legal Aid in Child Welfare: Evidence from a Randomized Trial of Mi Abogado,” is published in the American Economic Review.

The authors are Ryan Cooper, a professor and director of government innovation at the University of Chicago; Doyle, who is the Erwin H. Schell Professor of Management at MIT Sloan; and Andrés P. Hojman, a professor at the Pontifical Catholic University of Chile.

Rigorous design

To conduct the study, the scholars examined the Chilean government’s new program “Mi Abogado” — meaning, “My Lawyer” — which provided enhanced legal support to children in foster care, as well as access to psychologists and social workers. Legal advocates in the program were given a reduced caseload, for one thing, to help them focus further on each individual case.

Chile introduced Mi Abogado in 2017, with a feature that made it ripe for careful study: The program randomizes most of the participants selected, as part of how it was rolled out. From the pool of children in the foster care system, randomly being part of the program makes it easier to identify its causal impact on later outcomes.

“Very few foster-care redesigns are evaluated in such a rigorous way, and we need more of this innovative approach to policy improvement,” Doyle notes.

The experiment included 1,781 children who were in Chile’s foster care program in 2019, with 581 selected for the Mi Abogado services; it tracked their trajectories over more than two years. Almost all the participants were in group foster-care homes.

In addition to reduced time spent in foster care, the Chilean data showed that children in the Mi Abogado program had a subsequent 30 percent reduction in terms of contact with the criminal justice system and a 5 percent increase in school attendance, compared to children in foster care who did not participate in the program.

“They were getting involved with crime less and attending school more,” Doyle says.

As powerful as the results appear, Doyle acknowledges that he would like to be able to analyze further which elements of the Mi Abogado program had the biggest impact — legal help, counseling and therapy, or other factors.

“We would like to see more about what exactly they are doing for children to speed their exit from care,” Doyle says. “Is it mostly about therapy? Is it working with judges and cutting through red tape? We think the lawyer is a very important part. But the results suggest it is not just the lawyer that improves outcomes.”

More programs in other places?

The current paper is one of many studies Doyle has developed during his career that relate to foster care and related issues. In another forthcoming paper, Doyle and some co-authors find that about 5 percent of U.S. children spend some time in foster care — a number that appears to be fairly common internationally, too.

“People don’t appreciate how common child protective services and foster care are,” Doyle says. Moreover, he adds, “Children involved in these systems are particularly vulnerable.”

With a variety of U.S. jurisdictions running their own foster-care systems, Doyle notes that many people have the opportunity to usefully learn about the Mi Abogado program and consider if its principles might be worth testing. And while that requires some political will, Doyle expresses optimism that policymakers might be open to new ideas.

“It’s not really a partisan issue,” Doyle says. “Most people want to help protect kids, and, if an intervention is needed for kids, have an interest in making the intervention run well.”

After all, he notes, the impact of the Mi Abogado program appears to be both substantial and lasting, making it an interesting example to consider.

“Here we have a case where the child outcomes are improved and the government saved money,” Doyle observes. “I’d like to see more experimentation with programs like this in other places.”

Support for the research was provided in part by the MIT Sloan Latin America Office. Chile’s Studies Department of the Ministry of Education made data available from the education system.

Often when we listen to music, we just instinctually enjoy it. Sometimes, though, it’s worth dissecting a song or other composition to figure out how it’s built.

Take the 1953 jazz standard “Satin Doll,” written by Duke Ellington and Billy Strayhorn, whose subtle structure rewards a close listening. As it happens, MIT Professor Emeritus Samuel Jay Keyser, a distinguished linguist and an avid trombonist on the side, has given the song careful scrutiny.

To Keyser, “Satin Doll” is a glittering example of what he calls the “same/except” construction in art. A basic rhyme, like “rent” and “tent,” is another example of this construction, given the shared rhyming sound and the different starting consonants.

In “Satin Doll,” Keyser observes, both the music and words feature a “same/except” structure. For instance, the rhythm of the first two bars of “Satin Doll” is the same as the second two bars, but the pitch goes up a step in bars three and four. An intricate pattern of this prevails throughout the entire body of “Satin Doll,” which Keyser calls “a musical rhyme scheme.”

When lyricist Johnny Mercer wrote words for “Satin Doll,” he matched the musical rhyme scheme. One lyric for the first four bars is, “Cigarette holder / which wigs me / Over her shoulder / she digs me.” Other verses follow the same pattern.

“Both the lyrics and the melody have the same rhyme scheme in their separate mediums, words and music, namely, A-B-A-B,” says Keyser. “That’s how you write lyrics. If you understand the musical rhyme scheme, and write lyrics to match that, you are introducing a whole new level of repetition, one that enhances the experience.”

Now, Keyser has a new book out about repetition in art and its cognitive impact on us, scrutinizing “Satin Doll” along with many other works of music, poetry, painting, and photography. The volume, “Play It Again, Sam: Repetition in the Arts,” is published by the MIT Press. The title is partly a play on Keyser’s name.

Inspired by the Margulis experiment

The genesis of “Play It Again, Sam” dates back several years, when Keyser encountered an experiment conducted by musicologist Elizabeth Margulis, described in her 2014 book, “On Repeat.” Margulis found that when she altered modern atonal compositions to add repetition to them, audiences ranging from ordinary listeners to music theorists preferred these edited versions to the original works.

“The Margulis experiment really caused the ideas to materialize,” Keyser says. He then examined repetition across art forms that featured research on associated cognitive activity, especially music, poetry, and the visual arts. For instance, the brain has distinct locations dedicated to the recognition of faces, places, and bodies. Keyser suggests this is why, prior to the advent of modernism, painting was overwhelmingly mimetic.

Ideally, he suggests, it will be possible to more comprehensively study how our brains process art — to see if encountering repetition triggers an endorphin release, say. For now, Keyser postulates that repetition involves what he calls the 4 Ps: priming, parallelism, prediction, and pleasure. Essentially, hearing or seeing a motif sets the stage for it to be repeated, providing audiences with satisfaction when they discover the repetition.

With remarkable range, Keyser vigorously analyzes how artists deploy repetition and have thought about it, from “Beowulf” to Leonard Bernstein, from Gustave Caillebotte to Italo Calvino. Some artworks do deploy identical repetition of elements, such as the Homeric epics; others use the “same/except” technique.

Keyser is deeply interested in visual art displaying the “same/except” concept, such as Andy Warhol’s famous “Campbell Soup Cans” painting. It features four rows of eight soup cans, which are all the same — except for the kind of soup on each can.

“Discovering this ‘same/except’ repetition in a work of art brings pleasure,” Keyser says.

But why is this? Multiple experimental studies, Keyser notes, suggest that repeated exposure of a subject to an image — such as an infant’s exposure to its mother’s face — helps create a bond of affection. This is the “mere exposure” phenomenon, posited by social psychologist Robert Zajonc, who as Keyser notes in the book, studied in detail “the repetition of an arbitrary stimulus and the mild affection that people eventually have for it.”

This tendency also helps explain why product manufacturers create ads with just the name of their products in ads: Seen often enough, the viewer bonds with the name. However the mechanism connecting repetition with pleasure works, and whatever its original function, Keyser argues that many artists have successfully tapped into it, grasping that audiences like repetition in poetry, painting, and music.

A shadow dog in Albuquerque

In the book, Keyser’s emphasis on repetition generates some distinctive interpretive positions. In one chapter, he digs into Lee Friendlander’s well-known photo, “Albuquerque, New Mexico,” a street scene with a jumble of signs, wires, and buildings, often interpreted in symbolic terms: It’s the American West frontier being submerged under postwar concrete and commerce.

Keyser, however, has a really different view of the Friendlander photo. There is a dog sitting near the middle of it; to the right is the shadow of a street sign. Keyser believes the shadow resembles the dog, and thinks it creates playful repetition in the photo.

“This particular photograph is really two photographs that rhyme,” Keyser says.“They’re the same, except one is the dog and one is the shadow. And that’s why that photograph is pleasurable, because you see that, even if you may not be fully aware of it. Sensing repetition in a work of art brings pleasure.”

“Play It Again, Sam” has received praise from arts practitioners, among others. George Darrah, principal drummer and arranger of the Boston Pops Orchestra, has called the book “extraordinary” in its “demonstration of the ways that poetry, music, painting, and photography engender pleasure in their audiences by exploiting the ability of the brain to detect repetition.” He adds that “Keyser has an uncanny ability to simplify complex ideas so that difficult material is easily understandable.”

In certain ways “Play It Again, Sam” contains the classic intellectual outlook of an MIT linguist. For decades, MIT-linked linguistics research has identified the universal structures of human language, revealing important similarities despite the seemingly wild variation of global languages. And here too, Keyser finds patterns that help organize an apparently boundless world of art. “Play It Again, Sam” is a hunt for structure.

Asked about this, Keyser acknowledges the influence of his longtime field on his current intellectual explorations, while noting that his insights about art are part of a greater investigation into our works and minds.

“I’m bringing a linguistic habit of mind to art,” Keyser says. “But I’m also pointing an analytical lens in the direction of natural predilections of the brain. The idea is to investigate how our aesthetic sense depends on the way the mind works. I’m trying to show how art can exploit the brain’s capacity to produce pleasure from non-art related functions.”

Using an inexpensive electrode coated with DNA, MIT researchers have designed disposable diagnostics that could be adapted to detect a variety of diseases, including cancer or infectious diseases such as influenza and HIV.

These electrochemical sensors make use of a DNA-chopping enzyme found in the CRISPR gene-editing system. When a target such as a cancerous gene is detected by the enzyme, it begins shearing DNA from the electrode nonspecifically, like a lawnmower cutting grass, altering the electrical signal produced.

One of the main limitations of this type of sensing technology is that the DNA that coats the electrode breaks down quickly, so the sensors can’t be stored for very long and their storage conditions must be tightly controlled, limiting where they can be used. In a new study, MIT researchers stabilized the DNA with a polymer coating, allowing the sensors to be stored for up to two months, even at high temperatures. After storage, the sensors were able to detect a prostate cancer gene that is often used to diagnose the disease.

The DNA-based sensors, which cost only about 50 cents to make, could offer a cheaper way to diagnose many diseases in low-resource regions, says Ariel Furst, the Paul M. Cook Career Development Assistant Professor of Chemical Engineering at MIT and the senior author of the study.

“Our focus is on diagnostics that many people have limited access to, and our goal is to create a point-of-use sensor. People wouldn’t even need to be in a clinic to use it. You could do it at home,” Furst says.

MIT graduate student Xingcheng Zhou is the lead author of the paper, published June 30 in the journal ACS Sensors. Other authors of the paper are MIT undergraduate Jessica Slaughter, Smah Riki ’24, and graduate student Chao Chi Kuo.

An inexpensive sensor

Electrochemical sensors work by measuring changes in the flow of an electric current when a target molecule interacts with an enzyme. This is the same technology that glucose meters use to detect concentrations of glucose in a blood sample.

The electrochemical sensors developed in Furst’s lab consist of DNA adhered to an inexpensive gold leaf electrode, which is laminated onto a sheet of plastic. The DNA is attached to the electrode using a sulfur-containing molecule known as a thiol.

In a 2021 study, Furst’s lab showed that they could use these sensors to detect genetic material from HIV and human papillomavirus (HPV). The sensors detect their targets using a guide RNA strand, which can be designed to bind to nearly any DNA or RNA sequence. The guide RNA is linked to an enzyme called Cas12, which cleaves DNA nonspecifically when it is turned on and is in the same family of proteins as the Cas9 enzyme used for CRISPR genome editing.

If the target is present, it binds to the guide RNA and activates Cas12, which then cuts the DNA adhered to the electrode. That alters the current produced by the electrode, which can be measured using a potentiostat (the same technology used in handheld glucose meters).

“If Cas12 is on, it’s like a lawnmower that cuts off all the DNA on your electrode, and that turns off your signal,” Furst says.

In previous versions of the device, the DNA had to be added to the electrode just before it was used, because DNA doesn’t remain stable for very long. In the new study, the researchers found that they could increase the stability of the DNA by coating it with a polymer called polyvinyl alcohol (PVA).

This polymer, which costs less than 1 cent per coating, acts like a tarp that protects the DNA below it. Once deposited onto the electrode, the polymer dries to form a protective thin film.

“Once it’s dried, it seems to make a very strong barrier against the main things that can harm DNA, such as reactive oxygen species that can either damage the DNA itself or break the thiol bond with the gold and strip your DNA off the electrode,” Furst says.

Successful detection

The researchers showed that this coating could protect DNA on the sensors for at least two months, and it could also withstand temperatures up to about 150 degrees Fahrenheit. After two months, they rinsed off the polymer and demonstrated that the sensors could still detect PCA3, a prostate cancer gene that can be found in urine.

This type of test could be used with a variety of samples, including urine, saliva, or nasal swabs. The researchers hope to use this approach to develop cheaper diagnostics for infectious diseases, such as HPV or HIV, that could be used in a doctor’s office or at home. This approach could also be used to develop tests for emerging infectious diseases, the researchers say.

A group of researchers from Furst’s lab was recently accepted into delta v, MIT’s student venture accelerator, where they hope to launch a startup to further develop this technology. Now that the researchers can create tests with a much longer shelf-life, they hope to begin shipping them to locations where they could be tested with patient samples.

“Our goal is to continue to test with patient samples against different diseases in real world environments,” Furst says. “Our limitation before was that we had to make the sensors on site, but now that we can protect them, we can ship them. We don’t have to use refrigeration. That allows us to access a lot more rugged or non-ideal environments for testing.”

The research was funded, in part, by the MIT Research Support Committee and a MathWorks Fellowship.

A new imaging technique developed by MIT researchers could enable quality-control robots in a warehouse to peer through a cardboard shipping box and see that the handle of a mug buried under packing peanuts is broken.

Their approach leverages millimeter wave (mmWave) signals, the same type of signals used in Wi-Fi, to create accurate 3D reconstructions of objects that are blocked from view.

The waves can travel through common obstacles like plastic containers or interior walls, and reflect off hidden objects. The system, called mmNorm, collects those reflections and feeds them into an algorithm that estimates the shape of the object’s surface.

This new approach achieved 96 percent reconstruction accuracy on a range of everyday objects with complex, curvy shapes, like silverware and a power drill. State-of-the-art baseline methods achieved only 78 percent accuracy.

In addition, mmNorm does not require additional bandwidth to achieve such high accuracy. This efficiency could allow the method to be utilized in a wide range of settings, from factories to assisted living facilities.

For instance, mmNorm could enable robots working in a factory or home to distinguish between tools hidden in a drawer and identify their handles, so they could more efficiently grasp and manipulate the objects without causing damage.

“We’ve been interested in this problem for quite a while, but we’ve been hitting a wall because past methods, while they were mathematically elegant, weren’t getting us where we needed to go. We needed to come up with a very different way of using these signals than what has been used for more than half a century to unlock new types of applications,” says Fadel Adib, associate professor in the Department of Electrical Engineering and Computer Science, director of the Signal Kinetics group in the MIT Media Lab, and senior author of a paper on mmNorm.

Adib is joined on the paper by research assistants Laura Dodds, the lead author, and Tara Boroushaki, and former postdoc Kaichen Zhou. The research was recently presented at the Annual International Conference on Mobile Systems, Applications and Services.

Reflecting on reflections

Traditional radar techniques send mmWave signals and receive reflections from the environment to detect hidden or distant objects, a technique called back projection.

This method works well for large objects, like an airplane obscured by clouds, but the image resolution is too coarse for small items like kitchen gadgets that a robot might need to identify.

In studying this problem, the MIT researchers realized that existing back projection techniques ignore an important property known as specularity. When a radar system transmits mmWaves, almost every surface the waves strike acts like a mirror, generating specular reflections.

If a surface is pointed toward the antenna, the signal will reflect off the object to the antenna, but if the surface is pointed in a different direction, the reflection will travel away from the radar and won’t be received.

“Relying on specularity, our idea is to try to estimate not just the location of a reflection in the environment, but also the direction of the surface at that point,” Dodds says.

They developed mmNorm to estimate what is called a surface normal, which is the direction of a surface at a particular point in space, and use these estimations to reconstruct the curvature of the surface at that point.

Combining surface normal estimations at each point in space, mmNorm uses a special mathematical formulation to reconstruct the 3D object.

The researchers created an mmNorm prototype by attaching a radar to a robotic arm, which continually takes measurements as it moves around a hidden item. The system compares the strength of the signals it receives at different locations to estimate the curvature of the object’s surface.

For instance, the antenna will receive the strongest reflections from a surface pointed directly at it and weaker signals from surfaces that don’t directly face the antenna.

Because multiple antennas on the radar receive some amount of reflection, each antenna “votes” on the direction of the surface normal based on the strength of the signal it received.

“Some antennas might have a very strong vote, some might have a very weak vote, and we can combine all votes together to produce one surface normal that is agreed upon by all antenna locations,” Dodds says.

In addition, because mmNorm estimates the surface normal from all points in space, it generates many possible surfaces. To zero in on the right one, the researchers borrowed techniques from computer graphics, creating a 3D function that chooses the surface most representative of the signals received. They use this to generate a final 3D reconstruction.

Finer details

The team tested mmNorm’s ability to reconstruct more than 60 objects with complex shapes, like the handle and curve of a mug. It generated reconstructions with about 40 percent less error than state-of-the-art approaches, while also estimating the position of an object more accurately.

Their new technique can also distinguish between multiple objects, like a fork, knife, and spoon hidden in the same box. It also performed well for objects made from a range of materials, including wood, metal, plastic, rubber, and glass, as well as combinations of materials, but it does not work for objects hidden behind metal or very thick walls.

“Our qualitative results really speak for themselves. And the amount of improvement you see makes it easier to develop applications that use these high-resolution 3D reconstructions for new tasks,” Boroushaki says.

For instance, a robot can distinguish between multiple tools in a box, determine the precise shape and location of a hammer’s handle, and then plan to pick it up and use it for a task. One could also use mmNorm with an augmented reality headset, enabling a factory worker to see lifelike images of fully occluded objects.

It could also be incorporated into existing security and defense applications, generating more accurate reconstructions of concealed objects in airport security scanners or during military reconnaissance.

The researchers want to explore these and other potential applications in future work. They also want to improve the resolution of their technique, boost its performance for less reflective objects, and enable the mmWaves to effectively image  through thicker occlusions.

“This work really represents a paradigm shift in the way we are thinking about these signals and this 3D reconstruction process. We’re excited to see how the insights that we’ve gained here can have a broad impact,” Dodds says.

This work is supported, in part, by the National Science Foundation, the MIT Media Lab, and Microsoft.

Imagine that you want to know the plot of a movie, but you only have access to either the visuals or the sound. With visuals alone, you’ll miss all the dialogue. With sound alone, you will miss the action. Understanding our biology can be similar. Measuring one kind of data — such as which genes are being expressed — can be informative, but it only captures one facet of a multifaceted story. For many biological processes and disease mechanisms, the entire “plot” can’t be fully understood without combining data types.

However, capturing both the “visuals and sound” of biological data, such as gene expression and cell structure data, from the same cells requires researchers to develop new approaches. They also have to make sure that the data they capture accurately reflects what happens in living organisms, including how cells interact with each other and their environments.

Whitehead Institute for Biomedical Research and Harvard University researchers have taken on these challenges and developed Perturb-Multimodal (Perturb-Multi), a powerful new approach that simultaneously measures how genetic changes such as turning off individual genes affect both gene expression and cell structure in intact liver tissue. The method, described in Cell on June 12, aims to accelerate discovery of how genes control organ function and disease.

The research team, led by Whitehead Institute Member Jonathan Weissman and then-graduate student in his lab Reuben Saunders, along with Xiaowei Zhuang, the David B. Arnold Professor of Science at Harvard University, and then-postdoc in her lab Will Allen, created a system that can test hundreds of different genetic modifications within a single mouse liver while capturing multiple types of data from the same cells.

“Understanding how our organs work requires looking at many different aspects of cell biology at once,” Saunders says. “With Perturb-Multi, we can see how turning off specific genes changes not just what other genes are active, but also how proteins are distributed within cells, how cellular structures are organized, and where cells are located in the tissue. It’s like having multiple specialized microscopes all focused on the same experiment.”

“This approach accelerates discovery by both allowing us to test the functions of many different genes at once, and then for each gene, allowing us to measure many different functional outputs or cell properties at once — and we do that in intact tissue from animals,” says Zhuang, who is also a Howard Hughes Medical Institute (HHMI) investigator.

A more efficient approach to genetic studies

Traditional genetic studies in mice often turn off one gene and then observe what changes in that gene’s absence to learn about what the gene does. The researchers designed their approach to turn off hundreds of different genes across a single liver, while still only turning off one gene per cell — using what is known as a mosaic approach. This allowed them to study the roles of hundreds of individual genes at once in a single individual. The researchers then collected diverse types of data from cells across the same liver to get a full picture of the consequences of turning off the genes.

“Each cell serves as its own experiment, and because all the cells are in the same animal, we eliminate the variability that comes from comparing different mice,” Saunders says. “Every cell experiences the same physiological conditions, diet, and environment, making our comparisons much more precise.”

“The challenge we faced was that tissues, to perform their functions, rely on thousands of genes, expressed in many different cells, working together. Each gene, in turn, can control many aspects of a cell’s function. Testing these hundreds of genes in mice using current methods would be extremely slow and expensive — near impossible, in practice.” Allen says.

Revealing new biology through combined measurements

The team applied Perturb-Multi to study genetic controls of liver physiology and function. Their study led to discoveries in three important aspects of liver biology: fat accumulation in liver cells — a precursor to liver disease; stress responses; and hepatocyte zonation (how liver cells specialize, assuming different traits and functions, based on their location within the liver).

One striking finding emerged from studying genes that, when disrupted, cause fat accumulation in liver cells. The imaging data revealed that four different genes all led to similar fat droplet accumulation, but the sequencing data showed they did so through three completely different mechanisms.

“Without combining imaging and sequencing, we would have missed this complexity entirely,” Saunders says. “The imaging told us which genes affect fat accumulation, while the sequencing revealed whether this was due to increased fat production, cellular stress, or other pathways. This kind of mechanistic insight could be crucial for developing targeted therapies for fatty liver disease.”

The researchers also discovered new regulators of liver cell zonation. Unexpectedly, the newly discovered regulators include genes involved in modifying the extracellular matrix — the scaffolding between cells. “We found that cells can change their specialized functions without physically moving to a different zone,” Saunders says. “This suggests that liver cell identity is more flexible than previously thought.”

Technical innovation enables new science

Developing Perturb-Multi required solving several technical challenges. The team created new methods for preserving the content of interest in cells — RNA and proteins — during tissue processing, for collecting many types of imaging data and single-cell gene expression data from tissue samples that have been fixed with a preservative, and for integrating multiple types of data from the same cells.

“Overcoming the inherent complexity of biology in living animals required developing new tools that bridge multiple disciplines — including, in this case, genomics, imaging, and AI,” Allen says.

The two components of Perturb-Multi — the imaging and sequencing assays — together, applied to the same tissue, provide insights that are unattainable through either assay alone.

“Each component had to work perfectly while not interfering with the others,” says Weissman, who is also a professor of biology at MIT and an HHMI investigator. “The technical development took considerable effort, but the payoff is a system that can reveal biology we simply couldn’t see before.”

Expanding to new organs and other contexts

The researchers plan to expand Perturb-Multi to other organs, including the brain, and to study how genetic changes affect organ function under different conditions like disease states or dietary changes.

“We’re also excited about using the data we generate to train machine learning models,” adds Saunders. “With enough examples of how genetic changes affect cells, we could eventually predict the effects of mutations without having to test them experimentally — a ‘virtual cell’ that could accelerate both research and drug development.”

“Perturbation data are critical for training such AI models and the paucity of existing perturbation data represents a major hindrance in such ‘virtual cell’ efforts,” Zhuang says. “We hope Perturb-Multi will fill this gap by accelerating the collection of perturbation data.”

The approach is designed to be scalable, with the potential for genome-wide studies that test thousands of genes simultaneously. As sequencing and imaging technologies continue to improve, the researchers anticipate that Perturb-Multi will become even more powerful and accessible to the broader research community.

“Our goal is to keep scaling up. We plan to do genome-wide perturbations, study different physiological conditions, and look at different organs,” says Weissman. “That we can now collect so many types of data from so many cells, at speed, is going to be critical for building AI models like virtual cells, and I think it’s going to help us answer previously unsolvable questions about health and disease.”

Several researchers have taken a broad view of scientific progress over the last 50 years and come to the same troubling conclusion: Scientific productivity is declining. It’s taking more time, more funding, and larger teams to make discoveries that once came faster and cheaper. Although a variety of explanations have been offered for the slowdown, one is that, as research becomes more complex and specialized, scientists must spend more time reviewing publications, designing sophisticated experiments, and analyzing data.

Now, the philanthropically funded research lab FutureHouse is seeking to accelerate scientific research with an AI platform designed to automate many of the critical steps on the path toward scientific progress. The platform is made up of a series of AI agents specialized for tasks including information retrieval, information synthesis, chemical synthesis design, and data analysis.

FutureHouse founders Sam Rodriques PhD ’19 and Andrew White believe that by giving every scientist access to their AI agents, they can break through the biggest bottlenecks in science and help solve some of humanity’s most pressing problems.

“Natural language is the real language of science,” Rodriques says. “Other people are building foundation models for biology, where machine learning models speak the language of DNA or proteins, and that’s powerful. But discoveries aren’t represented in DNA or proteins. The only way we know how to represent discoveries, hypothesize, and reason is with natural language.”

Finding big problems

For his PhD research at MIT, Rodriques sought to understand the inner workings of the brain in the lab of Professor Ed Boyden.

“The entire idea behind FutureHouse was inspired by this impression I got during my PhD at MIT that even if we had all the information we needed to know about how the brain works, we wouldn’t know it because nobody has time to read all the literature,” Rodriques explains. “Even if they could read it all, they wouldn’t be able to assemble it into a comprehensive theory. That was a foundational piece of the FutureHouse puzzle.”

Rodriques wrote about the need for new kinds of large research collaborations as the last chapter of his PhD thesis in 2019, and though he spent some time running a lab at the Francis Crick Institute in London after graduation, he found himself gravitating toward broad problems in science that no single lab could take on.

“I was interested in how to automate or scale up science and what kinds of new organizational structures or technologies would unlock higher scientific productivity,” Rodriques says.

When Chat-GPT 3.5 was released in November 2022, Rodriques saw a path toward more powerful models that could generate scientific insights on their own. Around that time, he also met Andrew White, a computational chemist at the University of Rochester who had been granted early access to Chat-GPT 4. White had built the first large language agent for science, and the researchers joined forces to start FutureHouse.

The founders started out wanting to create distinct AI tools for tasks like literature searches, data analysis, and hypothesis generation. They began with data collection, eventually releasing PaperQA in September 2024, which Rodriques calls the best AI agent in the world for retrieving and summarizing information in scientific literature. Around the same time, they released Has Anyone, a tool that lets scientists determine if anyone has conducted specific experiments or explored specific hypotheses.

“We were just sitting around asking, ‘What are the kinds of questions that we as scientists ask all the time?’” Rodriques recalls.

When FutureHouse officially launched its platform on May 1 of this year, it rebranded some of its tools. Paper QA is now Crow, and Has Anyone is now called Owl. Falcon is an agent capable of compiling and reviewing more sources than Crow. Another new agent, Phoenix, can use specialized tools to help researchers plan chemistry experiments. And Finch is an agent designed to automate data driven discovery in biology.

On May 20, the company demonstrated a multi-agent scientific discovery workflow to automate key steps of the scientific process and identify a new therapeutic candidate for dry age-related macular degeneration (dAMD), a leading cause of irreversible blindness worldwide. In June, FutureHouse released ether0, a 24B open-weights reasoning model for chemistry.

“You really have to think of these agents as part of a larger system,” Rodriques says. “Soon, the literature search agents will be integrated with the data analysis agent, the hypothesis generation agent, an experiment planning agent, and they will all be engineered to work together seamlessly.”

Agents for everyone

Today anyone can access FutureHouse’s agents at platform.futurehouse.org. The company’s platform launch generated excitement in the industry, and stories have started to come in about scientists using the agents to accelerate research.

One of FutureHouse’s scientists used the agents to identify a gene that could be associated with polycystic ovary syndrome and come up with a new treatment hypothesis for the disease. Another researcher at the Lawrence Berkeley National Laboratory used Crow to create an AI assistant capable of searching the PubMed research database for information related to Alzheimer’s disease.

Scientists at another research institution have used the agents to conduct systematic reviews of genes relevant to Parkinson’s disease, finding FutureHouse’s agents performed better than general agents.

Rodriques says scientists who think of the agents less like Google Scholar and more like a smart assistant scientist get the most out of the platform.

“People who are looking for speculation tend to get more mileage out of Chat-GPT o3 deep research, while people who are looking for really faithful literature reviews tend to get more out of our agents,” Rodriques explains.

Rodriques also thinks FutureHouse will soon get to a point where its agents can use the raw data from research papers to test the reproducibility of its results and verify conclusions.

In the longer run, to keep scientific progress marching forward, Rodriques says FutureHouse is working on embedding its agents with tacit knowledge to be able to perform more sophisticated analyses while also giving the agents the ability to use computational tools to explore hypotheses.

“There have been so many advances around foundation models for science and around language models for proteins and DNA, that we now need to give our agents access to those models and all of the other tools people commonly use to do science,” Rodriques says. “Building the infrastructure to allow agents to use more specialized tools for science is going to be critical.”

Leveraging the strengths of two world-class research institutions, MIT and Mass General Brigham (MGB) recently celebrated the launch of the MIT-MGB Seed Program. The new initiative, which is supported by Analog Devices Inc. (ADI), will fund joint research projects led by researchers at MIT and Mass General Brigham. These collaborative projects will advance research in human health, with the goal of developing next-generation therapies, diagnostics, and digital tools that can improve lives at scale. 

The program represents a unique opportunity to dramatically accelerate innovations that address some of the most urgent challenges in human health. By supporting interdisciplinary teams from MIT and Mass General Brigham, including both researchers and clinicians, the seed program will foster groundbreaking work that brings together expertise in artificial intelligence, machine learning, and measurement and sensing technologies with pioneering clinical research and patient care.

“The power of this program is that it combines MIT’s strength in science, engineering, and innovation with Mass General Brigham’s world-class scientific and clinical research. With the support and incentive to work together, researchers and clinicians will have the freedom to tackle compelling problems and find novel ways to overcome them to achieve transformative changes in patient care,” says Sally Kornbluth, president of MIT.

“The MIT-MGB Seed Program will enable cross-disciplinary collaboration to advance transformative research and breakthrough science. By combining the collective strengths and expertise of our great institutions, we can transform medical care and drive innovation and discovery with speed,” says Anne Klibanski, president and CEO of Mass General Brigham.

The initiative is funded by a gift from ADI. Over the next three years, the ADI Fund for Health and Life Sciences will support approximately six joint projects annually, with funding split between the two institutions. 

“The converging domains of biology, medicine, and computing promise a new era of health-care efficacy, efficiency, and access. ADI has enjoyed a long and fruitful history of collaboration with MIT and Mass General Brigham, and we are excited by this new initiative’s potential to transform the future of patient care,” adds Vincent Roche, CEO and chair of the board of directors at ADI.

In addition to funding, teams selected for the program will have access to entrepreneurial workshops, including some hosted by The Engine — an MIT-built venture firm focused on tough tech. These sessions will connect researchers with company founders, investors, and industry leaders, helping them chart a path from breakthrough discoveries in the lab to real-world impact.

The program will launch an open call for proposals to researchers at MIT and Mass General Brigham. The first cohort of funded projects is expected to launch in fall 2025. Awardees will be selected by a joint review committee composed of MIT and Mass General Brigham experts.

According to MIT’s faculty lead for the MIT-MGB Seed Program, Alex K. Shalek, building collaborative research teams with leaders from both institutions could help fill critical gaps that often impede innovation in health and life sciences. Shalek also serves as director of the Institute for Medical Engineering & Science (IMES), the J. W. Kieckhefer Professor in IMES and Chemistry, and an extramural member of the Koch Institute for Integrative Cancer Research.

 “Clinicians often see where current interventions fall short, but may lack the scientific tools or engineering expertise needed to develop new ones. Conversely, MIT researchers may not fully grasp these clinical challenges or have access to the right patient data and samples,” explains Shalek, who is also a member of the Ragon Institute of Mass General Brigham, MIT, and Harvard. “By supporting bilateral collaborations and building a community across disciplines, this program is poised to drive critical advances in diagnostics, therapeutics, and AI-driven health applications.”

Emery Brown, a practicing anesthesiologist at Massachusetts General Hospital, will serve alongside Shalek as Mass General Brigham’s faculty lead for the program.

“The MIT-MGB Seed Program creates a perfect storm. The program will provide an opportunity for MIT faculty to bring novel science and engineering to attack and solve important clinical problems,” adds Brown, who is also the Edward Hood Taplin Professor of Medical Engineering and Computational Neuroscience at MIT. “The pursuit of solutions to important and challenging clinical problems by Mass General Brigham physicians and scientists will no doubt spur MIT scientists and engineers to develop new technologies, or find novel applications of existing technologies.”

The MIT-MGB Seed Program is a flagship initiative in the MIT Health and Life Sciences Collaborative (MIT HEALS). It reflects MIT HEALS’ core mission to establish MIT as a central hub for health and life sciences innovation and translation, and to leverage connections with other world-class research institutions in the Boston area.

“This program exemplifies the power of interdisciplinary research,” says Anantha Chandrakasan, MIT’s chief innovation and strategy officer, dean of engineering, and head of MIT HEALS. “It creates a critical bridge between clinical practice and technological innovation — two areas that must be deeply connected to advance real-world solutions.”

The program’s launch was celebrated at a special event at MIT’s Samberg Conference Center on March 31.

Diffusion models like OpenAI’s DALL-E are becoming increasingly useful in helping brainstorm new designs. Humans can prompt these systems to generate an image, create a video, or refine a blueprint, and come back with ideas they hadn’t considered before.

But did you know that generative artificial intelligence (GenAI) models are also making headway in creating working robots? Recent diffusion-based approaches have generated structures and the systems that control them from scratch. With or without a user’s input, these models can make new designs and then evaluate them in simulation before they’re fabricated.

A new approach from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) applies this generative know-how toward improving humans’ robotic designs. Users can draft a 3D model of a robot and specify which parts they’d like to see a diffusion model modify, providing its dimensions beforehand. GenAI then brainstorms the optimal shape for these areas and tests its ideas in simulation. When the system finds the right design, you can save and then fabricate a working, real-world robot with a 3D printer, without requiring additional tweaks.

The researchers used this approach to create a robot that leaps up an average of roughly 2 feet, or 41 percent higher than a similar machine they created on their own. The machines are nearly identical in appearance: They’re both made of a type of plastic called polylactic acid, and while they initially appear flat, they spring up into a diamond shape when a motor pulls on the cord attached to them. So what exactly did AI do differently?

A closer look reveals that the AI-generated linkages are curved, and resemble thick drumsticks (the musical instrument drummers use), whereas the standard robot’s connecting parts are straight and rectangular.

Better and better blobs

The researchers began to refine their jumping robot by sampling 500 potential designs using an initial embedding vector — a numerical representation that captures high-level features to guide the designs generated by the AI model. From these, they selected the top 12 options based on performance in simulation and used them to optimize the embedding vector.

This process was repeated five times, progressively guiding the AI model to generate better designs. The resulting design resembled a blob, so the researchers prompted their system to scale the draft to fit their 3D model. They then fabricated the shape, finding that it indeed improved the robot’s jumping abilities.

The advantage of using diffusion models for this task, according to co-lead author and CSAIL postdoc Byungchul Kim, is that they can find unconventional solutions to refine robots.

“We wanted to make our machine jump higher, so we figured we could just make the links connecting its parts as thin as possible to make them light,” says Kim. “However, such a thin structure can easily break if we just use 3D printed material. Our diffusion model came up with a better idea by suggesting a unique shape that allowed the robot to store more energy before it jumped, without making the links too thin. This creativity helped us learn about the machine’s underlying physics.”

The team then tasked their system with drafting an optimized foot to ensure it landed safely. They repeated the optimization process, eventually choosing the best-performing design to attach to the bottom of their machine. Kim and his colleagues found that their AI-designed machine fell far less often than its baseline, to the tune of an 84 percent improvement.

The diffusion model’s ability to upgrade a robot’s jumping and landing skills suggests it could be useful in enhancing how other machines are designed. For example, a company working on manufacturing or household robots could use a similar approach to improve their prototypes, saving engineers time normally reserved for iterating on those changes.

The balance behind the bounce

To create a robot that could jump high and land stably, the researchers recognized that they needed to strike a balance between both goals. They represented both jumping height and landing success rate as numerical data, and then trained their system to find a sweet spot between both embedding vectors that could help build an optimal 3D structure.

The researchers note that while this AI-assisted robot outperformed its human-designed counterpart, it could soon reach even greater new heights. This iteration involved using materials that were compatible with a 3D printer, but future versions would jump even higher with lighter materials.

Co-lead author and MIT PhD student and CSAIL affiliate Tsun-Hsuan “Johnson” Wang says the project is a jumping-off point for new robotics designs that generative AI could help with.

“We want to branch out to more flexible goals,” says Wang. “Imagine using natural language to guide a diffusion model to draft a robot that can pick up a mug, or operate an electric drill.”

Kim says that a diffusion model could also help to generate articulation and ideate on how parts connect, potentially improving how high the robot would jump. The team is also exploring the possibility of adding more motors to control which direction the machine jumps and perhaps improve its landing stability.

The researchers’ work was supported, in part, by the National Science Foundation’s Emerging Frontiers in Research and Innovation program, the Singapore-MIT Alliance for Research and Technology’s Mens, Manus and Machina program, and the Gwangju Institute of Science and Technology (GIST)-CSAIL Collaboration. They presented their work at the 2025 International Conference on Robotics and Automation.

Four MIT rising seniors have been selected to receive a 2025 Barry Goldwater Scholarship, including Avani Ahuja and Jacqueline Prawira in the School of Engineering and Julianna Lian and Alex Tang from the School of Science. An estimated 5,000 college sophomores and juniors from across the United States were nominated for the scholarships, of whom only 441 were selected.

The Goldwater Scholarships have been conferred since 1989 by the Barry Goldwater Scholarship and Excellence in Education Foundation. These scholarships have supported undergraduates who go on to become leading scientists, engineers, and mathematicians in their respective fields.

Avani Ahuja, a mechanical engineering and electrical engineering major, conducts research in the Conformable Decoders group, where she is focused on developing a “wearable conformable breast ultrasound patch” that makes ultrasounds for breast cancer more accessible.

“Doing research in the Media Lab has had a huge impact on me, especially in the ways that we think about inclusivity in research,” Ahuja says.

In her research group, Ahuja works under Canan Dagdeviren, the LG Career Development Professor of Media Arts and Sciences. Ahuja plans to pursue a PhD in electrical engineering. She aspires to conduct research in electromechanical systems for women’s health applications and teach at the university level.

“I want to thank Professor Dagdeviren for all her support. It’s an honor to receive this scholarship, and it’s amazing to see that women’s health research is getting recognized in this way,” Ahuja says.

Julianna Lian studies mechanochemistry, organic, and polymer chemistry in the lab of Professor Jeremiah Johnson, the A. Thomas Guertin Professor of Chemistry. In addition to her studies, she serves the MIT community as an emergency medical technician (EMT) with MIT Emergency Medical Services, is a member of MIT THINK, and a ClubChem mentorship chair.

“Receiving this award has been a tremendous opportunity to not only reflect on how much I have learned, but also on the many, many people I have had the chance to learn from,” says Lian. “I am deeply grateful for the guidance, support, and encouragement of these teachers, mentors, and friends. And I am excited to carry forward the lasting curiosity and excitement for chemistry that they have helped inspire in me.”

Lian’s career goals post-graduation include pursuing a PhD in organic chemistry, to conduct research at the interface of synthetic chemistry and materials science, aided by computation, and to teach at the university level.

Jacqueline Prawira, a materials science and engineering major, joined the Center of Decarbonization and Electrification of Industry as a first-year Undergraduate Research Opportunities Program student and became a co-inventor on a patent and a research technician at spinout company Rock Zero. She has also worked in collaboration with Indigenous farmers and Diné College students on the Navajo Nation.

“I’ve become significantly more cognizant of how I listen to people and stories, the tangled messiness of real-world challenges, and the critical skills needed to tackle complex sustainability issues,” Prawira says.

Prawira is mentored by Yet-Ming Chiang, professor of materials science and engineering. Her career goals are to pursue a PhD in materials science and engineering and to research sustainable materials and processes to solve environmental challenges and build a sustainable society.

“Receiving the prestigious title of 2025 Goldwater Scholar validates my current trajectory in innovating sustainable materials and demonstrates my growth as a researcher,” Prawira says. “This award signifies my future impact in building a society where sustainability is the norm, instead of just another option.”

Alex Tang studies the effects of immunotherapy and targeted molecular therapy on the tumor microenvironment in metastatic colorectal cancer patients. He is supervised by professors Jonathan Chen at Northwestern University and Nir Hacohen at the Broad Institute of MIT and Harvard.

“My mentors and collaborators have been instrumental to my growth since I joined the lab as a freshman. I am incredibly grateful for the generous mentorship and support of Professor Hacohen and Professor Chen, who have taught me how to approach scientific investigation with curiosity and rigor,” says Tang. “I’d also like to thank my advisor Professor Adam Martin and first-year advisor Professor Angela Belcher for their guidance throughout my undergraduate career thus far. I am excited to carry forward this work as I progress in my career.” Tang intends to pursue physician-scientist training following graduation.

The Scholarship Program honoring Senator Barry Goldwater was designed to identify, encourage, and financially support outstanding undergraduates interested in pursuing research careers in the sciences, engineering, and mathematics. The Goldwater Scholarship is the preeminent undergraduate award of its type in these fields.

A large language model (LLM) deployed to make treatment recommendations can be tripped up by nonclinical information in patient messages, like typos, extra white space, missing gender markers, or the use of uncertain, dramatic, and informal language, according to a study by MIT researchers.

They found that making stylistic or grammatical changes to messages increases the likelihood an LLM will recommend that a patient self-manage their reported health condition rather than come in for an appointment, even when that patient should seek medical care.

Their analysis also revealed that these nonclinical variations in text, which mimic how people really communicate, are more likely to change a model’s treatment recommendations for female patients, resulting in a higher percentage of women who were erroneously advised not to seek medical care, according to human doctors.

This work “is strong evidence that models must be audited before use in health care — which is a setting where they are already in use,” says Marzyeh Ghassemi, an associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS), a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems, and senior author of the study.

These findings indicate that LLMs take nonclinical information into account for clinical decision-making in previously unknown ways. It brings to light the need for more rigorous studies of LLMs before they are deployed for high-stakes applications like making treatment recommendations, the researchers say.

“These models are often trained and tested on medical exam questions but then used in tasks that are pretty far from that, like evaluating the severity of a clinical case. There is still so much about LLMs that we don’t know,” adds Abinitha Gourabathina, an EECS graduate student and lead author of the study.

They are joined on the paper, which will be presented at the ACM Conference on Fairness, Accountability, and Transparency, by graduate student Eileen Pan and postdoc Walter Gerych.

Mixed messages

Launched in February of this year, the MIT Generative AI Impact Consortium (MGAIC), a presidential initiative led by MIT’s Office of Innovation and Strategy and administered by the MIT Stephen A. Schwarzman College of Computing, issued a call for proposals, inviting researchers from across MIT to submit ideas for innovative projects studying high-impact uses of generative AI models.

The call received 180 submissions from nearly 250 faculty members, spanning all of MIT’s five schools and the college. The overwhelming response across the Institute exemplifies the growing interest in AI and follows in the wake of MIT’s Generative AI Week and call for impact papers. Fifty-five proposals were selected for MGAIC’s inaugural seed grants, with several more selected to be funded by the consortium’s founding company members.

Over 30 funding recipients presented their proposals to the greater MIT community at a kickoff event on May 13. Anantha P. Chandrakasan, chief innovation and strategy officer and dean of the School of Engineering who is head of the consortium, welcomed the attendees and thanked the consortium’s founding industry members.

“The amazing response to our call for proposals is an incredible testament to the energy and creativity that MGAIC has sparked at MIT. We are especially grateful to our founding members, whose support and vision helped bring this endeavor to life,” adds Chandrakasan. “One of the things that has been most remarkable about MGAIC is that this is a truly cross-Institute initiative. Deans from all five schools and the college collaborated in shaping and implementing it.”

Vivek F. Farias, the Patrick J. McGovern (1959) Professor at the MIT Sloan School of Management and co-faculty director of the consortium with Tim Kraska, associate professor of electrical engineering and computer science in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), emceed the afternoon of five-minute lightning presentations.

Presentation highlights include:

“AI-Driven Tutors and Open Datasets for Early Literacy Education,” presented by Ola Ozernov-Palchik, a research scientist at the McGovern Institute for Brain Research, proposed a refinement for AI-tutors for pK-7 students to potentially decrease literacy disparities.

“Developing jam_bots: Real-Time Collaborative Agents for Live Human-AI Musical Improvisation,” presented by Anna Huang, assistant professor of music and assistant professor of electrical engineering and computer science, and Joe Paradiso, the Alexander W. Dreyfoos (1954) Professor in Media Arts and Sciences at the MIT Media Lab, aims to enhance human-AI musical collaboration in real-time for live concert improvisation.

“GENIUS: GENerative Intelligence for Urban Sustainability,” presented by Norhan Bayomi, a postdoc at the MIT Environmental Solutions Initiative and a research assistant in the Urban Metabolism Group, which aims to address the critical gap of a standardized approach in evaluating and benchmarking cities’ climate policies.

Georgia Perakis, the John C Head III Dean (Interim) of the MIT Sloan School of Management and professor of operations management, operations research, and statistics, who serves as co-chair of the GenAI Dean’s oversight group with Dan Huttenlocher, dean of the MIT Schwarzman College of Computing, ended the event with closing remarks that emphasized “the readiness and eagerness of our community to lead in this space.”

“This is only the beginning,” she continued. “We are at the front edge of a historic moment — one where MIT has the opportunity, and the responsibility, to shape the future of generative AI with purpose, with excellence, and with care.”

Volcanic islands, such as the islands of Hawaii and the Caribbean, are surrounded by coral reefs that encircle an island in a labyrinthine, living ring. A coral reef is punctured at points by reef passes — wide channels that cut through the coral and serve as conduits for ocean water and nutrients to filter in and out. These watery passageways provide circulation throughout a reef, helping to maintain the health of corals by flushing out freshwater and transporting key nutrients.

Now, MIT scientists have found that reef passes are shaped by island rivers. In a study appearing today in the journal Geophysical Research Letters, the team shows that the locations of reef passes along coral reefs line up with where rivers funnel out from an island’s coast.

Their findings provide the first quantitative evidence of rivers forming reef passes.  Scientists and explorers had speculated that this may be the case: Where a river on a volcanic island meets the coast, the freshwater and sediment it carries flows toward the reef, where a strong enough flow can tunnel into the surrounding coral. This idea has been proposed from time to time but never quantitatively tested, until now.

“The results of this study help us to understand how the health of coral reefs depends on the islands they surround,” says study author Taylor Perron, the Cecil and Ida Green Professor of Earth, Atmospheric and Planetary Sciences at MIT.

“A lot of discussion around rivers and their impact on reefs today has been negative because of human impact and the effects of agricultural practices,” adds lead author Megan Gillen, a graduate student in the MIT-WHOI Joint Program in Oceanography. “This study shows the potential long-term benefits rivers can have on reefs, which I hope reshapes the paradigm and highlights the natural state of rivers interacting with reefs.”

The study’s other co-author is Andrew Ashton of the Woods Hole Oceanographic Institution.

Drawing the lines

The new study is based on the team’s analysis of the Society Islands, a chain of islands in the South Pacific Ocean that includes Tahiti and Bora Bora. Gillen, who joined the MIT-WHOI program in 2020, was interested in exploring connections between coral reefs and the islands they surround. With limited options for on-site work during the Covid-19 pandemic, she and Perron looked to see what they could learn through satellite images and maps of island topography. They did a quick search using Google Earth and zeroed in on the Society Islands for their uniquely visible reef and island features.

“The islands in this chain have these iconic, beautiful reefs, and we kept noticing these reef passes that seemed to align with deeply embayed portions of the coastline,” Gillen says. “We started asking ourselves, is there a correlation here?”

Viewed from above, the coral reefs that circle some islands bear what look to be notches, like cracks that run straight through a ring. These breaks in the coral are reef passes — large channels that run tens of meters deep and can be wide enough for some boats to pass through. On first look, Gillen noticed that the most obvious reef passes seemed to line up with flooded river valleys — depressions in the coastline that have been eroded over time by island rivers that flow toward the ocean. She wondered whether and to what extent island rivers might shape reef passes.

“People have examined the flow through reef passes to understand how ocean waves and seawater circulate in and out of lagoons, but there have been no claims of how these passes are formed,” Gillen says. “Reef pass formation has been mentioned infrequently in the literature, and people haven’t explored it in depth.”

Reefs unraveled

To get a detailed view of the topography in and around the Society Islands, the team used data from the NASA Shuttle Radar Topography Mission — two radar antennae that flew aboard the space shuttle in 1999 and measured the topography across 80 percent of the Earth’s surface.

The researchers used the mission’s topographic data in the Society Islands to create a map of every drainage basin along the coast of each island, to get an idea of where major rivers flow or once flowed. They also marked the locations of every reef pass in the surrounding coral reefs. They then essentially “unraveled” each island’s coastline and reef into a straight line, and compared the locations of basins versus reef passes.

“Looking at the unwrapped shorelines, we find a significant correlation in the spatial relationship between these big river basins and where the passes line up,” Gillen says. “So we can say that statistically, the alignment of reef passes and large rivers does not seem random. The big rivers have a role in forming passes.”

As for how rivers shape the coral conduits, the team has two ideas, which they call, respectively, reef incision and reef encroachment. In reef incision, they propose that reef passes can form in times when the sea level is relatively low, such that the reef is exposed above the sea surface and a river can flow directly over the reef. The water and sediment carried by the river can then erode the coral, progressively carving a path through the reef.

When sea level is relatively higher, the team suspects a reef pass can still form, through reef encroachment. Coral reefs naturally live close to the water surface, where there is light and opportunity for photosynthesis. When sea levels rise, corals naturally grow upward and inward toward an island, to try to “catch up” to the water line.

“Reefs migrate toward the islands as sea levels rise, trying to keep pace with changing average sea level,” Gillen says.

However, part of the encroaching reef can end up in old river channels that were previously carved out by large rivers and that are lower than the rest of the island coastline. The corals in these river beds end up deeper than light can extend into the water column, and inevitably drown, leaving a gap in the form of a reef pass.

“We don’t think it’s an either/or situation,” Gillen says. “Reef incision occurs when sea levels fall, and reef encroachment happens when sea levels rise. Both mechanisms, occurring over dozens of cycles of sea-level rise and island evolution, are likely responsible for the formation and maintenance of reef passes over time.”

The team also looked to see whether there were differences in reef passes in older versus younger islands. They observed that younger islands were surrounded by more reef passes that were spaced closer together, versus older islands that had fewer reef passes that were farther apart.

As islands age, they subside, or sink, into the ocean, which reduces the amount of land that funnels rainwater into rivers. Eventually, rivers are too weak to keep the reef passes open, at which point, the ocean likely takes over, and incoming waves could act to close up some passes.

Gillen is exploring ideas for how rivers, or river-like flow, can be engineered to create paths through coral reefs in ways that would promote circulation and benefit reef health.

“Part of me wonders: If you had a more persistent flow, in places where you don’t naturally have rivers interacting with the reef, could that potentially be a way to increase health, by incorporating that river component back into the reef system?” Gillen says. “That’s something we’re thinking about.”

This research was supported, in part, by the WHOI Watson and Von Damm fellowships.

Water makes up around 60 percent of the human body. More than half of this water sloshes around inside the cells that make up organs and tissues. Much of the remaining water flows in the nooks and crannies between cells, much like seawater between grains of sand.

Now, MIT engineers have found that this “intercellular” fluid plays a major role in how tissues respond when squeezed, pressed, or physically deformed. Their findings could help scientists understand how cells, tissues, and organs physically adapt to conditions such as aging, cancer, diabetes, and certain neuromuscular diseases.

In a paper appearing today in Nature Physics, the researchers show that when a tissue is pressed or squeezed, it is more compliant and relaxes more quickly when the fluid between its cells flows easily. When the cells are packed together and there is less room for intercellular flow, the tissue as a whole is stiffer and resists being pressed or squeezed.

The findings challenge conventional wisdom, which has assumed that a tissue’s compliance depends mainly on what’s inside, rather than around, a cell. Now that the researchers have shown that intercellular flow determines how tissues will adapt to physical forces, the results can be applied to understand a wide range of physiological conditions, including how muscles withstand exercise and recover from injury, and how a tissue’s physical adaptability may affect the progression of aging, cancer, and other medical conditions.

The team envisions the results could also inform the design of artificial tissues and organs. For instance, in engineering artificial tissue, scientists might optimize intercellular flow within the tissue to improve its function or resilience. The researchers suspect that intercellular flow could also be a route for delivering nutrients or therapies, either to heal a tissue or eradicate a tumor.

“People know there is a lot of fluid between cells in tissues, but how important that is, in particular in tissue deformation, is completely ignored,” says Ming Guo, associate professor of mechanical engineering at MIT. “Now we really show we can observe this flow. And as the tissue deforms, flow between cells dominates the behavior. So, let’s pay attention to this when we study diseases and engineer tissues.”

Guo is a co-author of the new study, which includes lead author and MIT postdoc Fan Liu PhD ’24, along with Bo Gao and Hui Li of Beijing Normal University, and Liran Lei and Shuainan Liu of Peking Union Medical College.

Pressed and squeezed

The tissues and organs in our body are constantly undergoing physical deformations, from the large stretch and strain of muscles during motion to the small and steady contractions of the heart. In some cases, how easily tissues adapt to deformation can relate to how quickly a person can recover from, for instance, an allergic reaction, a sports injury, or a brain stroke. However, exactly what sets a tissue’s response to deformation is largely unknown.

Guo and his group at MIT looked into the mechanics of tissue deformation, and the role of intercellular flow in particular, following a study they published in 2020. In that study, they focused on tumors and observed the way in which fluid can flow from the center of a tumor out to its edges, through the cracks and crevices between individual tumor cells. They found that when a tumor was squeezed or pressed, the intercellular flow increased, acting as a conveyor belt to transport fluid from the center to the edges. Intercellular flow, they found, could fuel tumor invasion into surrounding regions.

In their new study, the team looked to see what role this intercellular flow might play in other, noncancerous tissues.

“Whether you allow the fluid to flow between cells or not seems to have a major impact,” Guo says. “So we decided to look beyond tumors to see how this flow influences how other tissues respond to deformation.”

A fluid pancake

Guo, Liu, and their colleagues studied the intercellular flow in a variety of biological tissues, including cells derived from pancreatic tissue. They carried out experiments in which they first cultured small clusters of tissue, each measuring less than a quarter of a millimeter wide and numbering tens of thousands of individual cells. They placed each tissue cluster in a custom-designed testing platform that the team built specifically for the study.

“These microtissue samples are in this sweet zone where they are too large to see with atomic force microscopy techniques and too small for bulkier devices,” Guo says. “So, we decided to build a device.”

The researchers adapted a high-precision microbalance that measures minute changes in weight. They combined this with a step motor that is designed to press down on a sample with nanometer precision. The team placed tissue clusters one at a time on the balance and recorded each cluster’s changing weight as it relaxed from a sphere into the shape of a pancake in response to the compression. The team also took videos of the clusters as they were squeezed.

For each type of tissue, the team made clusters of varying sizes. They reasoned that if the tissue’s response is ruled by the flow between cells, then the bigger a tissue, the longer it should take for water to seep through, and therefore, the longer it should take the tissue to relax. It should take the same amount of time, regardless of size, if a tissue’s response is determined by the structure of the tissue rather than fluid.

Over multiple experiments with a variety of tissue types and sizes, the team observed a similar trend: The bigger the cluster, the longer it took to relax, indicating that intercellular flow dominates a tissue’s response to deformation.

“We show that this intercellular flow is a crucial component to be considered in the fundamental understanding of tissue mechanics and also applications in engineering living systems,” Liu says.

Going forward, the team plans to look into how intercellular flow influences brain function, particularly in disorders such as Alzheimer’s disease.

“Intercellular or interstitial flow can help you remove waste and deliver nutrients to the brain,” Liu adds. “Enhancing this flow in some cases might be a good thing.”

“As this work shows, as we apply pressure to a tissue, fluid will flow,” Guo says. “In the future, we can think of designing ways to massage a tissue to allow fluid to transport nutrients between cells.”

More than half of the nation’s 623,218 bridges are experiencing significant deterioration. Through an in-field case study conducted in western Massachusetts, a team led by the University of Massachusetts at Amherst in collaboration with researchers from the MIT Department of Mechanical Engineering (MechE) has just successfully demonstrated that 3D printing may provide a cost-effective, minimally disruptive solution.

“Anytime you drive, you go under or over a corroded bridge,” says Simos Gerasimidis, associate professor of civil and environmental engineering at UMass Amherst and former visiting professor in the Department of Civil and Environmental Engineering at MIT, in a press release. “They are everywhere. It’s impossible to avoid, and their condition often shows significant deterioration. We know the numbers.”

The numbers, according to the American Society of Civil Engineers’ 2025 Report Card for America’s Infrastructure, are staggering: Across the United States, 49.1 percent of the nation’s 623,218 bridges are in “fair” condition and 6.8 percent are in “poor” condition. The projected cost to restore all of these failing bridges exceeds $191 billion.

A proof-of-concept repair took place last month on a small, corroded section of a bridge in Great Barrington, Massachusetts. The technique, called cold spray, can extend the life of beams, reinforcing them with newly deposited steel. The process accelerates particles of powdered steel in heated, compressed gas, and then a technician uses an applicator to spray the steel onto the beam. Repeated sprays create multiple layers, restoring thickness and other structural properties.

This method has proven to be an effective solution for other large structures like submarines, airplanes, and ships, but bridges present a problem on a greater scale. Unlike movable vessels, stationary bridges cannot be brought to the 3D printer — the printer must be brought on-site — and, to lessen systemic impacts, repairs must also be made with minimal disruptions to traffic, which the new approach allows.

“Now that we’ve completed this proof-of-concept repair, we see a clear path to a solution that is much faster, less costly, easier, and less invasive,” says Gerasimidis. “To our knowledge, this is a first. Of course, there is some R&D that needs to be developed, but this is a huge milestone to that.”

“This is a tremendous collaboration where cutting-edge technology is brought to address a critical need for infrastructure in the commonwealth and across the United States,” says John Hart, Class of 1922 Professor and head of the Department of MechE at MIT. Hart and Haden Quinlan, senior program manager in the Center for Advanced Production Technologies at MIT, are leading MIT’s efforts in in the project. Hart is also faculty co-lead of the recently announced MIT Initiative for New Manufacturing.

“Integrating digital systems with advanced physical processing is the future of infrastructure,” says Quinlan. “We’re excited to have moved this technology beyond the lab and into the field, and grateful to our collaborators in making this work possible.”

UMass says the Massachusetts Department of Transportation (MassDOT) has been a valued research partner, helping to identify the problem and providing essential support for the development and demonstration of the technology. Technical guidance and funding support were provided by the MassDOT Highway Division and the Research and Technology Transfer Program.

Equipment for this project was supported through the Massachusetts Manufacturing Innovation Initiative, a statewide program led by the Massachusetts Technology Collaborative (MassTech)’s Center for Advanced Manufacturing that helps bridge the gap between innovation and commercialization in hard tech manufacturing.

“It’s a very Massachusetts success story,” Gerasimidis says. “It involves MassDOT being open-minded to new ideas. It involves UMass and MIT putting [together] the brains to do it. It involves MassTech to bring manufacturing back to Massachusetts. So, I think it’s a win-win for everyone involved here.”

The bridge in Great Barrington is scheduled for demolition in a few years. After demolition occurs, the recently-sprayed beams will be taken back to UMass for testing and measurement to study how well the deposited steel powder adhered to the structure in the field compared to in a controlled lab setting, if it corroded further after it was sprayed, and determine its mechanical properties.

This demonstration builds on several years of research by the UMass and MIT teams, including development of a “digital thread” approach to scan corroded beam surfaces and determine material deposition profiles, alongside laboratory studies of cold spray and other additive manufacturing approaches that are suited to field deployment.

Altogether, this work is a collaborative effort among UMass Amherst, MIT MechE, MassDOT, the Massachusetts Technology Collaborative (MassTech), the U.S. Department of Transportation, and the Federal Highway Administration. Research reports are available on the MassDOT website.

When the Earth froze over, where did life shelter? MIT scientists say one refuge may have been pools of melted ice that dotted the planet’s icy surface.

In a study appearing today in Nature Communications, the researchers report that 635 million to 720 million years ago, during periods known as “Snowball Earth,” when much of the planet was covered in ice, some of our ancient cellular ancestors could have waited things out in meltwater ponds.

The scientists found that eukaryotes — complex cellular lifeforms that eventually evolved into the diverse multicellular life we see today — could have survived the global freeze by living in shallow pools of water. These small, watery oases may have persisted atop relatively shallow ice sheets present in equatorial regions. There, the ice surface could accumulate dark-colored dust and debris from below, which enhanced its ability to melt into pools. At temperatures hovering around 0 degrees Celsius, the resulting meltwater ponds could have served as habitable environments for certain forms of early complex life.

The team drew its conclusions based on an analysis of modern-day meltwater ponds. Today in Antarctica, small pools of melted ice can be found along the margins of ice sheets. The conditions along these polar ice sheets are similar to what likely existed along ice sheets near the equator during Snowball Earth.

The researchers analyzed samples from a variety of meltwater ponds located on the McMurdo Ice Shelf in an area that was first described by members of Robert Falcon Scott's 1903 expedition as “dirty ice.” The MIT researchers discovered clear signatures of eukaryotic life in every pond. The communities of eukaryotes varied from pond to pond, revealing a surprising diversity of life across the setting. The team also found that salinity plays a key role in the kind of life a pond can host: Ponds that were more brackish or salty had more similar eukaryotic communities, which differed from those in ponds with fresher waters.

“We’ve shown that meltwater ponds are valid candidates for where early eukaryotes could have sheltered during these planet-wide glaciation events,” says lead author Fatima Husain, a graduate student in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “This shows us that diversity is present and possible in these sorts of settings. It’s really a story of life’s resilience.”

The study’s MIT co-authors include Schlumberger Professor of Geobiology Roger Summons and former postdoc Thomas Evans, along with Jasmin Millar of Cardiff University, Anne Jungblut at the Natural History Museum in London, and Ian Hawes of the University of Waikato in New Zealand.

Polar plunge

“Snowball Earth” is the colloquial term for periods of time in Earth history during which the planet iced over. It is often used as a reference to the two consecutive, multi-million-year glaciation events which took place during the Cryogenian Period, which geologists refer to as the time between 635 and 720 million years ago. Whether the Earth was more of a hardened snowball or a softer “slushball” is still up for debate. But scientists are certain of one thing: Most of the planet was plunged into a deep freeze, with average global temperatures of minus 50 degrees Celsius. The question has been: How and where did life survive?

“We’re interested in understanding the foundations of complex life on Earth. We see evidence for eukaryotes before and after the Cryogenian in the fossil record, but we largely lack direct evidence of where they may have lived during,” Husain says. “The great part of this mystery is, we know life survived. We’re just trying to understand how and where.”

There are a number of ideas for where organisms could have sheltered during Snowball Earth, including in certain patches of the open ocean (if such environments existed), in and around deep-sea hydrothermal vents, and under ice sheets. In considering meltwater ponds, Husain and her colleagues pursued the hypothesis that surface ice meltwaters may also have been capable of supporting early eukaryotic life at the time.

“There are many hypotheses for where life could have survived and sheltered during the Cryogenian, but we don’t have excellent analogs for all of them,” Husain notes. “Above-ice meltwater ponds occur on Earth today and are accessible, giving us the opportunity to really focus in on the eukaryotes which live in these environments.”

Small pond, big life

For their new study, the researchers analyzed samples taken from meltwater ponds in Antarctica. In 2018, Summons and colleagues from New Zealand traveled to a region of the McMurdo Ice Shelf in East Antarctica, known to host small ponds of melted ice, each just a few feet deep and a few meters wide. There, water freezes all the way to the seafloor, in the process trapping dark-colored sediments and marine organisms. Wind-driven loss of ice from the surface creates a sort of conveyer belt that brings this trapped debris to the surface over time, where it absorbs the sun’s warmth, causing ice to melt, while surrounding debris-free ice reflects incoming sunlight, resulting in the formation of shallow meltwater ponds.

The bottom of each pond is lined with mats of microbes that have built up over years to form layers of sticky cellular communities.

“These mats can be a few centimeters thick, colorful, and they can be very clearly layered,” Husain says.

These microbial mats are made up of cyanobacteria, prokaryotic, single-celled photosynthetic organisms that lack a cell nucleus or other organelles. While these ancient microbes are known to survive within some of the the harshest environments on Earth including meltwater ponds, the researchers wanted to know whether eukaryotes — complex organisms that evolved a cell nucleus and other membrane bound organelles — could also weather similarly challenging circumstances. Answering this question would take more than a microscope, as the defining characteristics of the microscopic eukaryotes present among the microbial mats are too subtle to distinguish by eye.

To characterize the eukaryotes, the team analyzed the mats for specific lipids they make called sterols, as well as genetic components called ribosomal ribonucleic acid (rRNA), both of which can be used to identify organisms with varying degrees of specificity. These two independent sets of analyses provided complementary fingerprints for certain eukaryotic groups. As part of the team’s lipid research, they found many sterols and rRNA genes closely associated with specific types of algae, protists, and microscopic animals among the microbial mats. The researchers were able to assess the types and relative abundance of lipids and rRNA genes from pond to pond, and found the ponds hosted a surprising diversity of eukaryotic life.

“No two ponds were alike,” Husain says. “There are repeating casts of characters, but they’re present in different abundances. And we found diverse assemblages of eukaryotes from all the major groups in all the ponds studied. These eukaryotes are the descendants of the eukaryotes that survived the Snowball Earth. This really highlights that meltwater ponds during Snowball Earth could have served as above-ice oases that nurtured the eukaryotic life that enabled the diversification and proliferation of complex life — including us — later on.”

This research was supported, in part, by the NASA Exobiology Program, the Simons Collaboration on the Origins of Life, and a MISTI grant from MIT-New Zealand.

Researchers at MIT and the Scripps Research Institute have shown that they can generate a strong immune response to HIV with just one vaccine dose, by adding two powerful adjuvants — materials that help stimulate the immune system.

In a study of mice, the researchers showed that this approach produced a much wider diversity of antibodies against an HIV antigen, compared to the vaccine given on its own or with just one of the adjuvants. The dual-adjuvant vaccine accumulated in the lymph nodes and remained there for up to a month, allowing the immune system to build up a much greater number of antibodies against the HIV protein.

This strategy could lead to the development of vaccines that only need to be given once, for infectious diseases including HIV or SARS-CoV-2, the researchers say.

“This approach is compatible with many protein-based vaccines, so it offers the opportunity to engineer new formulations for these types of vaccines across a wide range of different diseases, such as influenza, SARS-CoV-2, or other pandemic outbreaks,” says J. Christopher Love, the Raymond A. and Helen E. St. Laurent Professor of Chemical Engineering at MIT, and a member of the Koch Institute for Integrative Cancer Research and the Ragon Institute of MGH, MIT, and Harvard.

Love and Darrell Irvine, a professor of immunology and microbiology at the Scripps Research Institute, are the senior authors of the study, which appears today in Science Translational Medicine. Kristen Rodrigues PhD ’23 and Yiming Zhang PhD ’25 are the lead authors of the paper.

More powerful vaccines

Most vaccines are delivered along with adjuvants, which help to stimulate a stronger immune response to the antigen. One adjuvant commonly used with protein-based vaccines, including those for hepatitis A and B, is aluminum hydroxide, also known as alum. This adjuvant works by activating the innate immune response, helping the body to form a stronger memory of the vaccine antigen.

Several years ago, Irvine developed another adjuvant based on saponin, an FDA-approved adjuvant derived from the bark of the Chilean soapbark tree. His work showed that nanoparticles containing both saponin and a molecule called MPLA, which promotes inflammation, worked better than saponin on its own. That nanoparticle, known as SMNP, is now being used as an adjuvant for an HIV vaccine that is currently in clinical trials.

Irvine and Love then tried combining alum and SMNP and showed that vaccines containing both of those adjuvants could generate even more powerful immune responses against either HIV or SARS-CoV-2.

In the new paper, the researchers wanted to explore why these two adjuvants work so well together to boost the immune response, specifically the B cell response. B cells produce antibodies that can circulate in the bloodstream and recognize a pathogen if the body is exposed to it again.

For this study, the researchers used an HIV protein called MD39 as their vaccine antigen, and anchored dozens of these proteins to each alum particle, along with SMNP.

After vaccinating mice with these particles, the researchers found that the vaccine accumulated in the lymph nodes — structures where B cells encounter antigens and undergo rapid mutations that generate antibodies with high affinity for a particular antigen. This process takes place within clusters of cells known as germinal centers.

The researchers showed that SMNP and alum helped the HIV antigen to penetrate through the protective layer of cells surrounding the lymph nodes without being broken down into fragments. The adjuvants also helped the antigens to remain intact in the lymph nodes for up to 28 days.

“As a result, the B cells that are cycling in the lymph nodes are constantly being exposed to the antigen over that time period, and they get the chance to refine their solution to the antigen,” Love says.

This approach may mimic what occurs during a natural infection, when antigens can remain in the lymph nodes for weeks, giving the body time to build up an immune response.

Antibody diversity

Single-cell RNA sequencing of B cells from the vaccinated mice revealed that the vaccine containing both adjuvants generated a much more diverse repertoire of B cells and antibodies. Mice that received the dual-adjuvant vaccine produced two to three times more unique B cells than mice that received just one of the adjuvants.

That increase in B cell number and diversity boosts the chances that the vaccine could generate broadly neutralizing antibodies — antibodies that can recognize a variety of strains of a given virus, such as HIV.

“When you think about the immune system sampling all of the possible solutions, the more chances we give it to identify an effective solution, the better,” Love says. “Generating broadly neutralizing antibodies is something that likely requires both the kind of approach that we showed here, to get that strong and diversified response, as well as antigen design to get the right part of the immunogen shown.”

Using these two adjuvants together could also contribute to the development of more potent vaccines against other infectious diseases, with just a single dose.

“What’s potentially powerful about this approach is that you can achieve long-term exposures based on a combination of adjuvants that are already reasonably well-understood, so it doesn’t require a different technology. It’s just combining features of these adjuvants to enable low-dose or potentially even single-dose treatments,” Love says.

The research was funded by the National Institutes of Health; the Koch Institute Support (core) Grant from the National Cancer Institute; the Ragon Institute of MGH, MIT, and Harvard; and the Howard Hughes Medical Institute.

The advanced semiconductor material gallium nitride will likely be key for the next generation of high-speed communication systems and the power electronics needed for state-of-the-art data centers.

Unfortunately, the high cost of gallium nitride (GaN) and the specialization required to incorporate this semiconductor material into conventional electronics have limited its use in commercial applications.

Now, researchers from MIT and elsewhere have developed a new fabrication process that integrates high-performance GaN transistors onto standard silicon CMOS chips in a way that is low-cost and scalable, and compatible with existing semiconductor foundries.

Their method involves building many tiny transistors on the surface of a GaN chip, cutting out each individual transistor, and then bonding just the necessary number of transistors onto a silicon chip using a low-temperature process that preserves the functionality of both materials.

The cost remains minimal since only a tiny amount of GaN material is added to the chip, but the resulting device can receive a significant performance boost from compact, high-speed transistors. In addition, by separating the GaN circuit into discrete transistors that can be spread over the silicon chip, the new technology is able to reduce the temperature of the overall system.

The researchers used this process to fabricate a power amplifier, an essential component in mobile phones, that achieves higher signal strength and efficiencies than devices with silicon transistors. In a smartphone, this could improve call quality, boost wireless bandwidth, enhance connectivity, and extend battery life.

Because their method fits into standard procedures, it could improve electronics that exist today as well as future technologies. Down the road, the new integration scheme could even enable quantum applications, as GaN performs better than silicon at the cryogenic temperatures essential for many types of quantum computing.

“If we can bring the cost down, improve the scalability, and, at the same time, enhance the performance of the electronic device, it is a no-brainer that we should adopt this technology. We’ve combined the best of what exists in silicon with the best possible gallium nitride electronics. These hybrid chips can revolutionize many commercial markets,” says Pradyot Yadav, an MIT graduate student and lead author of a paper on this method.

He is joined on the paper by fellow MIT graduate students Jinchen Wang and Patrick Darmawi-Iskandar; MIT postdoc John Niroula; senior authors Ulrich L. Rohde, a visiting scientist at the Microsystems Technology Laboratories (MTL), and Ruonan Han, an associate professor in the Department of Electrical Engineering and Computer Science (EECS) and member of MTL; and Tomás Palacios, the Clarence J. LeBel Professor of EECS, and director of MTL; as well as collaborators at Georgia Tech and the Air Force Research Laboratory. The research was recently presented at the IEEE Radio Frequency Integrated Circuits Symposium.

Swapping transistors

Gallium nitride is the second most widely used semiconductor in the world, just after silicon, and its unique properties make it ideal for applications such as lighting, radar systems and power electronics.

The material has been around for decades and, to get access to its maximum performance, it is important for chips made of GaN to be connected to digital chips made of silicon, also called CMOS chips. To enable this, some integration methods bond GaN transistors onto a CMOS chip by soldering the connections, but this limits how small the GaN transistors can be. The tinier the transistors, the higher the frequency at which they can work.

Other methods integrate an entire gallium nitride wafer on top of a silicon wafer, but using so much material is extremely costly, especially since the GaN is only needed in a few tiny transistors. The rest of the material in the GaN wafer is wasted.

“We wanted to combine the functionality of GaN with the power of digital chips made of silicon, but without having to compromise on either cost of bandwidth. We achieved that by adding super-tiny discrete gallium nitride transistors right on top of the silicon chip,” Yadav explains.

The new chips are the result of a multistep process.

First, a tightly packed collection of miniscule transistors is fabricated across the entire surface of a GaN wafer. Using very fine laser technology, they cut each one down to just the size of the transistor, which is 240 by 410 microns, forming what they call a dielet. (A micron is one millionth of a meter.)

Each transistor is fabricated with tiny copper pillars on top, which they use to bond directly to the copper pillars on the surface of a standard silicon CMOS chip. Copper to copper bonding can be done at temperatures below 400 degrees Celsius, which is low enough to avoid damaging either material.

Current GaN integration techniques require bonds that utilize gold, an expensive material that needs much higher temperatures and stronger bonding forces than copper. Since gold can contaminate the tools used in most semiconductor foundries, it typically requires specialized facilities.

“We wanted a process that was low-cost, low-temperature, and low-force, and copper wins on all of those related to gold. At the same time, it has better conductivity,” Yadav says.

A new tool

To enable the integration process, they created a specialized new tool that can carefully integrate the extremely tiny GaN transistor with the silicon chips. The tool uses a vacuum to hold the dielet as it moves on top of a silicon chip, zeroing in on the copper bonding interface with nanometer precision.

They used advanced microscopy to monitor the interface, and then when the dielet is in the right position, they apply heat and pressure to bond the GaN transistor to the chip.

“For each step in the process, I had to find a new collaborator who knew how to do the technique that I needed, learn from them, and then integrate that into my platform. It was two years of constant learning,” Yadav says.

Once the researchers had perfected the fabrication process, they demonstrated it by developing power amplifiers, which are radio frequency circuits that boost wireless signals.

Their devices achieved higher bandwidth and better gain than devices made with traditional silicon transistors. Each compact chip has an area of less than half a square millimeter.

In addition, because the silicon chip they used in their demonstration is based on Intel 16 22nm FinFET state-of-the-art metallization and passive options, they were able to incorporate components often used in silicon circuits, such as neutralization capacitors. This significantly improved the gain of the amplifier, bringing it one step closer to enabling the next generation of wireless technologies.

“To address the slowdown of Moore’s Law in transistor scaling, heterogeneous integration has emerged as a promising solution for continued system scaling, reduced form factor, improved power efficiency, and cost optimization. Particularly in wireless technology, the tight integration of compound semiconductors with silicon-based wafers is critical to realizing unified systems of front-end integrated circuits, baseband processors, accelerators, and memory for next-generation antennas-to-AI platforms. This work makes a significant advancement by demonstrating 3D integration of multiple GaN chips with silicon CMOS and pushes the boundaries of current technological capabilities,” says Atom Watanabe, a research scientist at IBM who was not involved with this paper.

This work is supported, in part, by the U.S. Department of Defense through the National Defense Science and Engineering Graduate (NDSEG) Fellowship Program and CHIMES, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation Program by the Department of Defense and the Defense Advanced Research Projects Agency (DARPA). Fabrication was carried out using facilities at MIT.Nano, the Air Force Research Laboratory, and Georgia Tech.

Research has shown that large language models (LLMs) tend to overemphasize information at the beginning and end of a document or conversation, while neglecting the middle.

This “position bias” means that, if a lawyer is using an LLM-powered virtual assistant to retrieve a certain phrase in a 30-page affidavit, the LLM is more likely to find the right text if it is on the initial or final pages.

MIT researchers have discovered the mechanism behind this phenomenon.

They created a theoretical framework to study how information flows through the machine-learning architecture that forms the backbone of LLMs. They found that certain design choices which control how the model processes input data can cause position bias.

Their experiments revealed that model architectures, particularly those affecting how information is spread across input words within the model, can give rise to or intensify position bias, and that training data also contribute to the problem.

In addition to pinpointing the origins of position bias, their framework can be used to diagnose and correct it in future model designs.

This could lead to more reliable chatbots that stay on topic during long conversations, medical AI systems that reason more fairly when handling a trove of patient data, and code assistants that pay closer attention to all parts of a program.

“These models are black boxes, so as an LLM user, you probably don’t know that position bias can cause your model to be inconsistent. You just feed it your documents in whatever order you want and expect it to work. But by understanding the underlying mechanism of these black-box models better, we can improve them by addressing these limitations,” says Xinyi Wu, a graduate student in the MIT Institute for Data, Systems, and Society (IDSS) and the Laboratory for Information and Decision Systems (LIDS), and first author of a paper on this research.

Her co-authors include Yifei Wang, an MIT postdoc; and senior authors Stefanie Jegelka, an associate professor of electrical engineering and computer science (EECS) and a member of IDSS and the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Ali Jadbabaie, professor and head of the Department of Civil and Environmental Engineering, a core faculty member of IDSS, and a principal investigator in LIDS. The research will be presented at the International Conference on Machine Learning.

Analyzing attention

LLMs like Claude, Llama, and GPT-4 are powered by a type of neural network architecture known as a transformer. Transformers are designed to process sequential data, encoding a sentence into chunks called tokens and then learning the relationships between tokens to predict what words comes next.

These models have gotten very good at this because of the attention mechanism, which uses interconnected layers of data processing nodes to make sense of context by allowing tokens to selectively focus on, or attend to, related tokens.

But if every token can attend to every other token in a 30-page document, that quickly becomes computationally intractable. So, when engineers build transformer models, they often employ attention masking techniques which limit the words a token can attend to.

For instance, a causal mask only allows words to attend to those that came before it.

Engineers also use positional encodings to help the model understand the location of each word in a sentence, improving performance.

The MIT researchers built a graph-based theoretical framework to explore how these modeling choices, attention masks and positional encodings, could affect position bias.

“Everything is coupled and tangled within the attention mechanism, so it is very hard to study. Graphs are a flexible language to describe the dependent relationship among words within the attention mechanism and trace them across multiple layers,” Wu says.

Their theoretical analysis suggested that causal masking gives the model an inherent bias toward the beginning of an input, even when that bias doesn’t exist in the data.

If the earlier words are relatively unimportant for a sentence’s meaning, causal masking can cause the transformer to pay more attention to its beginning anyway.

“While it is often true that earlier words and later words in a sentence are more important, if an LLM is used on a task that is not natural language generation, like ranking or information retrieval, these biases can be extremely harmful,” Wu says.

As a model grows, with additional layers of attention mechanism, this bias is amplified because earlier parts of the input are used more frequently in the model’s reasoning process.

They also found that using positional encodings to link words more strongly to nearby words can mitigate position bias. The technique refocuses the model’s attention in the right place, but its effect can be diluted in models with more attention layers.

And these design choices are only one cause of position bias — some can come from training data the model uses to learn how to prioritize words in a sequence.

“If you know your data are biased in a certain way, then you should also finetune your model on top of adjusting your modeling choices,” Wu says.

Lost in the middle

After they’d established a theoretical framework, the researchers performed experiments in which they systematically varied the position of the correct answer in text sequences for an information retrieval task.

The experiments showed a “lost-in-the-middle” phenomenon, where retrieval accuracy followed a U-shaped pattern. Models performed best if the right answer was located at the beginning of the sequence. Performance declined the closer it got to the middle before rebounding a bit if the correct answer was near the end.

Ultimately, their work suggests that using a different masking technique, removing extra layers from the attention mechanism, or strategically employing positional encodings could reduce position bias and improve a model’s accuracy.

“By doing a combination of theory and experiments, we were able to look at the consequences of model design choices that weren’t clear at the time. If you want to use a model in high-stakes applications, you must know when it will work, when it won’t, and why,” Jadbabaie says.

In the future, the researchers want to further explore the effects of positional encodings and study how position bias could be strategically exploited in certain applications.

“These researchers offer a rare theoretical lens into the attention mechanism at the heart of the transformer model. They provide a compelling analysis that clarifies longstanding quirks in transformer behavior, showing that attention mechanisms, especially with causal masks, inherently bias models toward the beginning of sequences. The paper achieves the best of both worlds — mathematical clarity paired with insights that reach into the guts of real-world systems,” says Amin Saberi, professor and director of the Stanford University Center for Computational Market Design, who was not involved with this work.

This research is supported, in part, by the U.S. Office of Naval Research, the National Science Foundation, and an Alexander von Humboldt Professorship.

MIT researchers have designed a compact, low-power receiver for 5G-compatible smart devices that is about 30 times more resilient to a certain type of interference than some traditional wireless receivers.

The low-cost receiver would be ideal for battery-powered internet of things (IoT) devices like environmental sensors, smart thermostats, or other devices that need to run continuously for a long time, such as health wearables, smart cameras, or industrial monitoring sensors.

The researchers’ chip uses a passive filtering mechanism that consumes less than a milliwatt of static power while protecting both the input and output of the receiver’s amplifier from unwanted wireless signals that could jam the device.

Key to the new approach is a novel arrangement of precharged, stacked capacitors, which are connected by a network of tiny switches. These miniscule switches need much less power to be turned on and off than those typically used in IoT receivers.

The receiver’s capacitor network and amplifier are carefully arranged to leverage a phenomenon in amplification that allows the chip to use much smaller capacitors than would typically be necessary. 

“This receiver could help expand the capabilities of IoT gadgets. Smart devices like health monitors or industrial sensors could become smaller and have longer battery lives. They would also be more reliable in crowded radio environments, such as factory floors or smart city networks,” says Soroush Araei, an electrical engineering and computer science (EECS) graduate student at MIT and lead author of a paper on the receiver.

He is joined on the paper by Mohammad Barzgari, a postdoc in the MIT Research Laboratory of Electronics (RLE); Haibo Yang, an EECS graduate student; and senior author Negar Reiskarimian, the X-Window Consortium Career Development Assistant Professor in EECS at MIT and a member of the Microsystems Technology Laboratories and RLE. The research was recently presented at the IEEE Radio Frequency Integrated Circuits Symposium.

A new standard

A receiver acts as the intermediary between an IoT device and its environment. Its job is to detect and amplify a wireless signal, filter out any interference, and then convert it into digital data for processing.

Traditionally, IoT receivers operate on fixed frequencies and suppress interference using a single narrow-band filter, which is simple and inexpensive.

But the new technical specifications of the 5G mobile network enable reduced-capability devices that are more affordable and energy-efficient. This opens a range of IoT applications to the faster data speeds and increased network capability of 5G. These next-generation IoT devices need receivers that can tune across a wide range of frequencies while still being cost-effective and low-power.

“This is extremely challenging because now we need to not only think about the power and cost of the receiver, but also flexibility to address numerous interferers that exist in the environment,” Araei says.

To reduce the size, cost, and power consumption of an IoT device, engineers can’t rely on the bulky, off-chip filters that are typically used in devices that operate on a wide frequency range.

One solution is to use a network of on-chip capacitors that can filter out unwanted signals. But these capacitor networks are prone to special type of signal noise known as harmonic interference.

In prior work, the MIT researchers developed a novel switch-capacitor network that targets these harmonic signals as early as possible in the receiver chain, filtering out unwanted signals before they are amplified and converted into digital bits for processing.

Shrinking the circuit

Here, they extended that approach by using the novel switch-capacitor network as the feedback path in an amplifier with negative gain. This configuration leverages the Miller effect, a phenomenon that enables small capacitors to behave like much larger ones.

“This trick lets us meet the filtering requirement for narrow-band IoT without physically large components, which drastically shrinks the size of the circuit,” Araei says.

Their receiver has an active area of less than 0.05 square millimeters.

One challenge the researchers had to overcome was determining how to apply enough voltage to drive the switches while keeping the overall power supply of the chip at only 0.6 volts.

In the presence of interfering signals, such tiny switches can turn on and off in error, especially if the voltage required for switching is extremely low.

To address this, the researchers came up with a novel solution, using a special circuit technique called bootstrap clocking. This method boosts the control voltage just enough to ensure the switches operate reliably while using less power and fewer components than traditional clock boosting methods.

Taken together, these innovations enable the new receiver to consume less than a milliwatt of power while blocking about 30 times more harmonic interference than traditional IoT receivers.

“Our chip also is very quiet, in terms of not polluting the airwaves. This comes from the fact that our switches are very small, so the amount of signal that can leak out of the antenna is also very small,” Araei adds.

Because their receiver is smaller than traditional devices and relies on switches and precharged capacitors instead of more complex electronics, it could be more cost-effective to fabricate. In addition, since the receiver design can cover a wide range of signal frequencies, it could be implemented on a variety of current and future IoT devices.

Now that they have developed this prototype, the researchers want to enable the receiver to operate without a dedicated power supply, perhaps by harvesting Wi-Fi or Bluetooth signals from the environment to power the chip.

This research is supported, in part, by the National Science Foundation.

In 2023, about 4.4 percent (176 terawatt-hours) of total energy consumption in the United States was by data centers that are essential for processing large quantities of information. Of that 176 TWh, approximately 100 TWh (57 percent) was used by CPU and GPU equipment. Energy requirements have escalated substantially in the past decade and will only continue to grow, making the development of energy-efficient computing crucial. 

Superconducting electronics have arisen as a promising alternative for classical and quantum computing, although their full exploitation for high-end computing requires a dramatic reduction in the amount of wiring linking ambient temperature electronics and low-temperature superconducting circuits. To make systems that are both larger and more streamlined, replacing commonplace components such as semiconductors with superconducting versions could be of immense value. It’s a challenge that has captivated MIT Plasma Science and Fusion Center senior research scientist Jagadeesh Moodera and his colleagues, who described a significant breakthrough in a recent Nature Electronics paper, “Efficient superconducting diodes and rectifiers for quantum circuitry.”

Moodera was working on a stubborn problem. One of the critical long-standing requirements is the need for the efficient conversion of AC currents into DC currents on a chip while operating at the extremely cold cryogenic temperatures required for superconductors to work efficiently. For example, in superconducting “energy-efficient rapid single flux quantum” (ERSFQ) circuits, the AC-to-DC issue is limiting ERSFQ scalability and preventing their use in larger circuits with higher complexities. To respond to this need, Moodera and his team created superconducting diode (SD)-based superconducting rectifiers — devices that can convert AC to DC on the same chip. These rectifiers would allow for the efficient delivery of the DC current necessary to operate superconducting classical and quantum processors.

Quantum computer circuits can only operate at temperatures close to 0 kelvins (absolute zero), and the way power is supplied must be carefully controlled to limit the effects of interference introduced by too much heat or electromagnetic noise. Most unwanted noise and heat come from the wires connecting cold quantum chips to room-temperature electronics. Instead, using superconducting rectifiers to convert AC currents into DC within a cryogenic environment reduces the number of wires, cutting down on heat and noise and enabling larger, more stable quantum systems.

In a 2023 experiment, Moodera and his co-authors developed SDs that are made of very thin layers of superconducting material that display nonreciprocal (or unidirectional) flow of current and could be the superconducting counterpart to standard semiconductors. Even though SDs have garnered significant attention, especially since 2020, up until this point the research has focused only on individual SDs for proof of concept. The group’s 2023 paper outlined how they created and refined a method by which SDs could be scaled for broader application. 

Now, by building a diode bridge circuit, they demonstrated the successful integration of four SDs and realized AC-to-DC rectification at cryogenic temperatures. 

The new approach described in their recent Nature Electronics paper will significantly cut down on the thermal and electromagnetic noise traveling from ambient into cryogenic circuitry, enabling cleaner operation. The SDs could also potentially serve as isolators/circulators, assisting in insulating qubit signals from external influence. The successful assimilation of multiple SDs into the first integrated SD circuit represents a key step toward making superconducting computing a commercial reality. 

“Our work opens the door to the arrival of highly energy-efficient, practical superconductivity-based supercomputers in the next few years,” says Moodera. “Moreover, we expect our research to enhance the qubit stability while boosting the quantum computing program, bringing its realization closer." Given the multiple beneficial roles these components could play, Moodera and his team are already working toward the integration of such devices into actual superconducting logic circuits, including in dark matter detection circuits that are essential to the operation of experiments at CERN and LUX-ZEPLIN in at the Berkeley National Lab.

This work was partially funded by MIT Lincoln Laboratory’s Advanced Concepts Committee, the U.S. National Science Foundation, U.S. Army Research Office, and U.S. Air Force Office of Scientific Research. 

This work was carried out, in part, through the use of MIT.nano’s facilities.

In 1989, New York City opened a new jail. But not on dry land. The city leased a barge, then called the “Bibby Resolution,” which had been topped with five stories of containers made into housing, and anchored it in the East River. For five years, the vessel lodged inmates.

A floating detention center is a curiosity. But then, the entire history of this barge is curious. Built in 1979 in Sweden, it housed British troops during the Falkland Islands war with Argentina, became worker housing for Volkswagen employees in West Germany, got sent to New York, also became a detention center off the coast of England, then finally was deployed as oil worker housing off the coast of Nigeria. The barge has had nine names, several owners, and flown the flags of five countries.

In this one vessel, then, we can see many currents: globalization, the transience of economic activity, and the hazy world of transactions many analysts and observers call “the offshore,” the lightly regulated sphere of economic activity that encourages short-term actions.

“The offshore presents a quick and potentially cheap solution to a crisis,” says MIT lecturer Ian Kumekawa. “It is not a durable solution. The story of the barge is the story of it being used as a quick fix in all sorts of crises. Then these expediences become the norm, and people get used to them and have an expectation that this is the way the world works.”

Now Kumekawa, a historian who started teaching as a lecturer at MIT earlier this year, explores the ship’s entire history in “Empty Vessel: The Global Economy in One Barge,” just published by Knopf and John Murray. In it, he traces the barge’s trajectory and the many economic and geopolitical changes that helped create the ship’s distinctive deployments around the world.

“The book is about a barge, but it’s also about the developing, emerging offshore world, where you see these layers of globalization, financialization, privatization, and the dissolution of territoriality and orders,” Kumekawa says. “The barge is a vehicle through which I can tell the story of those layers together.”

“Never meant to be permanent”

Kumekawa first found out about the vessel several years ago; New York City obtained another floating detention center in the 1990s, which prompted Kumekawa to start looking into the past of the older jail ship, the former “Bibby Resolution,” from the 1990s. The more he found out about its distinctive past, the more curious he became.

“You start pulling on a thread, and you realize you can keep pulling,” Kumekawa says.

The barge Kumekawa follows in the book was built in Sweden in 1979 as the “Balder Scapa.” Even then, commerce was plenty globalized: The vessel was commissioned by a Norwegian shell company, with negotiations run by an expatriate Swedish shipping agent whose firm was registered in Panama and used a Miami bank.

The barge was built at an inflection point following the economic slowdown and oil shocks of the 1970s. Manufacturing was on the verge of declining in both Western Europe and the U.S.; about half as many people now work in manufacturing in those regions, compared to 1960. Companies were looking to find cheaper global locations for production, reinforcing the sense that economic activity was now less durable in any given place.

The barge became part of this transience. The five-story accommodation block was added in the early 1980s; in 1983 it was re-registered in the UK and sent to the Falkland Islands as a troop accommodation named the “COASTEL 3.” Then it was re-registered in the Bahamas and sent to Emden, West Germany, as housing for Volkswagen workers. The vessel then served its stints as inmate housing — first in New York, then off the coast of England from 1997 to 2005. By 2010, it had been re-re-re-registered, in St. Vincent and Grenadines, and was housing oil workers off the coast of Nigeria.

“Globalization is more about flow than about stocks, and the barge is a great example of that,” Kumekawa says. “It’s always on the move, and never meant to be a permanent container. It’s understood people are going to be passing through.”

As Kumekawa explores in the book, this sense of social dislocation overlapped with the shrinking of state capacity, as many states increasingly encouraged companies to pursue globalized production and lightly regulated financial activities in numerous jurisdictions, in the hope it would enhance growth. And it has, albeit with unresolved questions about who the benefits accrue to, the social dislocation of workers, and more.

“In a certain sense it’s not an erosion of state power at all,” Kumekawa says. “These states are making very active choices to use offshore tools, to circumvent certain roadblocks.” He adds: “What happens in the 1970s and certainly in the 1980s is that the offshore comes into its own as an entity, and didn’t exist in the same way even in the 1950s and 1960s. There’s a money interest in that, and there’s a political interest as well.”

Abstract forces, real materials and people

Kumekawa is a scholar with a strong interest in economic history; his previous book, “The First Serious Optimist: A.C. Pigou and the Birth of Welfare Economics,” was published in 2017. This coming fall, Kumekawa will be team-teaching a class on the relationship between economics and history, along with MIT economists Abhijit Banerjee and Jacob Moscona.

Working on “Empty Vessel” also necessitated that Kumekawa use a variety of research techniques, from archival work to journalistic interviews with people who knew the vessel well.

“I had a wonderful set of conversations with the man who was the last bargemaster,” Kumekawa says. “He was the person in effect steering the vessel for many years. He was so aware of all of the forces at play — the market for oil, the prices of accommodations, the regulations, the fact no one had reinforced the frame.”

“Empty Vessel” has already received critical acclaim. Reviewing it in The New York Times, Jennifer Szalai writes that this “elegant and enlightening book is an impressive feat.”

For his part, Kumekawa also took inspiration from a variety of writings about ships, voyages, commerce, and exploration, recognizing that these vessels contain stories and vignettes that illuminate the wider world.

“Ships work very well as devices connecting the global and the local,” he says. Using the barge as the organizing principle of his book, Kumekawa adds, “makes a whole bunch of abstract processes very concrete. The offshore itself is an abstraction, but it’s also entirely dependent on physical infrastructure and physical places. My hope for the book is it reinforces the material dimension of these abstract global forces.”

The long-term aspirational goal of the Paris Agreement on climate change is to cap global warming at 1.5 degrees Celsius above preindustrial levels, and thereby reduce the frequency and severity of floods, droughts, wildfires, and other extreme weather events. Achieving that goal will require a massive reduction in global carbon dioxide (CO₂) emissions across all economic sectors. A major roadblock, however, could be the industrial sector, which accounts for roughly 25 percent of global energy- and process-related CO₂ emissions — particularly within the iron and steel sector, industry’s largest emitter of CO₂.

Iron and steel production now relies heavily on fossil fuels (coal or natural gas) for heat, converting iron ore to iron, and making steel strong. Steelmaking could be decarbonized by a combination of several methods, including carbon capture technology, the use of low- or zero-carbon fuels, and increased use of recycled steel. Now a new study in the Journal of Cleaner Production systematically explores the viability of different iron-and-steel decarbonization strategies.

Today’s strategy menu includes improving energy efficiency, switching fuels and technologies, using more scrap steel, and reducing demand. Using the MIT Economic Projection and Policy Analysis model, a multi-sector, multi-region model of the world economy, researchers at MIT, the University of Illinois at Urbana-Champaign, and ExxonMobil Technology and Engineering Co. evaluate the decarbonization potential of replacing coal-based production processes with electric arc furnaces (EAF), along with either scrap steel or “direct reduced iron” (DRI), which is fueled by natural gas with carbon capture and storage (NG CCS DRI-EAF) or by hydrogen (H₂ DRI-EAF).

Under a global climate mitigation scenario aligned with the 1.5 C climate goal, these advanced steelmaking technologies could result in deep decarbonization of the iron and steel sector by 2050, as long as technology costs are low enough to enable large-scale deployment. Higher costs would favor the replacement of coal with electricity and natural gas, greater use of scrap steel, and reduced demand, resulting in a more-than-50-percent reduction in emissions relative to current levels. Lower technology costs would enable massive deployment of NG CCS DRI-EAF or H₂ DRI-EAF, reducing emissions by up to 75 percent.

Even without adoption of these advanced technologies, the iron-and-steel sector could significantly reduce its CO₂ emissions intensity (how much CO₂ is released per unit of production) with existing steelmaking technologies, primarily by replacing coal with gas and electricity (especially if it is generated by renewable energy sources), using more scrap steel, and implementing energy efficiency measures.

“The iron and steel industry needs to combine several strategies to substantially reduce its emissions by mid-century, including an increase in recycling, but investing in cost reductions in hydrogen pathways and carbon capture and sequestration will enable even deeper emissions mitigation in the sector,” says study supervising author Sergey Paltsev, deputy director of the MIT Center for Sustainability Science and Strategy (MIT CS3) and a senior research scientist at the MIT Energy Initiative (MITEI).

This study was supported by MIT CS3 and ExxonMobil through its membership in MITEI.

In 15 TED Talk-style presentations, MIT faculty recently discussed their pioneering research that incorporates social, ethical, and technical considerations and expertise, each supported by seed grants established by the Social and Ethical Responsibilities of Computing (SERC), a cross-cutting initiative of the MIT Schwarzman College of Computing. The call for proposals last summer was met with nearly 70 applications. A committee with representatives from every MIT school and the college convened to select the winning projects that received up to $100,000 in funding.

“SERC is committed to driving progress at the intersection of computing, ethics, and society. The seed grants are designed to ignite bold, creative thinking around the complex challenges and possibilities in this space,” said Nikos Trichakis, co-associate dean of SERC and the J.C. Penney Professor of Management. “With the MIT Ethics of Computing Research Symposium, we felt it important to not just showcase the breadth and depth of the research that’s shaping the future of ethical computing, but to invite the community to be part of the conversation as well.”

“What you’re seeing here is kind of a collective community judgment about the most exciting work when it comes to research, in the social and ethical responsibilities of computing being done at MIT,” said Caspar Hare, co-associate dean of SERC and professor of philosophy.

The full-day symposium on May 1 was organized around four key themes: responsible health-care technology, artificial intelligence governance and ethics, technology in society and civic engagement, and digital inclusion and social justice. Speakers delivered thought-provoking presentations on a broad range of topics, including algorithmic bias, data privacy, the social implications of artificial intelligence, and the evolving relationship between humans and machines. The event also featured a poster session, where student researchers showcased projects they worked on throughout the year as SERC Scholars.

Highlights from the MIT Ethics of Computing Research Symposium in each of the theme areas, many of which are available to watch on YouTube, included:

Making the kidney transplant system fairer

Policies regulating the organ transplant system in the United States are made by a national committee that often takes more than six months to create, and then years to implement, a timeline that many on the waiting list simply can’t survive.

Dimitris Bertsimas, vice provost for open learning, associate dean of business analytics, and Boeing Professor of Operations Research, shared his latest work in analytics for fair and efficient kidney transplant allocation. Bertsimas’ new algorithm examines criteria like geographic location, mortality, and age in just 14 seconds, a monumental change from the usual six hours.

Bertsimas and his team work closely with the United Network for Organ Sharing (UNOS), a nonprofit that manages most of the national donation and transplant system through a contract with the federal government. During his presentation, Bertsimas shared a video from James Alcorn, senior policy strategist at UNOS, who offered this poignant summary of the impact the new algorithm has:

“This optimization radically changes the turnaround time for evaluating these different simulations of policy scenarios. It used to take us a couple months to look at a handful of different policy scenarios, and now it takes a matter of minutes to look at thousands and thousands of scenarios. We are able to make these changes much more rapidly, which ultimately means that we can improve the system for transplant candidates much more rapidly.”

The ethics of AI-generated social media content

As AI-generated content becomes more prevalent across social media platforms, what are the implications of disclosing (or not disclosing) that any part of a post was created by AI? Adam Berinsky, Mitsui Professor of Political Science, and Gabrielle Péloquin-Skulski, PhD student in the Department of Political Science, explored this question in a session that examined recent studies on the impact of various labels on AI-generated content.

In a series of surveys and experiments affixing labels to AI-generated posts, the researchers looked at how specific words and descriptions impacted users’ perception of deception, their intent to engage with the post, and ultimately if the post was true or false.

“The big takeaway from our initial set of findings is that one size doesn’t fit all,” said Péloquin-Skulski. “We found that labeling AI-generated images with a process-oriented label reduces belief in both false and true posts. This is quite problematic, as labeling intends to reduce people’s belief in false information, not necessarily true information. This suggests that labels combining both process and veracity might be better at countering AI-generated misinformation.”

Using AI to increase civil discourse online

“Our research aims to address how people increasingly want to have a say in the organizations and communities they belong to,” Lily Tsai explained in a session on experiments in generative AI and the future of digital democracy. Tsai, Ford Professor of Political Science and director of the MIT Governance Lab, is conducting ongoing research with Alex Pentland, Toshiba Professor of Media Arts arts Sciences, and a larger team.

Online deliberative platforms have recently been rising in popularity across the United States in both public- and private-sector settings. Tsai explained that with technology, it’s now possible for everyone to have a say — but doing so can be overwhelming, or even feel unsafe. First, too much information is available, and secondly, online discourse has become increasingly “uncivil.”

The group focuses on “how we can build on existing technologies and improve them with rigorous, interdisciplinary research, and how we can innovate by integrating generative AI to enhance the benefits of online spaces for deliberation.” They have developed their own AI-integrated platform for deliberative democracy, DELiberation.io, and rolled out four initial modules. All studies have been in the lab so far, but they are also working on a set of forthcoming field studies, the first of which will be in partnership with the government of the District of Columbia.

Tsai told the audience, “If you take nothing else from this presentation, I hope that you’ll take away this — that we should all be demanding that technologies that are being developed are assessed to see if they have positive downstream outcomes, rather than just focusing on maximizing the number of users.”

A public think tank that considers all aspects of AI

When Catherine D’Ignazio, associate professor of urban science and planning, and Nikko Stevens, postdoc at the Data + Feminism Lab at MIT, initially submitted their funding proposal, they weren’t intending to develop a think tank, but a framework — one that articulated how artificial intelligence and machine learning work could integrate community methods and utilize participatory design.

In the end, they created Liberatory AI, which they describe as a “rolling public think tank about all aspects of AI.” D’Ignazio and Stevens gathered 25 researchers from a diverse array of institutions and disciplines who authored more than 20 position papers examining the most current academic literature on AI systems and engagement. They intentionally grouped the papers into three distinct themes: the corporate AI landscape, dead ends, and ways forward.

“Instead of waiting for Open AI or Google to invite us to participate in the development of their products, we’ve come together to contest the status quo, think bigger-picture, and reorganize resources in this system in hopes of a larger societal transformation,” said D’Ignazio.

As more connected devices demand an increasing amount of bandwidth for tasks like teleworking and cloud computing, it will become extremely challenging to manage the finite amount of wireless spectrum available for all users to share.

Engineers are employing artificial intelligence to dynamically manage the available wireless spectrum, with an eye toward reducing latency and boosting performance. But most AI methods for classifying and processing wireless signals are power-hungry and can’t operate in real-time.

Now, MIT researchers have developed a novel AI hardware accelerator that is specifically designed for wireless signal processing. Their optical processor performs machine-learning computations at the speed of light, classifying wireless signals in a matter of nanoseconds.

The photonic chip is about 100 times faster than the best digital alternative, while converging to about 95 percent accuracy in signal classification. The new hardware accelerator is also scalable and flexible, so it could be used for a variety of high-performance computing applications. At the same time, it is smaller, lighter, cheaper, and more energy-efficient than digital AI hardware accelerators.

The device could be especially useful in future 6G wireless applications, such as cognitive radios that optimize data rates by adapting wireless modulation formats to the changing wireless environment.

By enabling an edge device to perform deep-learning computations in real-time, this new hardware accelerator could provide dramatic speedups in many applications beyond signal processing. For instance, it could help autonomous vehicles make split-second reactions to environmental changes or enable smart pacemakers to continuously monitor the health of a patient’s heart.

“There are many applications that would be enabled by edge devices that are capable of analyzing wireless signals. What we’ve presented in our paper could open up many possibilities for real-time and reliable AI inference. This work is the beginning of something that could be quite impactful,” says Dirk Englund, a professor in the MIT Department of Electrical Engineering and Computer Science, principal investigator in the Quantum Photonics and Artificial Intelligence Group and the Research Laboratory of Electronics (RLE), and senior author of the paper.

He is joined on the paper by lead author Ronald Davis III PhD ’24; Zaijun Chen, a former MIT postdoc who is now an assistant professor at the University of Southern California; and Ryan Hamerly, a visiting scientist at RLE and senior scientist at NTT Research. The research appears today in Science Advances.

Light-speed processing  

State-of-the-art digital AI accelerators for wireless signal processing convert the signal into an image and run it through a deep-learning model to classify it. While this approach is highly accurate, the computationally intensive nature of deep neural networks makes it infeasible for many time-sensitive applications.

Optical systems can accelerate deep neural networks by encoding and processing data using light, which is also less energy intensive than digital computing. But researchers have struggled to maximize the performance of general-purpose optical neural networks when used for signal processing, while ensuring the optical device is scalable.

By developing an optical neural network architecture specifically for signal processing, which they call a multiplicative analog frequency transform optical neural network (MAFT-ONN), the researchers tackled that problem head-on.

The MAFT-ONN addresses the problem of scalability by encoding all signal data and performing all machine-learning operations within what is known as the frequency domain — before the wireless signals are digitized.

The researchers designed their optical neural network to perform all linear and nonlinear operations in-line. Both types of operations are required for deep learning.

Thanks to this innovative design, they only need one MAFT-ONN device per layer for the entire optical neural network, as opposed to other methods that require one device for each individual computational unit, or “neuron.”

“We can fit 10,000 neurons onto a single device and compute the necessary multiplications in a single shot,” Davis says.   

The researchers accomplish this using a technique called photoelectric multiplication, which dramatically boosts efficiency. It also allows them to create an optical neural network that can be readily scaled up with additional layers without requiring extra overhead.

Results in nanoseconds

MAFT-ONN takes a wireless signal as input, processes the signal data, and passes the information along for later operations the edge device performs. For instance, by classifying a signal’s modulation, MAFT-ONN would enable a device to automatically infer the type of signal to extract the data it carries.

One of the biggest challenges the researchers faced when designing MAFT-ONN was determining how to map the machine-learning computations to the optical hardware.

“We couldn’t just take a normal machine-learning framework off the shelf and use it. We had to customize it to fit the hardware and figure out how to exploit the physics so it would perform the computations we wanted it to,” Davis says.

When they tested their architecture on signal classification in simulations, the optical neural network achieved 85 percent accuracy in a single shot, which can quickly converge to more than 99 percent accuracy using multiple measurements.  MAFT-ONN only required about 120 nanoseconds to perform entire process.

“The longer you measure, the higher accuracy you will get. Because MAFT-ONN computes inferences in nanoseconds, you don’t lose much speed to gain more accuracy,” Davis adds.

While state-of-the-art digital radio frequency devices can perform machine-learning inference in a microseconds, optics can do it in nanoseconds or even picoseconds.

Moving forward, the researchers want to employ what are known as multiplexing schemes so they could perform more computations and scale up the MAFT-ONN. They also want to extend their work into more complex deep learning architectures that could run transformer models or LLMs.

This work was funded, in part, by the U.S. Army Research Laboratory, the U.S. Air Force, MIT Lincoln Laboratory, Nippon Telegraph and Telephone, and the National Science Foundation.

Art restoration takes steady hands and a discerning eye. For centuries, conservators have restored paintings by identifying areas needing repair, then mixing an exact shade to fill in one area at a time. Often, a painting can have thousands of tiny regions requiring individual attention. Restoring a single painting can take anywhere from a few weeks to over a decade.

In recent years, digital restoration tools have opened a route to creating virtual representations of original, restored works. These tools apply techniques of computer vision, image recognition, and color matching, to generate a “digitally restored” version of a painting relatively quickly.

Still, there has been no way to translate digital restorations directly onto an original work, until now. In a paper appearing today in the journal Nature, Alex Kachkine, a mechanical engineering graduate student at MIT, presents a new method he’s developed to physically apply a digital restoration directly onto an original painting.

The restoration is printed on a very thin polymer film, in the form of a mask that can be aligned and adhered to an original painting. It can also be easily removed. Kachkine says that a digital file of the mask can be stored and referred to by future conservators, to see exactly what changes were made to restore the original painting.

“Because there’s a digital record of what mask was used, in 100 years, the next time someone is working with this, they’ll have an extremely clear understanding of what was done to the painting,” Kachkine says. “And that’s never really been possible in conservation before.”

As a demonstration, he applied the method to a highly damaged 15th century oil painting. The method automatically identified 5,612 separate regions in need of repair, and filled in these regions using 57,314 different colors. The entire process, from start to finish, took 3.5 hours, which he estimates is about 66 times faster than traditional restoration methods.

Kachkine acknowledges that, as with any restoration project, there are ethical issues to consider, in terms of whether a restored version is an appropriate representation of an artist’s original style and intent. Any application of his new method, he says, should be done in consultation with conservators with knowledge of a painting’s history and origins.

“There is a lot of damaged art in storage that might never be seen,” Kachkine says. “Hopefully with this new method, there’s a chance we’ll see more art, which I would be delighted by.”

Digital connections

The new restoration process started as a side project. In 2021, as Kachkine made his way to MIT to start his PhD program in mechanical engineering, he drove up the East Coast and made a point to visit as many art galleries as he could along the way.

“I’ve been into art for a very long time now, since I was a kid,” says Kachkine, who restores paintings as a hobby, using traditional hand-painting techniques. As he toured galleries, he came to realize that the art on the walls is only a fraction of the works that galleries hold. Much of the art that galleries acquire is stored away because the works are aged or damaged, and take time to properly restore.

“Restoring a painting is fun, and it’s great to sit down and infill things and have a nice evening,” Kachkine says. “But that’s a very slow process.”

As he has learned, digital tools can significantly speed up the restoration process. Researchers have developed artificial intelligence algorithms that quickly comb through huge amounts of data. The algorithms learn connections within this visual data, which they apply to generate a digitally restored version of a particular painting, in a way that closely resembles the style of an artist or time period. However, such digital restorations are usually displayed virtually or printed as stand-alone works and cannot be directly applied to retouch original art.

“All this made me think: If we could just restore a painting digitally, and effect the results physically, that would resolve a lot of pain points and drawbacks of a conventional manual process,” Kachkine says.

“Align and restore”

For the new study, Kachkine developed a method to physically apply a digital restoration onto an original painting, using a 15th-century painting that he acquired when he first came to MIT. His new method involves first using traditional techniques to clean a painting and remove any past restoration efforts.

“This painting is almost 600 years old and has gone through conservation many times,” he says. “In this case there was a fair amount of overpainting, all of which has to be cleaned off to see what’s actually there to begin with.”

He scanned the cleaned painting, including the many regions where paint had faded or cracked. He then used existing artificial intelligence algorithms to analyze the scan and create a virtual version of what the painting likely looked like in its original state.

Then, Kachkine developed software that creates a map of regions on the original painting that require infilling, along with the exact colors needed to match the digitally restored version. This map is then translated into a physical, two-layer mask that is printed onto thin polymer-based films. The first layer is printed in color, while the second layer is printed in the exact same pattern, but in white.

“In order to fully reproduce color, you need both white and color ink to get the full spectrum,” Kachkine explains. “If those two layers are misaligned, that’s very easy to see. So I also developed a few computational tools, based on what we know of human color perception, to determine how small of a region we can practically align and restore.”

Kachkine used high-fidelity commercial inkjets to print the mask’s two layers, which he carefully aligned and overlaid by hand onto the original painting and adhered with a thin spray of conventional varnish. The printed films are made from materials that can be easily dissolved with conservation-grade solutions, in case conservators need to reveal the original, damaged work. The digital file of the mask can also be saved as a detailed record of what was restored.

For the painting that Kachkine used, the method was able to fill in thousands of losses in just a few hours. “A few years ago, I was restoring this baroque Italian painting with probably the same order magnitude of losses, and it took me nine months of part-time work,” he recalls. “The more losses there are, the better this method is.”

He estimates that the new method can be orders of magnitude faster than traditional, hand-painted approaches. If the method is adopted widely, he emphasizes that conservators should be involved at every step in the process, to ensure that the final work is in keeping with an artist’s style and intent.

“It will take a lot of deliberation about the ethical challenges involved at every stage in this process to see how can this be applied in a way that’s most consistent with conservation principles,” he says. “We’re setting up a framework for developing further methods. As others work on this, we’ll end up with methods that are more precise.”

This work was supported, in part, by the John O. and Katherine A. Lutz Memorial Fund. The research was carried out, in part, through the use of equipment and facilities at MIT.Nano, with additional support from the MIT Microsystems Technology Laboratories, the MIT Department of Mechanical Engineering, and the MIT Libraries.

Today, 2.2 billion people in the world lack access to safe drinking water. In the United States, more than 46 million people experience water insecurity, living with either no running water or water that is unsafe to drink. The increasing need for drinking water is stretching traditional resources such as rivers, lakes, and reservoirs.

To improve access to safe and affordable drinking water, MIT engineers are tapping into an unconventional source: the air. The Earth’s atmosphere contains millions of billions of gallons of water in the form of vapor. If this vapor can be efficiently captured and condensed, it could supply clean drinking water in places where traditional water resources are inaccessible.

With that goal in mind, the MIT team has developed and tested a new atmospheric water harvester and shown that it efficiently captures water vapor and produces safe drinking water across a range of relative humidities, including dry desert air.

The new device is a black, window-sized vertical panel, made from a water-absorbent hydrogel material, enclosed in a glass chamber coated with a cooling layer. The hydrogel resembles black bubble wrap, with small dome-shaped structures that swell when the hydrogel soaks up water vapor. When the captured vapor evaporates, the domes shrink back down in an origami-like transformation. The evaporated vapor then condenses on the the glass, where it can flow down and out through a tube, as clean and drinkable water.

The system runs entirely on its own, without a power source, unlike other designs that require batteries, solar panels, or electricity from the grid. The team ran the device for over a week in Death Valley, California — the driest region in North America. Even in very low-humidity conditions, the device squeezed drinking water from the air at rates of up to 160 milliliters (about two-thirds of a cup) per day.

The team estimates that multiple vertical panels, set up in a small array, could passively supply a household with drinking water, even in arid desert environments. What’s more, the system’s water production should increase with humidity, supplying drinking water in temperate and tropical climates.

“We have built a meter-scale device that we hope to deploy in resource-limited regions, where even a solar cell is not very accessible,” says Xuanhe Zhao, the Uncas and Helen Whitaker Professor of Mechanical Engineering and Civil and Environmental Engineering at MIT. “It’s a test of feasibility in scaling up this water harvesting technology. Now people can build it even larger, or make it into parallel panels, to supply drinking water to people and achieve real impact.”

Zhao and his colleagues present the details of the new water harvesting design in a paper appearing today in the journal Nature Water. The study’s lead author is former MIT postdoc “Will” Chang Liu, who is currently an assistant professor at the National University of Singapore (NUS). MIT co-authors include Xiao-Yun Yan, Shucong Li, and Bolei Deng, along with collaborators from multiple other institutions.

Carrying capacity

Hydrogels are soft, porous materials that are made mainly from water and a microscopic network of interconnecting polymer fibers. Zhao’s group at MIT has primarily explored the use of hydrogels in biomedical applications, including adhesive coatings for medical implants, soft and flexible electrodes, and noninvasive imaging stickers.

“Through our work with soft materials, one property we know very well is the way hydrogel is very good at absorbing water from air,” Zhao says.

Researchers are exploring a number of ways to harvest water vapor for drinking water. Among the most efficient so far are devices made from metal-organic frameworks, or MOFs — ultra-porous materials that have also been shown to capture water from dry desert air. But the MOFs do not swell or stretch when absorbing water, and are limited in vapor-carrying capacity.

Water from air

The group’s new hydrogel-based water harvester addresses another key problem in similar designs. Other groups have designed water harvesters out of micro- or nano-porous hydrogels. But the water produced from these designs can be salty, requiring additional filtering. Salt is a naturally absorbent material, and researchers embed salts — typically, lithium chloride — in hydrogel to increase the material’s water absorption. The drawback, however, is that this salt can leak out with the water when it is eventually collected.

The team’s new design significantly limits salt leakage. Within the hydrogel itself, they included an extra ingredient: glycerol, a liquid compound that naturally stabilizes salt, keeping it within the gel rather than letting it crystallize and leak out with the water. The hydrogel itself has a microstructure that lacks nanoscale pores, which further prevents salt from escaping the material. The salt levels in the water they collected were below the standard threshold for safe drinking water, and significantly below the levels produced by many other hydrogel-based designs.

In addition to tuning the hydrogel’s composition, the researchers made improvements to its form. Rather than keeping the gel as a flat sheet, they molded it into a pattern of small domes resembling bubble wrap, that act to increase the gel’s surface area, along with the amount of water vapor it can absorb.

The researchers fabricated a half-square-meter of hydrogel and encased the material in a window-like glass chamber. They coated the exterior of the chamber with a special polymer film, which helps to cool the glass and stimulates any water vapor in the hydrogel to evaporate and condense onto the glass. They installed a simple tubing system to collect the water as it flows down the glass.

In November 2023, the team traveled to Death Valley, California, and set up the device as a vertical panel. Over seven days, they took measurements as the hydrogel absorbed water vapor during the night (the time of day when water vapor in the desert is highest). In the daytime, with help from the sun, the harvested water evaporated out from the hydrogel and condensed onto the glass.

Over this period, the device worked across a range of humidities, from 21 to 88 percent, and produced between 57 and 161.5 milliliters of drinking water per day. Even in the driest conditions, the device harvested more water than other passive and some actively powered designs.

“This is just a proof-of-concept design, and there are a lot of things we can optimize,” Liu says. “For instance, we could have a multipanel design. And we’re working on a next generation of the material to further improve its intrinsic properties.”

“We imagine that you could one day deploy an array of these panels, and the footprint is very small because they are all vertical,” says Zhao, who has plans to further test the panels in many resource-limited regions. “Then you could have many panels together, collecting water all the time, at household scale.”

This work was supported, in part, by the MIT J-WAFS Water and Food Seed Grant, the MIT-Chinese University of Hong Kong collaborative research program, and the UM6P-MIT collaborative research program.

The human brain is very good at solving complicated problems. One reason for that is that humans can break problems apart into manageable subtasks that are easy to solve one at a time.

This allows us to complete a daily task like going out for coffee by breaking it into steps: getting out of our office building, navigating to the coffee shop, and once there, obtaining the coffee. This strategy helps us to handle obstacles easily. For example, if the elevator is broken, we can revise how we get out of the building without changing the other steps.

While there is a great deal of behavioral evidence demonstrating humans’ skill at these complicated tasks, it has been difficult to devise experimental scenarios that allow precise characterization of the computational strategies we use to solve problems.

In a new study, MIT researchers have successfully modeled how people deploy different decision-making strategies to solve a complicated task — in this case, predicting how a ball will travel through a maze when the ball is hidden from view. The human brain cannot perform this task perfectly because it is impossible to track all of the possible trajectories in parallel, but the researchers found that people can perform reasonably well by flexibly adopting two strategies known as hierarchical reasoning and counterfactual reasoning.

The researchers were also able to determine the circumstances under which people choose each of those strategies.

“What humans are capable of doing is to break down the maze into subsections, and then solve each step using relatively simple algorithms. Effectively, when we don’t have the means to solve a complex problem, we manage by using simpler heuristics that get the job done,” says Mehrdad Jazayeri, a professor of brain and cognitive sciences, a member of MIT’s McGovern Institute for Brain Research, an investigator at the Howard Hughes Medical Institute, and the senior author of the study.

Mahdi Ramadan PhD ’24 and graduate student Cheng Tang are the lead authors of the paper, which appears today in Nature Human Behavior. Nicholas Watters PhD ’25 is also a co-author.

Rational strategies

When humans perform simple tasks that have a clear correct answer, such as categorizing objects, they perform extremely well. When tasks become more complex, such as planning a trip to your favorite cafe, there may no longer be one clearly superior answer. And, at each step, there are many things that could go wrong. In these cases, humans are very good at working out a solution that will get the task done, even though it may not be the optimal solution.

Those solutions often involve problem-solving shortcuts, or heuristics. Two prominent heuristics humans commonly rely on are hierarchical and counterfactual reasoning. Hierarchical reasoning is the process of breaking down a problem into layers, starting from the general and proceeding toward specifics. Counterfactual reasoning involves imagining what would have happened if you had made a different choice. While these strategies are well-known, scientists don’t know much about how the brain decides which one to use in a given situation.

“This is really a big question in cognitive science: How do we problem-solve in a suboptimal way, by coming up with clever heuristics that we chain together in a way that ends up getting us closer and closer until we solve the problem?” Jazayeri says.

To overcome this, Jazayeri and his colleagues devised a task that is just complex enough to require these strategies, yet simple enough that the outcomes and the calculations that go into them can be measured.

The task requires participants to predict the path of a ball as it moves through four possible trajectories in a maze. Once the ball enters the maze, people cannot see which path it travels. At two junctions in the maze, they hear an auditory cue when the ball reaches that point. Predicting the ball’s path is a task that is impossible for humans to solve with perfect accuracy.

“It requires four parallel simulations in your mind, and no human can do that. It’s analogous to having four conversations at a time,” Jazayeri says. “The task allows us to tap into this set of algorithms that the humans use, because you just can’t solve it optimally.”

The researchers recruited about 150 human volunteers to participate in the study. Before each subject began the ball-tracking task, the researchers evaluated how accurately they could estimate timespans of several hundred milliseconds, about the length of time it takes the ball to travel along one arm of the maze.

For each participant, the researchers created computational models that could predict the patterns of errors that would be seen for that participant (based on their timing skill) if they were running parallel simulations, using hierarchical reasoning alone, counterfactual reasoning alone, or combinations of the two reasoning strategies.

The researchers compared the subjects’ performance with the models’ predictions and found that for every subject, their performance was most closely associated with a model that used hierarchical reasoning but sometimes switched to counterfactual reasoning.

That suggests that instead of tracking all the possible paths that the ball could take, people broke up the task. First, they picked the direction (left or right), in which they thought the ball turned at the first junction, and continued to track the ball as it headed for the next turn. If the timing of the next sound they heard wasn’t compatible with the path they had chosen, they would go back and revise their first prediction — but only some of the time.

Switching back to the other side, which represents a shift to counterfactual reasoning, requires people to review their memory of the tones that they heard. However, it turns out that these memories are not always reliable, and the researchers found that people decided whether to go back or not based on how good they believed their memory to be.

“People rely on counterfactuals to the degree that it’s helpful,” Jazayeri says. “People who take a big performance loss when they do counterfactuals avoid doing them. But if you are someone who’s really good at retrieving information from the recent past, you may go back to the other side.”

Human limitations

To further validate their results, the researchers created a machine-learning neural network and trained it to complete the task. A machine-learning model trained on this task will track the ball’s path accurately and make the correct prediction every time, unless the researchers impose limitations on its performance.

When the researchers added cognitive limitations similar to those faced by humans, they found that the model altered its strategies. When they eliminated the model’s ability to follow all possible trajectories, it began to employ hierarchical and counterfactual strategies like humans do. If the researchers reduced the model’s memory recall ability, it began to switch to counterfactual only if it thought its recall would be good enough to get the right answer — just as humans do.

“What we found is that networks mimic human behavior when we impose on them those computational constraints that we found in human behavior,” Jazayeri says. “This is really saying that humans are acting rationally under the constraints that they have to function under.”

By slightly varying the amount of memory impairment programmed into the models, the researchers also saw hints that the switching of strategies appears to happen gradually, rather than at a distinct cut-off point. They are now performing further studies to try to determine what is happening in the brain as these shifts in strategy occur.

The research was funded by a Lisa K. Yang ICoN Fellowship, a Friends of the McGovern Institute Student Fellowship, a National Science Foundation Graduate Research Fellowship, the Simons Foundation, the Howard Hughes Medical Institute, and the McGovern Institute.

For many patients with schizophrenia, other psychiatric illnesses, or diseases such as hypertension and asthma, it can be difficult to take their medicine every day. To help overcome that challenge, MIT researchers have developed a pill that can be taken just once a week and gradually releases medication from within the stomach.

In a phase 3 clinical trial conducted by MIT spinout Lyndra Therapeutics, the researchers used the once-a-week pill to deliver a widely used medication for managing the symptoms of schizophrenia. They found that this treatment regimen maintained consistent levels of the drug in patients’ bodies and controlled their symptoms just as well as daily doses of the drug. The results are published today in Lancet Psychiatry.

“We’ve converted something that has to be taken once a day to once a week, orally, using a technology that can be adapted for a variety of medications,” says Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, an associate member of the Broad Institute, and an author of the study. “The ability to provide a sustained level of drug for a prolonged period, in an easy-to-administer system, makes it easier to ensure patients are receiving their medication.”

Traverso’s lab began developing the ingestible capsule studied in this trial more than 10 years ago, as part of an ongoing effort to make medications easier for patients to take. The capsule is about the size of a multivitamin, and once swallowed, it expands into a star shape that helps it remain in the stomach until all of the drug is released.

Richard Scranton, chief medical officer of Lyndra Therapeutics, is the senior author of the paper, and Leslie Citrome, a clinical professor of psychiatry and behavioral sciences at New York Medical College School of Medicine, is the lead author. Nayana Nagaraj, medical director at Lyndra Therapeutics, and Todd Dumas, senior director of pharmacometrics at Certara, are also authors.

Sustained delivery

Over the past decade, Traverso’s lab has been working on a variety of capsules that can be swallowed and remain in the digestive tract for days or weeks, slowly releasing their drug payload. In 2016, his team reported the star-shaped device, which was then further developed by Lyndra for clinical trials in patients with schizophrenia.

The device contains six arms that can be folded in, allowing it to fit inside a capsule. The capsule dissolves when the device reaches the stomach, allowing the arms to spring out. Once the arms are extended, the device becomes too large to pass through the pylorus (the exit of the stomach), so it remains freely floating in the stomach as drugs are slowly released from the arms. After about a week, the arms break off on their own, and each segment exits the stomach and passes through the digestive tract.

For the clinical trials, the capsule was loaded with risperidone, a commonly prescribed medication used to treat schizophrenia. Most patients take the drug orally once a day. There are also injectable versions that can be given every two weeks, every month, or every two months, but they require administration by a health care provider and are not always acceptable to patients.

The MIT and Lyndra team chose to focus on schizophrenia in hopes that a drug regimen that could be administered less frequently, through oral delivery, could make treatment easier for patients and their caregivers.

“One of the areas of unmet need that was recognized early on is neuropsychiatric conditions, where the illness can limit or impair one’s ability to remember to take their medication,” Traverso says. “With that in mind, one of the conditions that has been a big focus has been schizophrenia.”

The phase 3 trial was coordinated by researchers at Lyndra and enrolled 83 patients at five different sites around the United States. Forty-five of those patients completed the full five weeks of the study, in which they took one risperidone-loaded capsule per week.

Throughout the study, the researchers measured the amount of drug in each patient’s bloodstream. Each week, they found a sharp increase on the day the pill was given, followed by a slow decline over the next week. The levels were all within the optimal range, and there was less variation over time than is seen when patients take a pill each day.

Effective treatment

Using an evaluation known as the Positive and Negative Syndrome Scale (PANSS), the researchers also found that the patients’ symptoms remained stable throughout the study.

“One of the biggest obstacles in the care of people with chronic illnesses in general is that medications are not taken consistently. This leads to worsening symptoms, and in the case of schizophrenia, potential relapse and hospitalization,” Citrome says. “Having the option to take medication by mouth once a week represents an important option that can assist with adherence for the many patients who would prefer oral medications versus injectable formulations.”

Side effects from the treatment were minimal, the researchers found. Some patients experienced mild acid reflux and constipation early in the study, but these did not last long. The results, showing effectiveness of the capsule and few side effects, represent a major milestone in this approach to drug delivery, Traverso says.

“This really demonstrates that what we had hypothesized a decade ago, which is that a single capsule providing a drug depot within the GI tract could be possible,” he says. “Here what you see is that the capsule can achieve the drug levels that were predicted, and also control symptoms in a sizeable cohort of patients with schizophrenia.”

The investigators now hope to complete larger phase 3 studies before applying for FDA approval of this delivery approach for risperidone. They are also preparing for phase 1 trials using this capsule to deliver other drugs, including contraceptives.

“We are delighted that this technology which started at MIT has reached the point of phase 3 clinical trials,” says Robert Langer, the David H. Koch Institute Professor at MIT, who was an author of the original study on the star capsule and is a co-founder of Lyndra Therapeutics.

The research was funded by Lyndra Therapeutics.

Travel agents help to provide end-to-end logistics — like transportation, accommodations, meals, and lodging — for businesspeople, vacationers, and everyone in between. For those looking to make their own arrangements, large language models (LLMs) seem like they would be a strong tool to employ for this task because of their ability to iteratively interact using natural language, provide some commonsense reasoning, collect information, and call other tools in to help with the task at hand. However, recent work has found that state-of-the-art LLMs struggle with complex logistical and mathematical reasoning, as well as problems with multiple constraints, like trip planning, where they’ve been found to provide viable solutions 4 percent or less of the time, even with additional tools and application programming interfaces (APIs).

Subsequently, a research team from MIT and the MIT-IBM Watson AI Lab reframed the issue to see if they could increase the success rate of LLM solutions for complex problems. “We believe a lot of these planning problems are naturally a combinatorial optimization problem,” where you need to satisfy several constraints in a certifiable way, says Chuchu Fan, associate professor in the MIT Department of Aeronautics and Astronautics (AeroAstro) and the Laboratory for Information and Decision Systems (LIDS). She is also a researcher in the MIT-IBM Watson AI Lab. Her team applies machine learning, control theory, and formal methods to develop safe and verifiable control systems for robotics, autonomous systems, controllers, and human-machine interactions.

Noting the transferable nature of their work for travel planning, the group sought to create a user-friendly framework that can act as an AI travel broker to help develop realistic, logical, and complete travel plans. To achieve this, the researchers combined common LLMs with algorithms and a complete satisfiability solver. Solvers are mathematical tools that rigorously check if criteria can be met and how, but they require complex computer programming for use. This makes them natural companions to LLMs for problems like these, where users want help planning in a timely manner, without the need for programming knowledge or research into travel options. Further, if a user’s constraint cannot be met, the new technique can identify and articulate where the issue lies and propose alternative measures to the user, who can then choose to accept, reject, or modify them until a valid plan is formulated, if one exists.

“Different complexities of travel planning are something everyone will have to deal with at some point. There are different needs, requirements, constraints, and real-world information that you can collect,” says Fan. “Our idea is not to ask LLMs to propose a travel plan. Instead, an LLM here is acting as a translator to translate this natural language description of the problem into a problem that a solver can handle [and then provide that to the user],” says Fan.

Co-authoring a paper on the work with Fan are Yang Zhang of MIT-IBM Watson AI Lab, AeroAstro graduate student Yilun Hao, and graduate student Yongchao Chen of MIT LIDS and Harvard University. This work was recently presented at the Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics.

Breaking down the solver

Math tends to be domain-specific. For example, in natural language processing, LLMs perform regressions to predict the next token, a.k.a. “word,” in a series to analyze or create a document. This works well for generalizing diverse human inputs. LLMs alone, however, wouldn’t work for formal verification applications, like in aerospace or cybersecurity, where circuit connections and constraint tasks need to be complete and proven, otherwise loopholes and vulnerabilities can sneak by and cause critical safety issues. Here, solvers excel, but they need fixed formatting inputs and struggle with unsatisfiable queries.  A hybrid technique, however, provides an opportunity to develop solutions for complex problems, like trip planning, in a way that’s intuitive for everyday people.

“The solver is really the key here, because when we develop these algorithms, we know exactly how the problem is being solved as an optimization problem,” says Fan. Specifically, the research group used a solver called satisfiability modulo theories (SMT), which determines whether a formula can be satisfied. “With this particular solver, it’s not just doing optimization. It’s doing reasoning over a lot of different algorithms there to understand whether the planning problem is possible or not to solve. That’s a pretty significant thing in travel planning. It’s not a very traditional mathematical optimization problem because people come up with all these limitations, constraints, restrictions,” notes Fan.

Translation in action

The “travel agent” works in four steps that can be repeated, as needed. The researchers used GPT-4, Claude-3, or Mistral-Large as the method’s LLM. First, the LLM parses a user’s requested travel plan prompt into planning steps, noting preferences for budget, hotels, transportation, destinations, attractions, restaurants, and trip duration in days, as well as any other user prescriptions. Those steps are then converted into executable Python code (with a natural language annotation for each of the constraints), which calls APIs like CitySearch, FlightSearch, etc. to collect data, and the SMT solver to begin executing the steps laid out in the constraint satisfaction problem. If a sound and complete solution can be found, the solver outputs the result to the LLM, which then provides a coherent itinerary to the user.

If one or more constraints cannot be met, the framework begins looking for an alternative. The solver outputs code identifying the conflicting constraints (with its corresponding annotation) that the LLM then provides to the user with a potential remedy. The user can then decide how to proceed, until a solution (or the maximum number of iterations) is reached.

Generalizable and robust planning

The researchers tested their method using the aforementioned LLMs against other baselines: GPT-4 by itself, OpenAI o1-preview by itself, GPT-4 with a tool to collect information, and a search algorithm that optimizes for total cost. Using the TravelPlanner dataset, which includes data for viable plans, the team looked at multiple performance metrics: how frequently a method could deliver a solution, if the solution satisfied commonsense criteria like not visiting two cities in one day, the method’s ability to meet one or more constraints, and a final pass rate indicating that it could meet all constraints. The new technique generally achieved over a 90 percent pass rate, compared to 10 percent or lower for the baselines. The team also explored the addition of a JSON representation within the query step, which further made it easier for the method to provide solutions with 84.4-98.9 percent pass rates.

The MIT-IBM team posed additional challenges for their method. They looked at how important each component of their solution was — such as removing human feedback or the solver — and how that affected plan adjustments to unsatisfiable queries within 10 or 20 iterations using a new dataset they created called UnsatChristmas, which includes unseen constraints, and a modified version of TravelPlanner. On average, the MIT-IBM group’s framework achieved 78.6  and 85 percent success, which rises to 81.6 and 91.7 percent with additional plan modification rounds. The researchers analyzed how well it handled new, unseen constraints and paraphrased query-step and step-code prompts. In both cases, it performed very well, especially with an 86.7 percent pass rate for the paraphrasing trial.

Lastly, the MIT-IBM researchers applied their framework to other domains with tasks like block picking, task allocation, the traveling salesman problem, and warehouse. Here, the method must select numbered, colored blocks and maximize its score; optimize robot task assignment for different scenarios; plan trips minimizing distance traveled; and robot task completion and optimization.

“I think this is a very strong and innovative framework that can save a lot of time for humans, and also, it’s a very novel combination of the LLM and the solver,” says Hao.

This work was funded, in part, by the Office of Naval Research and the MIT-IBM Watson AI Lab.

Suppose you were shown that an artificial intelligence tool offers accurate predictions about some stocks you own. How would you feel about using it? Now, suppose you are applying for a job at a company where the HR department uses an AI system to screen resumes. Would you be comfortable with that?

A new study finds that people are neither entirely enthusiastic nor totally averse to AI. Rather than falling into camps of techno-optimists and Luddites, people are discerning about the practical upshot of using AI, case by case.

“We propose that AI appreciation occurs when AI is perceived as being more capable than humans and personalization is perceived as being unnecessary in a given decision context,” says MIT Professor Jackson Lu, co-author of a newly published paper detailing the study’s results. “AI aversion occurs when either of these conditions is not met, and AI appreciation occurs only when both conditions are satisfied.”

The paper, “AI Aversion or Appreciation? A Capability–Personalization Framework and a Meta-Analytic Review,” appears in Psychological Bulletin. The paper has eight co-authors, including Lu, who is the Career Development Associate Professor of Work and Organization Studies at the MIT Sloan School of Management.

New framework adds insight

People’s reactions to AI have long been subject to extensive debate, often producing seemingly disparate findings. An influential 2015 paper on “algorithm aversion” found that people are less forgiving of AI-generated errors than of human errors, whereas a widely noted 2019 paper on “algorithm appreciation” found that people preferred advice from AI, compared to advice from humans.

To reconcile these mixed findings, Lu and his co-authors conducted a meta-analysis of 163 prior studies that compared people’s preferences for AI versus humans. The researchers tested whether the data supported their proposed “Capability–Personalization Framework” — the idea that in a given context, both the perceived capability of AI and the perceived necessity for personalization shape our preferences for either AI or humans.

Across the 163 studies, the research team analyzed over 82,000 reactions to 93 distinct “decision contexts” — for instance, whether or not participants would feel comfortable with AI being used in cancer diagnoses. The analysis confirmed that the Capability–Personalization Framework indeed helps account for people’s preferences.

“The meta-analysis supported our theoretical framework,” Lu says. “Both dimensions are important: Individuals evaluate whether or not AI is more capable than people at a given task, and whether the task calls for personalization. People will prefer AI only if they think the AI is more capable than humans and the task is nonpersonal.”

He adds: “The key idea here is that high perceived capability alone does not guarantee AI appreciation. Personalization matters too.”

For example, people tend to favor AI when it comes to detecting fraud or sorting large datasets — areas where AI’s abilities exceed those of humans in speed and scale, and personalization is not required. But they are more resistant to AI in contexts like therapy, job interviews, or medical diagnoses, where they feel a human is better able to recognize their unique circumstances.

“People have a fundamental desire to see themselves as unique and distinct from other people,” Lu says. “AI is often viewed as impersonal and operating in a rote manner. Even if the AI is trained on a wealth of data, people feel AI can’t grasp their personal situations. They want a human recruiter, a human doctor who can see them as distinct from other people.”

Context also matters: From tangibility to unemployment

The study also uncovered other factors that influence individuals’ preferences for AI. For instance, AI appreciation is more pronounced for tangible robots than for intangible algorithms.

Economic context also matters. In countries with lower unemployment, AI appreciation is more pronounced.

“It makes intuitive sense,” Lu says. “If you worry about being replaced by AI, you’re less likely to embrace it.”  

Lu is continuing to examine people’s complex and evolving attitudes toward AI. While he does not view the current meta-analysis as the last word on the matter, he hopes the Capability–Personalization Framework offers a valuable lens for understanding how people evaluate AI across different contexts.

“We’re not claiming perceived capability and personalization are the only two dimensions that matter, but according to our meta-analysis, these two dimensions capture much of what shapes people’s preferences for AI versus humans across a wide range of studies,” Lu concludes.

In addition to Lu, the paper’s co-authors are Xin Qin, Chen Chen, Hansen Zhou, Xiaowei Dong, and Limei Cao of Sun Yat-sen University; Xiang Zhou of Shenzhen University; and Dongyuan Wu of Fudan University.

The research was supported, in part, by grants to Qin and Wu from the National Natural Science Foundation of China. 

MIT has an unparalleled history of bringing together interdisciplinary teams to solve pressing problems — think of the development of radar during World War II, or leading the international coalition that cracked the code of the human genome — but the challenge of climate change could demand a scale of collaboration unlike any that’s come before at MIT.

“Solving climate change is not just about new technologies or better models. It’s about forging new partnerships across campus and beyond — between scientists and economists, between architects and data scientists, between policymakers and physicists, between anthropologists and engineers, and more,” MIT Vice President for Energy and Climate Evelyn Wang told an energetic crowd of faculty, students, and staff on May 6. “Each of us holds a piece of the solution — but only together can we see the whole.”

Undeterred by heavy rain, approximately 300 campus community members filled the atrium in the Tina and Hamid Moghadam Building (Building 55) for a spring gathering hosted by Wang and the Climate Project at MIT. The initiative seeks to direct the full strength of MIT to address climate change, which Wang described as one of the defining challenges of this moment in history — and one of its greatest opportunities.

“It calls on us to rethink how we power our world, how we build, how we live — and how we work together,” Wang said. “And there is no better place than MIT to lead this kind of bold, integrated effort. Our culture of curiosity, rigor, and relentless experimentation makes us uniquely suited to cross boundaries — to break down silos and build something new.”

The Climate Project is organized around six missions, thematic areas in which MIT aims to make significant impact, ranging from decarbonizing industry to new policy approaches to designing resilient cities. The faculty leaders of these missions posed challenges to the crowd before circulating among the crowd to share their perspectives and to discuss community questions and ideas.

Wang and the Climate Project team were joined by a number of research groups, startups, and MIT offices conducting relevant work today on issues related to energy and climate. For example, the MIT Office of Sustainability showcased efforts to use the MIT campus as a living laboratory; MIT spinouts such as Forma Systems, which is developing high-performance, low-carbon building systems, and Addis Energy, which envisions using the earth as a reactor to produce clean ammonia, presented their technologies; and visitors learned about current projects in MIT labs, including DebunkBot, an artificial intelligence-powered chatbot that can persuade people to shift their attitudes about conspiracies, developed by David Rand, the Erwin H. Schell Professor at the MIT Sloan School of Management.

Benedetto Marelli, an associate professor in the Department of Civil and Environmental Engineering who leads the Wild Cards Mission, said the energy and enthusiasm that filled the room was inspiring — but that the individual conversations were equally valuable.

“I was especially pleased to see so many students come out. I also spoke with other faculty, talked to staff from across the Institute, and met representatives of external companies interested in collaborating with MIT,” Marelli said. “You could see connections being made all around the room, which is exactly what we need as we build momentum for the Climate Project.”

Researchers from the Disruptive and Sustainable Technologies for Agricultural Precision (DiSTAP) interdisciplinary research group within the Singapore-MIT Alliance for Research and Technology have developed the world’s first near-infrared fluorescent nanosensor capable of real-time, nondestructive, and species-agnostic detection of indole-3-acetic acid (IAA) — the primary bioactive auxin hormone that controls the way plants develop, grow, and respond to stress.

Auxins, particularly IAA, play a central role in regulating key plant processes such as cell division, elongation, root and shoot development, and response to environmental cues like light, heat, and drought. External factors like light affect how auxin moves within the plant, temperature influences how much is produced, and a lack of water can disrupt hormone balance. When plants cannot effectively regulate auxins, they may not grow well, adapt to changing conditions, or produce as much food. 

Existing IAA detection methods, such as liquid chromatography, require taking plant samples from the plant — which harms or removes part of it. Conventional methods also measure the effects of IAA rather than detecting it directly, and cannot be used universally across different plant types. In addition, since IAA are small molecules that cannot be easily tracked in real time, biosensors that contain fluorescent proteins need to be inserted into the plant’s genome to measure auxin, making it emit a fluorescent signal for live imaging.

SMART’s newly developed nanosensor enables direct, real-time tracking of auxin levels in living plants with high precision. The sensor uses near infrared imaging to monitor IAA fluctuations non-invasively across tissues like leaves, roots, and cotyledons, and it is capable of bypassing chlorophyll interference to ensure highly reliable readings even in densely pigmented tissues. The technology does not require genetic modification and can be integrated with existing agricultural systems — offering a scalable precision tool to advance both crop optimization and fundamental plant physiology research. 

By providing real-time, precise measurements of auxin, the sensor empowers farmers with earlier and more accurate insights into plant health. With these insights and comprehensive data, farmers can make smarter, data-driven decisions on irrigation, nutrient delivery, and pruning, tailored to the plant’s actual needs — ultimately improving crop growth, boosting stress resilience, and increasing yields.

“We need new technologies to address the problems of food insecurity and climate change worldwide. Auxin is a central growth signal within living plants, and this work gives us a way to tap it to give new information to farmers and researchers,” says Michael Strano, co-lead principal investigator at DiSTAP, Carbon P. Dubbs Professor of Chemical Engineering at MIT, and co-corresponding author of the paper. “The applications are many, including early detection of plant stress, allowing for timely interventions to safeguard crops. For urban and indoor farms, where light, water, and nutrients are already tightly controlled, this sensor can be a valuable tool in fine-tuning growth conditions with even greater precision to optimize yield and sustainability.”

The research team documented the nanosensor’s development in a paper titled, “A Near-Infrared Fluorescent Nanosensor for Direct and Real-Time Measurement of Indole-3-Acetic Acid in Plants,” published in the journal ACS Nano. The sensor comprises single-walled carbon nanotubes wrapped in a specially designed polymer, which enables it to detect IAA through changes in near infrared fluorescence intensity. Successfully tested across multiple species, including Arabidopsis, Nicotiana benthamiana, choy sum, and spinach, the nanosensor can map IAA responses under various environmental conditions such as shade, low light, and heat stress. 

“This sensor builds on DiSTAP’s ongoing work in nanotechnology and the CoPhMoRe technique, which has already been used to develop other sensors that can detect important plant compounds such as gibberellins and hydrogen peroxide. By adapting this approach for IAA, we’re adding to our inventory of novel, precise, and nondestructive tools for monitoring plant health. Eventually, these sensors can be multiplexed, or combined, to monitor a spectrum of plant growth markers for more complete insights into plant physiology,” says Duc Thinh Khong, research scientist at DiSTAP and co-first author of the paper.

“This small but mighty nanosensor tackles a long-standing challenge in agriculture: the need for a universal, real-time, and noninvasive tool to monitor plant health across various species. Our collaborative achievement not only empowers researchers and farmers to optimize growth conditions and improve crop yield and resilience, but also advances our scientific understanding of hormone pathways and plant-environment interactions,” says In-Cheol Jang, senior principal investigator at TLL, principal investigator at DiSTAP, and co-corresponding author of the paper.

Looking ahead, the research team is looking to combine multiple sensing platforms to simultaneously detect IAA and its related metabolites to create a comprehensive hormone signaling profile, offering deeper insights into plant stress responses and enhancing precision agriculture. They are also working on using microneedles for highly localized, tissue-specific sensing, and collaborating with industrial urban farming partners to translate the technology into practical, field-ready solutions. 

The research was carried out by SMART, and supported by the National Research Foundation of Singapore under its Campus for Research Excellence And Technological Enterprise program. The universal nanosensor was developed in collaboration with Temasek Life Sciences Laboratory (TLL) and MIT.

An autonomous drone carrying water to help extinguish a wildfire in the Sierra Nevada might encounter swirling Santa Ana winds that threaten to push it off course. Rapidly adapting to these unknown disturbances inflight presents an enormous challenge for the drone’s flight control system.

To help such a drone stay on target, MIT researchers developed a new, machine learning-based adaptive control algorithm that could minimize its deviation from its intended trajectory in the face of unpredictable forces like gusty winds.

Unlike standard approaches, the new technique does not require the person programming the autonomous drone to know anything in advance about the structure of these uncertain disturbances. Instead, the control system’s artificial intelligence model learns all it needs to know from a small amount of observational data collected from 15 minutes of flight time.

Importantly, the technique automatically determines which optimization algorithm it should use to adapt to the disturbances, which improves tracking performance. It chooses the algorithm that best suits the geometry of specific disturbances this drone is facing.

The researchers train their control system to do both things simultaneously using a technique called meta-learning, which teaches the system how to adapt to different types of disturbances.

Taken together, these ingredients enable their adaptive control system to achieve 50 percent less trajectory tracking error than baseline methods in simulations and perform better with new wind speeds it didn’t see during training.

In the future, this adaptive control system could help autonomous drones more efficiently deliver heavy parcels despite strong winds or monitor fire-prone areas of a national park.

“The concurrent learning of these components is what gives our method its strength. By leveraging meta-learning, our controller can automatically make choices that will be best for quick adaptation,” says Navid Azizan, who is the Esther and Harold E. Edgerton Assistant Professor in the MIT Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), a principal investigator of the Laboratory for Information and Decision Systems (LIDS), and the senior author of a paper on this control system.

Azizan is joined on the paper by lead author Sunbochen Tang, a graduate student in the Department of Aeronautics and Astronautics, and Haoyuan Sun, a graduate student in the Department of Electrical Engineering and Computer Science. The research was recently presented at the Learning for Dynamics and Control Conference.

Finding the right algorithm

Typically, a control system incorporates a function that models the drone and its environment, and includes some existing information on the structure of potential disturbances. But in a real world filled with uncertain conditions, it is often impossible to hand-design this structure in advance.

Many control systems use an adaptation method based on a popular optimization algorithm, known as gradient descent, to estimate the unknown parts of the problem and determine how to keep the drone as close as possible to its target trajectory during flight. However, gradient descent is only one algorithm in a larger family of algorithms available to choose, known as mirror descent.

“Mirror descent is a general family of algorithms, and for any given problem, one of these algorithms can be more suitable than others. The name of the game is how to choose the particular algorithm that is right for your problem. In our method, we automate this choice,” Azizan says.

In their control system, the researchers replaced the function that contains some structure of potential disturbances with a neural network model that learns to approximate them from data. In this way, they don’t need to have an a priori structure of the wind speeds this drone could encounter in advance.

Their method also uses an algorithm to automatically select the right mirror-descent function while learning the neural network model from data, rather than assuming a user has the ideal function picked out already. The researchers give this algorithm a range of functions to pick from, and it finds the one that best fits the problem at hand.

“Choosing a good distance-generating function to construct the right mirror-descent adaptation matters a lot in getting the right algorithm to reduce the tracking error,” Tang adds.

Learning to adapt

While the wind speeds the drone may encounter could change every time it takes flight, the controller’s neural network and mirror function should stay the same so they don’t need to be recomputed each time.

To make their controller more flexible, the researchers use meta-learning, teaching it to adapt by showing it a range of wind speed families during training.

“Our method can cope with different objectives because, using meta-learning, we can learn a shared representation through different scenarios efficiently from data,” Tang explains.

In the end, the user feeds the control system a target trajectory and it continuously recalculates, in real-time, how the drone should produce thrust to keep it as close as possible to that trajectory while accommodating the uncertain disturbance it encounters.

In both simulations and real-world experiments, the researchers showed that their method led to significantly less trajectory tracking error than baseline approaches with every wind speed they tested.

“Even if the wind disturbances are much stronger than we had seen during training, our technique shows that it can still handle them successfully,” Azizan adds.

In addition, the margin by which their method outperformed the baselines grew as the wind speeds intensified, showing that it can adapt to challenging environments.

The team is now performing hardware experiments to test their control system on real drones with varying wind conditions and other disturbances.

They also want to extend their method so it can handle disturbances from multiple sources at once. For instance, changing wind speeds could cause the weight of a parcel the drone is carrying to shift in flight, especially when the drone is carrying sloshing payloads.

They also want to explore continual learning, so the drone could adapt to new disturbances without the need to also be retrained on the data it has seen so far.

“Navid and his collaborators have developed breakthrough work that combines meta-learning with conventional adaptive control to learn nonlinear features and the suitable adaptation law from data. Key to their approach is the use of mirror descent techniques that exploit the underlying geometry of the problem and do so automatically. Their work can contribute significantly to the design of autonomous systems that need to operate in complex and uncertain environments,” says Babak Hassibi, the Mose and Lillian S. Bohn Professor of Electrical Engineering and Computing and Mathematical Sciences at Caltech, who was not involved with this work.

This research was supported, in part, by MathWorks, the MIT-IBM Watson AI Lab, the MIT-Amazon Science Hub, and the MIT-Google Program for Computing Innovation.

Fusion energy has the potential to enable the energy transition from fossil fuels, enhance domestic energy security, and power artificial intelligence. Private companies have already invested more than $8 billion to develop commercial fusion and seize the opportunities it offers. An urgent challenge, however, is the discovery and evaluation of cost-effective materials that can withstand extreme conditions for extended periods, including 150-million-degree plasmas and intense particle bombardment.

To meet this challenge, MIT’s Plasma Science and Fusion Center (PSFC) has launched the Schmidt Laboratory for Materials in Nuclear Technologies, or LMNT (pronounced “element”). Backed by a philanthropic consortium led by Eric and Wendy Schmidt, LMNT is designed to speed up the discovery and selection of materials for a variety of fusion power plant components. 

By drawing on MIT's expertise in fusion and materials science, repurposing existing research infrastructure, and tapping into its close collaborations with leading private fusion companies, the PSFC aims to drive rapid progress in the materials that are necessary for commercializing fusion energy on rapid timescales. LMNT will also help develop and assess materials for nuclear power plants, next-generation particle physics experiments, and other science and industry applications.

Zachary Hartwig, head of LMNT and an associate professor in the Department of Nuclear Science and Engineering (NSE), says, “We need technologies today that will rapidly develop and test materials to support the commercialization of fusion energy. LMNT’s mission includes discovery science but seeks to go further, ultimately helping select the materials that will be used to build fusion power plants in the coming years.”

A different approach to fusion materials

For decades, researchers have worked to understand how materials behave under fusion conditions using methods like exposing test specimens to low-energy particle beams, or placing them in the core of nuclear fission reactors. These approaches, however, have significant limitations. Low-energy particle beams only irradiate the thinnest surface layer of materials, while fission reactor irradiation doesn’t accurately replicate the mechanism by which fusion damages materials. Fission irradiation is also an expensive, multiyear process that requires specialized facilities.

To overcome these obstacles, researchers at MIT and peer institutions are exploring the use of energetic beams of protons to simulate the damage materials undergo in fusion environments. Proton beams can be tuned to match the damage expected in fusion power plants, and protons penetrate deep enough into test samples to provide insights into how exposure can affect structural integrity. They also offer the advantage of speed: first, intense proton beams can rapidly damage dozens of material samples at once, allowing researchers to test them in days, rather than years. Second, high-energy proton beams can be generated with a type of particle accelerator known as a cyclotron commonly used in the health-care industry. As a result, LMNT will be built around a cost-effective, off-the-shelf cyclotron that is easy to obtain and highly reliable.

LMNT will surround its cyclotron with four experimental areas dedicated to materials science research. The lab is taking shape inside the large shielded concrete vault at PSFC that once housed the Alcator C-Mod tokamak, a record-setting fusion experiment that ran at the PSFC from 1992 to 2016. By repurposing C-Mod’s former space, the center is skipping the need for extensive, costly new construction and accelerating the research timeline significantly. The PSFC’s veteran team — who have led major projects like the Alcator tokamaks and advanced high-temperature superconducting magnet development — are overseeing the facilities design, construction, and operation, ensuring LMNT moves quickly from concept to reality. The PSFC expects to receive the cyclotron by the end of 2025, with experimental operations starting in early 2026.

“LMNT is the start of a new era of fusion research at MIT, one where we seek to tackle the most complex fusion technology challenges on timescales commensurate with the urgency of the problem we face: the energy transition,” says Nuno Loureiro, director of the PSFC, a professor of nuclear science and engineering, and the Herman Feshbach Professor of Physics. “It’s ambitious, bold, and critical — and that’s exactly why we do it.”

“What’s exciting about this project is that it aligns the resources we have today — substantial research infrastructure, off-the-shelf technologies, and MIT expertise — to address the key resource we lack in tackling climate change: time. Using the Schmidt Laboratory for Materials in Nuclear Technologies, MIT researchers advancing fusion energy, nuclear power, and other technologies critical to the future of energy will be able to act now and move fast,” says Elsa Olivetti, the Jerry McAfee Professor in Engineering and a mission director of MIT’s Climate Project.

In addition to advancing research, LMNT will provide a platform for educating and training students in the increasingly important areas of fusion technology. LMNT’s location on MIT’s main campus gives students the opportunity to lead research projects and help manage facility operations. It also continues the hands-on approach to education that has defined the PSFC, reinforcing that direct experience in large-scale research is the best approach to create fusion scientists and engineers for the expanding fusion industry workforce.

Benoit Forget, head of NSE and the Korea Electric Power Professor of Nuclear Engineering, notes, “This new laboratory will give nuclear science and engineering students access to a unique research capability that will help shape the future of both fusion and fission energy.”

Accelerating progress on big challenges

Philanthropic support has helped LMNT leverage existing infrastructure and expertise to move from concept to facility in just one-and-a-half years — a fast timeline for establishing a major research project.

“I’m just as excited about this research model as I am about the materials science. It shows how focused philanthropy and MIT’s strengths can come together to build something that’s transformational — a major new facility that helps researchers from the public and private sectors move fast on fusion materials,” emphasizes Hartwig.

By utilizing this approach, the PSFC is executing a major public-private partnership in fusion energy, realizing a research model that the U.S. fusion community has only recently started to explore, and demonstrating the crucial role that universities can play in the acceleration of the materials and technology required for fusion energy.

“Universities have long been at the forefront of tackling society’s biggest challenges, and the race to identify new forms of energy and address climate change demands bold, high-risk, high-reward approaches,” says Ian Waitz, MIT’s vice president for research. “LMNT is helping turn fusion energy from a long-term ambition into a near-term reality.”

When navigating a place that we’re only somewhat familiar with, we often rely on unique landmarks to help make our way. However, if we’re looking for an office in a brick building, and there are many brick buildings along our route, we might use a rule like looking for the second building on a street, rather than relying on distinguishing the building itself.

Until that ambiguity is resolved, we must hold in mind that there are multiple possibilities (or hypotheses) for where we are in relation to our destination. In a study of mice, MIT neuroscientists have now discovered that these hypotheses are explicitly represented in the brain by distinct neural activity patterns.

This is the first time that neural activity patterns that encode simultaneous hypotheses have been seen in the brain. The researchers found that these representations, which were observed in the brain’s retrosplenial cortex (RSC), not only encode hypotheses but also could be used by the animals to choose the correct way to go.

“As far as we know, no one has shown in a complex reasoning task that there’s an area in association cortex that holds two hypotheses in mind and then uses one of those hypotheses, once it gets more information, to actually complete the task,” says Mark Harnett, an associate professor of brain and cognitive sciences, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the study.

Jakob Voigts PhD ’17, a former postdoc in Harnett’s lab and now a group leader at the Howard Hughes Medical Institute Janelia Research Campus, is the lead author of the paper, which appears today in Nature Neuroscience.

Ambiguous landmarks

The RSC receives input from the visual cortex, the hippocampal formation, and the anterior thalamus, which it integrates to help guide navigation.

In a 2020 paper, Harnett’s lab found that the RSC uses both visual and spatial information to encode landmarks used for navigation. In that study, the researchers showed that neurons in the RSC of mice integrate visual information about the surrounding environment with spatial feedback of the mice’s own position along a track, allowing them to learn where to find a reward based on landmarks that they saw.

In their new study, the researchers wanted to delve further into how the RSC uses spatial information and situational context to guide navigational decision-making. To do that, the researchers devised a much more complicated navigational task than typically used in mouse studies. They set up a large, round arena, with 16 small openings, or ports, along the side walls. One of these openings would give the mice a reward when they stuck their nose through it. In the first set of experiments, the researchers trained the mice to go to different reward ports indicated by dots of light on the floor that were only visible when the mice get close to them.

Once the mice learned to perform this relatively simple task, the researchers added a second dot. The two dots were always the same distance from each other and from the center of the arena. But now the mice had to go to the port by the counterclockwise dot to get the reward. Because the dots were identical and only became visible at close distances, the mice could never see both dots at once and could not immediately determine which dot was which.

To solve this task, mice therefore had to remember where they expected a dot to show up, integrating their own body position, the direction they were heading, and path they took to figure out which landmark is which. By measuring RSC activity as the mice approached the ambiguous landmarks, the researchers could determine whether the RSC encodes hypotheses about spatial location. The task was carefully designed to require the mice to use the visual landmarks to obtain rewards, instead of other strategies like odor cues or dead reckoning.

“What is important about the behavior in this case is that mice need to remember something and then use that to interpret future input,” says Voigts, who worked on this study while a postdoc in Harnett’s lab. “It’s not just remembering something, but remembering it in such a way that you can act on it.”

The researchers found that as the mice accumulated information about which dot might be which, populations of RSC neurons displayed distinct activity patterns for incomplete information. Each of these patterns appears to correspond to a hypothesis about where the mouse thought it was with respect to the reward.

When the mice get close enough to figure out which dot was indicating the reward port, these patterns collapsed into the one that represents the correct hypothesis. The findings suggest that these patterns not only passively store hypotheses, they can also be used to compute how to get to the correct location, the researchers say.

“We show that RSC has the required information for using this short-term memory to distinguish the ambiguous landmarks. And we show that this type of hypothesis is encoded and processed in a way that allows the RSC to use it to solve the computation,” Voigts says.

Interconnected neurons

When analyzing their initial results, Harnett and Voigts consulted with MIT Professor Ila Fiete, who had run a study about 10 years ago using an artificial neural network to perform a similar navigation task.

That study, previously published on bioRxiv, showed that the neural network displayed activity patterns that were conceptually similar to those seen in the animal studies run by Harnett’s lab. The neurons of the artificial neural network ended up forming highly interconnected low-dimensional networks, like the neurons of the RSC.

“That interconnectivity seems, in ways that we still don’t understand, to be key to how these dynamics emerge and how they’re controlled. And it’s a key feature of how the RSC holds these two hypotheses in mind at the same time,” Harnett says.

In his lab at Janelia, Voigts now plans to investigate how other brain areas involved in navigation, such as the prefrontal cortex, are engaged as mice explore and forage in a more naturalistic way, without being trained on a specific task.

“We’re looking into whether there are general principles by which tasks are learned,” Voigts says. “We have a lot of knowledge in neuroscience about how brains operate once the animal has learned a task, but in comparison we know extremely little about how mice learn tasks or what they choose to learn when given freedom to behave naturally.”

The research was funded, in part, by the National Institutes of Health, a Simons Center for the Social Brain at MIT postdoctoral fellowship, the National Institute of General Medical Sciences, and the Center for Brains, Minds, and Machines at MIT, funded by the National Science Foundation.

Animators could create more realistic bouncy, stretchy, and squishy characters for movies and video games thanks to a new simulation method developed by researchers at MIT.

Their approach allows animators to simulate rubbery and elastic materials in a way that preserves the physical properties of the material and avoids pitfalls like instability.

The technique simulates elastic objects for animation and other applications, with improved reliability compared to other methods. In comparison, many existing simulation techniques can produce elastic animations that become erratic or sluggish or can even break down entirely.

To achieve this improvement, the MIT researchers uncovered a hidden mathematical structure in equations that capture how elastic materials deform on a computer. By leveraging this property, known as convexity, they designed a method that consistently produces accurate, physically faithful simulations.

“The way animations look often depends on how accurately we simulate the physics of the problem,” says Leticia Mattos Da Silva, an MIT graduate student and lead author of a paper on this research. “Our method aims to stay true to physical laws while giving more control and stability to animation artists.”

Beyond 3D animation, the researchers also see potential future uses in the design of real elastic objects, such as flexible shoes, garments, or toys. The method could be extended to help engineers explore how stretchy objects will perform before they are built.

She is joined on the paper by Silvia Sellán, an assistant professor of computer science at Columbia University; Natalia Pacheco-Tallaj, an MIT graduate student; and senior author Justin Solomon, an associate professor in the MIT Department of Electrical Engineering and Computer Science and leader of the Geometric Data Processing Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the SIGGRAPH conference.

Truthful to physics

If you drop a rubber ball on a wooden floor, it bounces back up. Viewers expect to see the same behavior in an animated world, but recreating such dynamics convincingly can be difficult. Many existing techniques simulate elastic objects using fast solvers that trade physical realism for speed, which can result in excessive energy loss or even simulation failure.

More accurate approaches, including a class of techniques called variational integrators, preserve the physical properties of the object, such as its total energy or momentum, and, in this way, mimic real-world behavior more closely. But these methods are often unreliable because they depend on complex equations that are hard to solve efficiently.

The MIT researchers tackled this problem by rewriting the equations of variational integrators to reveal a hidden convex structure. They broke the deformation of elastic materials into a stretch component and a rotation component, and found that the stretch portion forms a convex problem that is well-suited for stable optimization algorithms.

“If you just look at the original formulation, it seems fully non-convex. But because we can rewrite it so that is convex in at least some of its variables, we can inherit some advantages of convex optimization algorithms,” she says.

These convex optimization algorithms, when applied under the right conditions, come with guarantees of convergence, meaning they are more likely to find the correct answer to the problem. This generates more stable simulations over time, avoiding issues like a bouncing rubber ball losing too much energy or exploding mid-animation.

One of the biggest challenges the researchers faced was reinterpreting the formulation so they could extract that hidden convexity. Some other works explored hidden convexity in static problems, but it was not clear whether the structures remained solid for dynamic problems like simulating elastic objects in motion, Mattos Da Silva says.

Stability and efficiency

In experiments, their solver was able to simulate a wide range of elastic behavior, from bouncing shapes to squishy characters, with preservation of important physical properties and stability over long periods of time. Other simulation methods quickly ran into trouble: Some became unstable, causing erratic behavior, while others showed visible damping.

“Because our method demonstrates more stability, it can give animators more reliability and confidence when simulating anything elastic, whether it’s something from the real world or even something completely imaginary,” she says.

While the solver is not as fast as some simulation tools that prioritize speed over accuracy, it avoids many of the trade-offs those methods make. Compared to other physics-based approaches, it also avoids the need for complex, nonlinear solvers that can be sensitive and prone to failure.

In the future, the researchers want to explore techniques to further reduce computational cost. In addition, they want to explore applications of this technique in fabrication and engineering, where reliable simulations of elastic materials could support the design of real-world objects, like garments and toys.

“We were able to revive an old class of integrators in our work. My guess is there are other examples where researchers can revisit a problem to find a hidden convexity structure that could offer a lot of advantages,” she says.

This research is funded, in part, by a MathWorks Engineering Fellowship, the Army Research Office, the National Science Foundation, the CSAIL Future of Data Program, the MIT-IBM Watson AI Laboratory, the Wistron Corporation, and the Toyota-CSAIL Joint Research Center.

Academic research groups and startups are essential drivers of scientific progress. But some projects, like the Hubble Space Telescope or the Human Genome Project, are too big for any one academic lab or loose consortium. They’re also not immediately profitable enough for industry to take on.

That’s the gap researchers at MIT were trying to fill when they created the concept of focused research organizations, or FROs. They describe a FRO as a new type of entity, often philanthropically funded, that undertakes large research efforts using tightly coordinated teams to create a public good that accelerates scientific progress.

The original idea for focused research organizations came out of talks among researchers, most of whom were working to map the brain in MIT Professor Ed Boyden’s lab. After they began publishing their ideas, however, the researchers realized FROs could be a powerful tool to unlock scientific advances across many other applications.

“We were quite pleasantly surprised by the range of fields where we see FRO-shaped problems,” says Adam Marblestone, a former MIT research scientist who co-founded the nonprofit Convergent Research to help launch FROs in 2021. “Convergent has FRO proposals from climate, materials science, chemistry, biology — we even have launched a FRO on software for math. You wouldn’t expect math to be something with a large-scale technological research bottleneck, but it turns out even there, we found a software engineering bottleneck that needed to be solved.”

Marblestone helped formulate the idea for focused research organizations at MIT with a group including Andrew Payne SM ’17, PhD ’21 and Sam Rodriques PhD ’19, who were PhD students in Boyden’s lab at the time. Since then, the FRO concept has caught on. Convergent has helped attract philanthropic funding for FROs working to decode the immune system, identify the unintended targets of approved drugs, and understand the impacts of carbon dioxide removal in our oceans.

In total, Convergent has supported the creation of 10 FROs since its founding in 2021. Many of those groups have already released important tools for better understanding our world — and their leaders believe the best is yet to come.

“We’re starting to see these first open-source tools released in important areas,” Marblestone says. “We’re seeing the first concrete evidence that FROs are effective, because no other entity could have released these tools, and I think 2025 is going to be a significant year in terms of our newer FROs putting out new datasets and tools.”

A new model

Marblestone joined Boyden’s lab in 2014 as a research scientist after completing his PhD at Harvard University. He also worked in a new position called director of scientific architecting at the MIT Media Lab, which Boyden helped create, through which he tried to organize individual research efforts into larger projects. His own research focused on overcoming the challenges of measuring brain activity across large scales.

Marblestone discussed this and other large-scale neuroscience problems with Payne and Rodriques, and the researchers began thinking about gaps in scientific funding more broadly.

“The combination of myself, Sam, Andrew, Ed, and others’ experiences trying to start various large brain-mapping projects convinced us of the gap in support for medium-sized science and engineering teams with startup-inspired structures, built for the nonprofit purpose of building scientific infrastructure,” Marblestone says.

Through MIT, the researchers also connected with Tom Kalil, who was at the time chief innovation officer at Schmidt Futures, a philanthropic initiative of Eric and Wendy Schmidt. Rodriques wrote about the concept of a focused research organization as the last chapter of his PhD thesis in 2019.

“Ed always encouraged us to dream very, very big,” Rodriques says. “We were always trying to think about the hardest problems in biology and how to tackle them. My thesis basically ended with me explaining why we needed a new structure that is like a company, but nonprofit and dedicated to science.”

As part of a fellowship with the Federation of American Scientists in 2020, and working with Kalil, Marblestone interviewed scientists in dozens of fields outside of neuroscience and learned that the funding gap existed across disciplines.

When Rodriques and Marblestone published an essay about their findings, it helped attract philanthropic funding, which Marblestone, Kalil, and co-founder Anastasia Gamick used to launch Convergent Research, a nonprofit science studio for launching FROs.

“I see Ed’s lab as a melting pot where myself, Ed, Sam, and others worked on articulating a need and identifying specific projects that might make sense as FROs,” Marblestone says. “All those ideas later got crystallized when we created Convergent Research.”

In 2021, Convergent helped launch the first FROs: E11 Bio, which is led by Payne and committed to developing tools to understand how the brain is wired, and Cultivarium, a FRO making microorganisms more accessible for work in synthetic biology.

“From our brain mapping work we started asking the question, ‘Are there other projects that look like this that aren’t getting funded?’” Payne says. “We realized there was a gap in the research ecosystem, where some of these interdisciplinary, team science projects were being systematically overlooked. We knew a lot of amazing things would come out of getting those projects funded.”

Tools to advance science

Early progress from the first focused research organizations has strengthened Marblestone’s conviction that they’re filling a gap.

[C]Worthy is the FRO building tools to ensure safe, ocean-based carbon dioxide removal. It recently released an interactive map of alkaline activity to improve our understanding of one method for sequestering carbon known as ocean alkalinity enhancement. Last year, a math FRO, Lean, released a programming language and proof assistant that was used by Google’s DeepMind AI lab to solve problems in the International Mathematical Olympiad, achieving the same level as a silver medalist in the competition for the first time. The synthetic biology FRO Cultivarium, in turn, has already released software that can predict growth conditions for microbes based on their genome.

Last year, E11 Bio previewed a new method for mapping the brain called PRISM, which it has used to map out a portion of the mouse hippocampus. It will be making the data and mapping tool available to all researchers in coming months.

“A lot of this early work has proven you can put a really talented team together and move fast to go from zero to one,” Payne says. “The next phase is proving FROs can continue to build on that momentum and develop even more datasets and tools, establish even bigger collaborations, and scale their impact.”

Payne credits Boyden for fostering an ecosystem where researchers could think about problems beyond their narrow area of study.

“Ed’s lab was a really intellectually stimulating, collaborative environment,” Payne says. “He trains his students to think about impact first and work backward. It was a bunch of people thinking about how they were going to change the world, and that made it a particularly good place to develop the FRO idea.”

Marblestone says supporting FROs has been the highest-impact thing he’s been able to do in his career. Still, he believes the success of FROs should be judged over closer to 10-year periods and will depend on not just the tools they produce but also whether they spin out companies, partner with other institutes, and create larger, long-lasting initiatives to deploy what they built.

“We were initially worried people wouldn’t be willing to join these organizations because it doesn’t offer tenure and it doesn’t offer equity in a startup,” Marblestone says. “But we’ve been able to recruit excellent leaders, scientists, engineers, and others to create highly motivated teams. That’s good evidence this is working. As we get strong projects and good results, I hope it will create this flywheel where it becomes easier to fund these ideas, more scientists will come up with them, and I think we’re starting to get there.”

At the level of molecules and cells, ketamine and dexmedetomidine work very differently, but in the operating room they do the same exact thing: anesthetize the patient. By demonstrating how these distinct drugs achieve the same result, a new study in animals by neuroscientists at The Picower Institute for Learning and Memory at MIT identifies a potential signature of unconsciousness that is readily measurable to improve anesthesiology care.

What the two drugs have in common, the researchers discovered, is the way they push around brain waves, which are produced by the collective electrical activity of neurons. When brain waves are in phase, meaning the peaks and valleys of the waves are aligned, local groups of neurons in the brain’s cortex can share information to produce conscious cognitive functions such as attention, perception, and reasoning, says Picower Professor Earl K. Miller, senior author of the new study in Cell Reports. When brain waves fall out of phase, local communications, and therefore functions, fall apart, producing unconsciousness.

The finding, led by graduate student Alexandra Bardon, not only adds to scientists’ understanding of the dividing line between consciousness and unconsciousness, Miller says, but also could provide a common new measure for anesthesiologists who use a variety of different anesthetics to maintain patients on the proper side of that line during surgery.

“If you look at the way phase is shifted in our recordings, you can barely tell which drug it was,” says Miller, a faculty member in the Picower Institute and MIT’s Department of Brain and Cognitive Sciences. “That’s valuable for medical practice. Plus if unconsciousness has a universal signature, it could also reveal the mechanisms that generate consciousness.”

If more anesthetic drugs are also shown to affect phase in the same way, then anesthesiologists might be able to use brain wave phase alignment as a reliable marker of unconsciousness as they titrate doses of anesthetic drugs, Miller says, regardless of which particular mix of drugs they are using. That insight could aid efforts to build closed-loop systems that can aid anesthesiologists by constantly adjusting drug dose based on brain wave measurements of the patient’s unconsciousness.

Miller has been collaborating with study co-author Emery N. Brown, an anesthesiologist and Edward Hood Taplin Professor of Computational Neuroscience and Medical Engineering in the Picower Institute, on building such a system. In a recent clinical trial with colleagues in Japan, Brown demonstrated that monitoring brain wave power signals using EEG enabled an anesthesiologist to use much less sevoflurane during surgery with young children. The reduced doses proved safe and were associated with many improved clinical outcomes, including a reduced incidence of post-operative delirium.

Phase findings

Neuroscientists studying anesthesia have rarely paid attention to phase, but in the new study, Bardon, Brown, and Miller’s team made a point of it as they anesthetized two animals.

After the animals lost consciousness, the measurements indicated a substantial increase in “phase locking,” especially at low frequencies. Phase locking means that the relative differences in phase remained more stable. But what caught the researchers’ attention were the differences that became locked in: within each hemisphere, regardless of which anesthetic they used, brain wave phase became misaligned between the dorsolateral and ventrolateral regions of the prefrontal cortex.

Surprisingly, brain wave phase across hemispheres became more aligned, not less. But Miller notes that case is still a big shift from the conscious state, in which brain hemispheres are typically not aligned well, so the finding is a further indication that major changes of phase alignment, albeit in different ways at different distances, are a correlate of unconsciousness compared to wakefulness.

“The increase in interhemispheric alignment of activity by anesthetics seems to reverse the pattern observed in the awake, cognitively engaged brain,” the Bardon and Miller team wrote in Cell Reports.

Determined by distance

Distance proved to be a major factor in determining the change in phase alignment. Even across the 2.5 millimeters of a single electrode array, low-frequency waves moved 20-30 degrees out of alignment. Across the 20 or so millimeters between arrays in the upper (dorsolateral) and lower (ventrolateral) regions within a hemisphere, that would mean a roughly 180-degree shift in phase alignment, which is a complete offset of the waves.

The dependence on distance is consistent with the idea of waves traveling across the cortex, Miller says. Indeed, in a 2022 study, Miller and Brown’s labs showed that the anesthetic propofol induced a powerful low-frequency traveling wave that swept straight across the cortex, overwhelming higher-frequency straight and rotating waves.

The new results raise many opportunities for follow-up studies, Miller says. Does propofol also produce this signature of changed phase alignment? What role do traveling waves play in the phenomenon? And given that sleep is also characterized by increased power in slow wave frequencies, but is definitely not the same state as anesthesia-induced unconsciousness, could phase alignment explain the difference?

In addition to Bardon, Brown, and Miller, the paper’s other authors are Jesus Ballesteros, Scott Brincat, Jefferson Roy, Meredith Mahnke, and Yumiko Ishizawa.

The U.S. Department of Energy, the National Institutes of Health, the Simons Center for the Social Brain, the Freedom Together Foundation, and the Picower Institute provided support for the research.

Ready for that long-awaited summer vacation? First, you’ll need to pack all items required for your trip into a suitcase, making sure everything fits securely without crushing anything fragile.

Because humans possess strong visual and geometric reasoning skills, this is usually a straightforward problem, even if it may take a bit of finagling to squeeze everything in.

To a robot, though, it is an extremely complex planning challenge that requires thinking simultaneously about many actions, constraints, and mechanical capabilities. Finding an effective solution could take the robot a very long time — if it can even come up with one.

Researchers from MIT and NVIDIA Research have developed a novel algorithm that dramatically speeds up the robot’s planning process. Their approach enables a robot to “think ahead” by evaluating thousands of possible solutions in parallel and then refining the best ones to meet the constraints of the robot and its environment.

Instead of testing each potential action one at a time, like many existing approaches, this new method considers thousands of actions simultaneously, solving multistep manipulation problems in a matter of seconds.

The researchers harness the massive computational power of specialized processors called graphics processing units (GPUs) to enable this speedup.

In a factory or warehouse, their technique could enable robots to rapidly determine how to manipulate and tightly pack items that have different shapes and sizes without damaging them, knocking anything over, or colliding with obstacles, even in a narrow space.

“This would be very helpful in industrial settings where time really does matter and you need to find an effective solution as fast as possible. If your algorithm takes minutes to find a plan, as opposed to seconds, that costs the business money,” says MIT graduate student William Shen SM ’23, lead author of the paper on this technique.

He is joined on the paper by Caelan Garrett ’15, MEng ’15, PhD ’21, a senior research scientist at NVIDIA Research; Nishanth Kumar, an MIT graduate student; Ankit Goyal, a NVIDIA research scientist; Tucker Hermans, a NVIDIA research scientist and associate professor at the University of Utah; Leslie Pack Kaelbling, the Panasonic Professor of Computer Science and Engineering at MIT and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); Tomás Lozano-Pérez, an MIT professor of computer science and engineering and a member of CSAIL; and Fabio Ramos, principal research scientist at NVIDIA and a professor at the University of Sydney. The research will be presented at the Robotics: Science and Systems Conference.

Planning in parallel

The researchers’ algorithm is designed for what is called task and motion planning (TAMP). The goal of a TAMP algorithm is to come up with a task plan for a robot, which is a high-level sequence of actions, along with a motion plan, which includes low-level action parameters, like joint positions and gripper orientation, that complete that high-level plan.

To create a plan for packing items in a box, a robot needs to reason about many variables, such as the final orientation of packed objects so they fit together, as well as how it is going to pick them up and manipulate them using its arm and gripper.

It must do this while determining how to avoid collisions and achieve any user-specified constraints, such as a certain order in which to pack items.

With so many potential sequences of actions, sampling possible solutions at random and trying one at a time could take an extremely long time.

“It is a very large search space, and a lot of actions the robot does in that space don’t actually achieve anything productive,” Garrett adds.

Instead, the researchers’ algorithm, called cuTAMP, which is accelerated using a parallel computing platform called CUDA, simulates and refines thousands of solutions in parallel. It does this by combining two techniques, sampling and optimization.

Sampling involves choosing a solution to try. But rather than sampling solutions randomly, cuTAMP limits the range of potential solutions to those most likely to satisfy the problem’s constraints. This modified sampling procedure allows cuTAMP to broadly explore potential solutions while narrowing down the sampling space.

“Once we combine the outputs of these samples, we get a much better starting point than if we sampled randomly. This ensures we can find solutions more quickly during optimization,” Shen says.

Once cuTAMP has generated that set of samples, it performs a parallelized optimization procedure that computes a cost, which corresponds to how well each sample avoids collisions and satisfies the motion constraints of the robot, as well as any user-defined objectives.

It updates the samples in parallel, chooses the best candidates, and repeats the process until it narrows them down to a successful solution.

Harnessing accelerated computing

The researchers leverage GPUs, specialized processors that are far more powerful for parallel computation and workloads than general-purpose CPUs, to scale up the number of solutions they can sample and optimize simultaneously. This maximized the performance of their algorithm.

“Using GPUs, the computational cost of optimizing one solution is the same as optimizing hundreds or thousands of solutions,” Shen explains.

When they tested their approach on Tetris-like packing challenges in simulation, cuTAMP took only a few seconds to find successful, collision-free plans that might take sequential planning approaches much longer to solve.

And when deployed on a real robotic arm, the algorithm always found a solution in under 30 seconds.

The system works across robots and has been tested on a robotic arm at MIT and a humanoid robot at NVIDIA. Since cuTAMP is not a machine-learning algorithm, it requires no training data, which could enable it to be readily deployed in many situations.

“You can give it a brand-new problem and it will provably solve it,” Garrett says.

The algorithm is generalizable to situations beyond packing, like a robot using tools. A user could incorporate different skill types into the system to expand a robot’s capabilities automatically.

In the future, the researchers want to leverage large language models and vision language models within cuTAMP, enabling a robot to formulate and execute a plan that achieves specific objectives based on voice commands from a user.

This work is supported, in part, by the National Science Foundation (NSF), Air Force Office for Scientific Research, Office of Naval Research, MIT Quest for Intelligence, NVIDIA, and the Robotics and Artificial Intelligence Institute.

MIT physicists have demonstrated a new form of magnetism that could one day be harnessed to build faster, denser, and less power-hungry “spintronic” memory chips.

The new magnetic state is a mash-up of two main forms of magnetism: the ferromagnetism of everyday fridge magnets and compass needles, and antiferromagnetism, in which materials have magnetic properties at the microscale yet are not macroscopically magnetized.

Now, the MIT team has demonstrated a new form of magnetism, termed “p-wave magnetism.”

Physicists have long observed that electrons of atoms in regular ferromagnets share the same orientation of “spin,” like so many tiny compasses pointing in the same direction. This spin alignment generates a magnetic field, which gives a ferromagnet its inherent magnetism. Electrons belonging to magnetic atoms in an antiferromagnet also have spin, although these spins alternate, with electrons orbiting neighboring atoms aligning their spins antiparallel to each other. Taken together, the equal and opposite spins cancel out, and the antiferromagnet does not exhibit macroscopic magnetization.

The team discovered the new p-wave magnetism in nickel iodide (NiI₂), a two-dimensional crystalline material that they synthesized in the lab. Like a ferromagnet, the electrons exhibit a preferred spin orientation, and, like an antiferromagnet, equal populations of opposite spins result in a net cancellation. However, the spins on the nickel atoms exhibit a unique pattern, forming spiral-like configurations within the material that are mirror-images of each other, much like the left hand is the right hand’s mirror image.

What’s more, the researchers found this spiral spin configuration enabled them to carry out “spin switching”: Depending on the direction of spiraling spins in the material, they could apply a small electric field in a related direction to easily flip a left-handed spiral of spins into a right-handed spiral of spins, and vice-versa.

The ability to switch electron spins is at the heart of “spintronics,” which is a proposed alternative to conventional electronics. With this approach, data can be written in the form of an electron’s spin, rather than its electronic charge, potentially allowing orders of magnitude more data to be packed onto a device while using far less power to write and read that data.   

“We showed that this new form of magnetism can be manipulated electrically,” says Qian Song, a research scientist in MIT’s Materials Research Laboratory. “This breakthrough paves the way for a new class of ultrafast, compact, energy-efficient, and nonvolatile magnetic memory devices.”

Song and his colleagues published their results May 28 in the journal Nature. MIT co-authors include Connor Occhialini, Batyr Ilyas, Emre Ergeçen, Nuh Gedik, and Riccardo Comin, along with Rafael Fernandes at the University of Illinois Urbana-Champaign, and collaborators from multiple other institutions.

Connecting the dots

The discovery expands on work by Comin’s group in 2022. At that time, the team probed the magnetic properties of the same material, nickel iodide. At the microscopic level, nickel iodide resembles a triangular lattice of nickel and iodine atoms. Nickel is the material’s main magnetic ingredient, as the electrons on the nickel atoms exhibit spin, while those on iodine atoms do not.

In those experiments, the team observed that the spins of those nickel atoms were arranged in a spiral pattern throughout the material’s lattice, and that this pattern could spiral in two different orientations.

At the time, Comin had no idea that this unique pattern of atomic spins could enable precise switching of spins in surrounding electrons. This possibility was later raised by collaborator Rafael Fernandes, who along with other theorists was intrigued by a recently proposed idea for a new, unconventional, “p-wave” magnet, in which electrons moving along opposite directions in the material would have their spins aligned in opposite directions.

Fernandes and his colleagues recognized that if the spins of atoms in a material form the geometric spiral arrangement that Comin observed in nickel iodide, that would be a realization of a “p-wave” magnet. Then, when an electric field is applied to switch the “handedness” of the spiral, it should also switch the spin alignment of the electrons traveling along the same direction.

In other words, such a p-wave magnet could enable simple and controllable switching of electron spins, in a way that could be harnessed for spintronic applications.

“It was a completely new idea at the time, and we decided to test it experimentally because we realized nickel iodide was a good candidate to show this kind of p-wave magnet effect,” Comin says.

Spin current

For their new study, the team synthesized single-crystal flakes of nickel iodide by first depositing powders of the respective elements on a crystalline substrate, which they placed in a high-temperature furnace. The process causes the elements to settle into layers, each arranged microscopically in a triangular lattice of nickel and iodine atoms.

“What comes out of the oven are samples that are several millimeters wide and thin, like cracker bread,” Comin says. “We then exfoliate the material, peeling off even smaller flakes, each several microns wide, and a few tens of nanometers thin.”

The researchers wanted to know if, indeed, the spiral geometry of the nickel atoms’s spins would force electrons traveling in opposite directions to have opposite spins, like what Fernandes expected a p-wave magnet should exhibit. To observe this, the group applied to each flake a beam of circularly polarized light — light that produces an electric field that rotates in a particular direction, for instance, either clockwise or counterclockwise.

They reasoned that if travelling electrons interacting with the spin spirals have a spin that is aligned in the same direction, then incoming light, polarized in that same direction, should resonate and produce a characteristic signal. Such a signal would confirm that the traveling electrons’ spins align because of the spiral configuration, and furthermore, that the material does in fact exhibit p-wave magnetism.

And indeed, that’s what the group found. In experiments with multiple nickel iodide flakes, the researchers directly observed that the direction of the electron’s spin was correlated to the handedness of the light used to excite those electrons. Such is a telltale signature of p-wave magnetism, here observed for the first time.

Going a step further, they looked to see whether they could switch the spins of the electrons by applying an electric field, or a small amount of voltage, along different directions through the material. They found that when the direction of the electric field was in line with the direction of the spin spiral, the effect switched electrons along the route to spin in the same direction, producing a current of like-spinning electrons.

“With such a current of spin, you can do interesting things at the device level, for instance, you could flip magnetic domains that can be used for control of a magnetic bit,” Comin explains. “These spintronic effects are more efficient than conventional electronics because you’re just moving spins around, rather than moving charges. That means you’re not subject to any dissipation effects that generate heat, which is essentially the reason computers heat up.”

“We just need a small electric field to control this magnetic switching,” Song adds. “P-wave magnets could save five orders of magnitude of energy. Which is huge.”

“We are excited to see these cutting-edge experiments confirm our prediction of p-wave spin polarized states,” says Libor Šmejkal, head of the Max Planck Research Group in Dresden, Germany, who is one of the authors of the theoretical work that proposed the concept of p-wave magnetism but was not involved in the new paper. “The demonstration of electrically switchable p-wave spin polarization also highlights the promising applications of unconventional magnetic states.”

The team observed p-wave magnetism in nickel iodide flakes, only at ultracold temperatures of about 60 kelvins.

“That’s below liquid nitrogen, which is not necessarily practical for applications,” Comin says. “But now that we’ve realized this new state of magnetism, the next frontier is finding a material with these properties, at room temperature. Then we can apply this to a spintronic device.”

This research was supported, in part, by the National Science Foundation, the Department of Energy, and the Air Force Office of Scientific Research.

The accumulation of microplastics in the environment, and within our bodies, is an increasingly worrisome issue. But predicting where these ubiquitous particles will accumulate, and therefore where remediation efforts should be focused, has been difficult because of the many factors that contribute to their dispersal and deposition.

New research from MIT shows that one key factor in determining where microparticles are likely to build up has to do with the presence of biofilms. These thin, sticky biopolymer layers are shed by microorganisms and can accumulate on surfaces, including along sandy riverbeds or seashores. The study found that, all other conditions being equal, microparticles are less likely to accumulate in sediment infused with biofilms, because if they land there, they are more likely to be resuspended by flowing water and carried away.

The open-access findings appear in the journal Geophysical Research Letters, in a paper by MIT postdoc Hyoungchul Park and professor of civil and environmental engineering Heidi Nepf. “Microplastics are definitely in the news a lot,” Nepf says, “and we don’t fully understand where the hotspots of accumulation are likely to be. This work gives a little bit of guidance” on some of the factors that can cause these particles, and small particles in general, to accumulate in certain locations.

Most experiments looking at the ways microparticles are transported and deposited have been conducted over bare sand, Park says. “But in nature, there are a lot of microorganisms, such as bacteria, fungi, and algae, and when they adhere to the stream bed they generate some sticky things.” These substances are known as extracellular polymeric substances, or EPS, and they “can significantly affect the channel bed characteristics,” he says. The new research focused on determining exactly how these substances affected the transport of microparticles, including microplastics.

The research involved a flow tank with a bottom lined with fine sand, and sometimes with vertical plastic tubes simulating the presence of mangrove roots. In some experiments the bed consisted of pure sand, and in others the sand was mixed with a biological material to simulate the natural biofilms found in many riverbed and seashore environments.

Water mixed with tiny plastic particles was pumped through the tank for three hours, and then the bed surface was photographed under ultraviolet light that caused the plastic particles to fluoresce, allowing a quantitative measurement of their concentration.

The results revealed two different phenomena that affected how much of the plastic accumulated on the different surfaces. Immediately around the rods that stood in for above-ground roots, turbulence prevented particle deposition. In addition, as the amount of simulated biofilms in the sediment bed increased, the accumulation of particles also decreased.

Nepf and Park concluded that the biofilms filled up the spaces between the sand grains, leaving less room for the microparticles to fit in. The particles were more exposed because they penetrated less deeply in between the sand grains, and as a result they were much more easily resuspended and carried away by the flowing water.

“These biological films fill the pore spaces between the sediment grains,” Park explains, “and that makes the deposited particles — the particles that land on the bed — more exposed to the forces generated by the flow, which makes it easier for them to be resuspended. What we found was that in a channel with the same flow conditions and the same vegetation and the same sand bed, if one is without EPS and one is with EPS, then the one without EPS has a much higher deposition rate than the one with EPS.”

Nepf adds: “The biofilm is blocking the plastics from accumulating in the bed because they can’t go deep into the bed. They just stay right on the surface, and then they get picked up and moved elsewhere. So, if I spilled a large amount of microplastic in two rivers, and one had a sandy or gravel bottom, and one was muddier with more biofilm, I would expect more of the microplastics to be retained in the sandy or gravelly river.”

All of this is complicated by other factors, such as the turbulence of the water or the roughness of the bottom surface, she says. But it provides a “nice lens” to provide some suggestions for people who are trying to study the impacts of microplastics in the field. “They’re trying to determine what kinds of habitats these plastics are in, and this gives a framework for how you might categorize those habitats,” she says. “It gives guidance to where you should go to find more plastics versus less.”

As an example, Park suggests, in mangrove ecosystems, microplastics may preferentially accumulate in the outer edges, which tend to be sandy, while the interior zones have sediment with more biofilm. Thus, this work suggests “the sandy outer regions may be potential hotspots for microplastic accumulation,” he says, and can make this a priority zone for monitoring and protection.

“This is a highly relevant finding,” says Isabella Schalko, a research scientist at ETH Zurich, who was not associated with this research. “It suggests that restoration measures such as re-vegetation or promoting biofilm growth could help mitigate microplastic accumulation in aquatic systems. It highlights the powerful role of biological and physical features in shaping particle transport processes.”

The work was supported by Shell International Exploration and Production through the MIT Energy Initiative.

Hydrogen has the potential to be a climate-friendly fuel since it doesn’t release carbon dioxide when used as an energy source. Currently, however, most methods for producing hydrogen involve fossil fuels, making hydrogen less of a “green” fuel over its entire life cycle.

A new process developed by MIT engineers could significantly shrink the carbon footprint associated with making hydrogen.

Last year, the team reported that they could produce hydrogen gas by combining seawater, recycled soda cans, and caffeine. The question then was whether the benchtop process could be applied at an industrial scale, and at what environmental cost.

Now, the researchers have carried out a “cradle-to-grave” life cycle assessment, taking into account every step in the process at an industrial scale. For instance, the team calculated the carbon emissions associated with acquiring and processing aluminum, reacting it with seawater to produce hydrogen, and transporting the fuel to gas stations, where drivers could tap into hydrogen tanks to power engines or fuel cell cars. They found that, from end to end, the new process could generate a fraction of the carbon emissions that is associated with conventional hydrogen production.

In a study appearing today in Cell Reports Sustainability, the team reports that for every kilogram of hydrogen produced, the process would generate 1.45 kilograms of carbon dioxide over its entire life cycle. In comparison, fossil-fuel-based processes emit 11 kilograms of carbon dioxide per kilogram of hydrogen generated.

The low-carbon footprint is on par with other proposed “green hydrogen” technologies, such as those powered by solar and wind energy.

“We’re in the ballpark of green hydrogen,” says lead author Aly Kombargi PhD ’25, who graduated this spring from MIT with a doctorate in mechanical engineering. “This work highlights aluminum’s potential as a clean energy source and offers a scalable pathway for low-emission hydrogen deployment in transportation and remote energy systems.”

The study’s MIT co-authors are Brooke Bao, Enoch Ellis, and professor of mechanical engineering Douglas Hart.

Gas bubble

Dropping an aluminum can in water won’t normally cause much of a chemical reaction. That’s because when aluminum is exposed to oxygen, it instantly forms a shield-like layer. Without this layer, aluminum exists in its pure form and can readily react when mixed with water. The reaction that occurs involves aluminum atoms that efficiently break up molecules of water, producing aluminum oxide and pure hydrogen. And it doesn’t take much of the metal to bubble up a significant amount of the gas.

“One of the main benefits of using aluminum is the energy density per unit volume,” Kombargi says. “With a very small amount of aluminum fuel, you can conceivably supply much of the power for a hydrogen-fueled vehicle.”

Last year, he and Hart developed a recipe for aluminum-based hydrogen production. They found they could puncture aluminum’s natural shield by treating it with a small amount of gallium-indium, which is a rare-metal alloy that effectively scrubs aluminum into its pure form. The researchers then mixed pellets of pure aluminum with seawater and observed that the reaction produced pure hydrogen. What’s more, the salt in the water helped to precipitate gallium-indium, which the team could subsequently recover and reuse to generate more hydrogen, in a cost-saving, sustainable cycle.

“We were explaining the science of this process in conferences, and the questions we would get were, ‘How much does this cost?’ and, ‘What’s its carbon footprint?’” Kombargi says. “So we wanted to look at the process in a comprehensive way.”

A sustainable cycle

For their new study, Kombargi and his colleagues carried out a life cycle assessment to estimate the environmental impact of aluminum-based hydrogen production, at every step of the process, from sourcing the aluminum to transporting the hydrogen after production. They set out to calculate the amount of carbon associated with generating 1 kilogram of hydrogen — an amount that they chose as a practical, consumer-level illustration.

“With a hydrogen fuel cell car using 1 kilogram of hydrogen, you can go between 60 to 100 kilometers, depending on the efficiency of the fuel cell,” Kombargi notes.

They performed the analysis using Earthster — an online life cycle assessment tool that draws data from a large repository of products and processes and their associated carbon emissions. The team considered a number of scenarios to produce hydrogen using aluminum, from starting with “primary” aluminum mined from the Earth, versus “secondary” aluminum that is recycled from soda cans and other products, and using various methods to transport the aluminum and hydrogen.

After running life cycle assessments for about a dozen scenarios, the team identified one scenario with the lowest carbon footprint. This scenario centers on recycled aluminum — a source that saves a significant amount of emissions compared with mining aluminum — and seawater — a natural resource that also saves money by recovering gallium-indium. They found that this scenario, from start to finish, would generate about 1.45 kilograms of carbon dioxide for every kilogram of hydrogen produced. The cost of the fuel produced, they calculated, would be about $9 per kilogram, which is comparable to the price of hydrogen that would be generated with other green technologies such as wind and solar energy.

The researchers envision that if the low-carbon process were ramped up to a commercial scale, it would look something like this: The production chain would start with scrap aluminum sourced from a recycling center. The aluminum would be shredded into pellets and treated with gallium-indium. Then, drivers could transport the pretreated pellets as aluminum “fuel,” rather than directly transporting hydrogen, which is potentially volatile. The pellets would be transported to a fuel station that ideally would be situated near a source of seawater, which could then be mixed with the aluminum, on demand, to produce hydrogen. A consumer could then directly pump the gas into a car with either an internal combustion engine or a fuel cell.

The entire process does produce an aluminum-based byproduct, boehmite, which is a mineral that is commonly used in fabricating semiconductors, electronic elements, and a number of industrial products. Kombargi says that if this byproduct were recovered after hydrogen production, it could be sold to manufacturers, further bringing down the cost of the process as a whole.

“There are a lot of things to consider,” Kombargi says. “But the process works, which is the most exciting part. And we show that it can be environmentally sustainable.”

The group is continuing to develop the process. They recently designed a small reactor, about the size of a water bottle, that takes in aluminum pellets and seawater to generate hydrogen, enough to power an electric bike for several hours. They previously demonstrated that the process can produce enough hydrogen to fuel a small car. The team is also exploring underwater applications, and are designing a hydrogen reactor that would take in surrounding seawater to power a small boat or underwater vehicle.

This research was supported, in part, by the MIT Portugal Program.

Hearing aids, mouth guards, dental implants, and other highly tailored structures are often products of 3D printing. These structures are typically made via vat photopolymerization — a form of 3D printing that uses patterns of light to shape and solidify a resin, one layer at a time.

The process also involves printing structural supports from the same material to hold the product in place as it’s printed. Once a product is fully formed, the supports are removed manually and typically thrown out as unusable waste.

MIT engineers have found a way to bypass this last finishing step, in a way that could significantly speed up the 3D-printing process. They developed a resin that turns into two different kinds of solids, depending on the type of light that shines on it: Ultraviolet light cures the resin into an highly resilient solid, while visible light turns the same resin into a solid that is easily dissolvable in certain solvents.

The team exposed the new resin simultaneously to patterns of UV light to form a sturdy structure, as well as patterns of visible light to form the structure’s supports. Instead of having to carefully break away the supports, they simply dipped the printed material into solution that dissolved the supports away, revealing the sturdy, UV-printed part.

The supports can dissolve in a variety of food-safe solutions, including baby oil. Interestingly, the supports could even dissolve in the main liquid ingredient of the original resin, like a cube of ice in water. This means that the material used to print structural supports could be continuously recycled: Once a printed structure’s supporting material dissolves, that mixture can be blended directly back into fresh resin and used to print the next set of parts — along with their dissolvable supports.

The researchers applied the new method to print complex structures, including functional gear trains and intricate lattices.

“You can now print — in a single print — multipart, functional assemblies with moving or interlocking parts, and you can basically wash away the supports,” says graduate student Nicholas Diaco. “Instead of throwing out this material, you can recycle it on site and generate a lot less waste. That’s the ultimate hope.”

He and his colleagues report the details of the new method in a paper appearing today in Advanced Materials Technologies. The MIT study’s co-authors include Carl Thrasher, Max Hughes, Kevin Zhou, Michael Durso, Saechow Yap, Professor Robert Macfarlane, and Professor A. John Hart, head of MIT’s Department of Mechanical Engineering.

Waste removal

Conventional vat photopolymerization (VP) begins with a 3D computer model of a structure to be printed — for instance, of two interlocking gears. Along with the gears themselves, the model includes small support structures around, under, and between the gears to keep every feature in place as the part is printed. This computer model is then sliced into many digital layers that are sent to a VP printer for printing.

A standard VP printer includes a small vat of liquid resin that sits over a light source. Each slice of the model is translated into a matching pattern of light that is projected onto the liquid resin, which solidifies into the same pattern. Layer by layer, a solid, light-printed version of the model’s gears and supports forms on the build platform. When printing is finished, the platform lifts the completed part above the resin bath. Once excess resin is washed away, a person can go in by hand to remove the intermediary supports, usually by clipping and filing, and the support material is ultimately thrown away.

“For the most part, these supports end up generating a lot of waste,” Diaco says.

Print and dip

Diaco and the team looked for a way to simplify and speed up the removal of printed supports and, ideally, recycle them in the process. They came up with a general concept for a resin that, depending on the type of light that it is exposed to, can take on one of two phases: a resilient phase that would form the desired 3D structure and a secondary phase that would function as a supporting material but also be easily dissolved away.

After working out some chemistry, the team found they could make such a two-phase resin by mixing two commercially available monomers, the chemical building blocks that are found in many types of plastic. When ultraviolet light shines on the mixture, the monomers link together into a tightly interconnected network, forming a tough solid that resists dissolution. When the same mixture is exposed to visible light, the same monomers still cure, but at the molecular scale the resulting monomer strands remain separate from one another. This solid can quickly dissolve when placed in certain solutions.

In benchtop tests with small vials of the new resin, the researchers found the material did transform into both the insoluble and soluble forms in response to ultraviolet and visible light, respectively. But when they moved to a 3D printer with LEDs dimmer than the benchtop setup, the UV-cured material fell apart in solution. The weaker light only partially linked the monomer strands, leaving them too loosely tangled to hold the structure together.

Diaco and his colleagues found that adding a small amount of a third “bridging” monomer could link the two original monomers together under UV light, knitting them into a much sturdier framework. This fix enabled the researchers to simultaneously print resilient 3D structures and dissolvable supports using timed pulses of UV and visible light in one run.

The team applied the new method to print a variety of intricate structures, including interlocking gears, intricate lattices, a ball within a square frame, and, for fun, a small dinosaur encased in an egg-shaped support that dissolved away when dipped in solution.

“With all these structures, you need a lattice of supports inside and out while printing,” Diaco says. “Removing those supports normally requires careful, manual removal. This shows we can print multipart assemblies with a lot of moving parts, and detailed, personalized products like hearing aids and dental implants, in a way that’s fast and sustainable.”

“We’ll continue studying the limits of this process, and we want to develop additional resins with this wavelength-selective behavior and mechanical properties necessary for durable products,” says professor of mechanical engineering John Hart. “Along with automated part handling and closed-loop reuse of the dissolved resin, this is an exciting path to resource-efficient and cost-effective polymer 3D printing at scale.”

This research was supported, in part, by the Center for Perceptual and Interactive Intelligence (InnoHK) in Hong Kong, the U.S. National Science Foundation, the U.S. Office of Naval Research, and the U.S. Army Research Office.

For weeks, the whiteboard in the lab was crowded with scribbles, diagrams, and chemical formulas. A research team across the Olivetti Group and the MIT Concrete Sustainability Hub (CSHub) was working intensely on a key problem: How can we reduce the amount of cement in concrete to save on costs and emissions? 

The question was certainly not new; materials like fly ash, a byproduct of coal production, and slag, a byproduct of steelmaking, have long been used to replace some of the cement in concrete mixes. However, the demand for these products is outpacing supply as industry looks to reduce its climate impacts by expanding their use, making the search for alternatives urgent. The challenge that the team discovered wasn’t a lack of candidates; the problem was that there were too many to sort through.

On May 17, the team, led by postdoc Soroush Mahjoubi, published an open-access paper in Nature’s Communications Materials outlining their solution. “We realized that AI was the key to moving forward,” notes Mahjoubi. “There is so much data out there on potential materials — hundreds of thousands of pages of scientific literature. Sorting through them would have taken many lifetimes of work, by which time more materials would have been discovered!”

With large language models, like the chatbots many of us use daily, the team built a machine-learning framework that evaluates and sorts candidate materials based on their physical and chemical properties. 

“First, there is hydraulic reactivity. The reason that concrete is strong is that cement — the ‘glue’ that holds it together — hardens when exposed to water. So, if we replace this glue, we need to make sure the substitute reacts similarly,” explains Mahjoubi. “Second, there is pozzolanicity. This is when a material reacts with calcium hydroxide, a byproduct created when cement meets water, to make the concrete harder and stronger over time.  We need to balance the hydraulic and pozzolanic materials in the mix so the concrete performs at its best.”

Analyzing scientific literature and over 1 million rock samples, the team used the framework to sort candidate materials into 19 types, ranging from biomass to mining byproducts to demolished construction materials. Mahjoubi and his team found that suitable materials were available globally — and, more impressively, many could be incorporated into concrete mixes just by grinding them. This means it’s possible to extract emissions and cost savings without much additional processing. 

“Some of the most interesting materials that could replace a portion of cement are ceramics,” notes Mahjoubi. “Old tiles, bricks, pottery — all these materials may have high reactivity. That’s something we’ve observed in ancient Roman concrete, where ceramics were added to help waterproof structures. I’ve had many interesting conversations on this with Professor Admir Masic, who leads a lot of the ancient concrete studies here at MIT.”

The potential of everyday materials like ceramics and industrial materials like mine tailings is an example of how materials like concrete can help enable a circular economy. By identifying and repurposing materials that would otherwise end up in landfills, researchers and industry can help to give these materials a second life as part of our buildings and infrastructure.

Looking ahead, the research team is planning to upgrade the framework to be capable of assessing even more materials, while experimentally validating some of the best candidates. “AI tools have gotten this research far in a short time, and we are excited to see how the latest developments in large language models enable the next steps,” says Professor Elsa Olivetti, senior author on the work and member of the MIT Department of Materials Science and Engineering. She serves as an MIT Climate Project mission director, a CSHub principal investigator, and the leader of the Olivetti Group.

“Concrete is the backbone of the built environment,” says Randolph Kirchain, co-author and CSHub director. “By applying data science and AI tools to material design, we hope to support industry efforts to build more sustainably, without compromising on strength, safety, or durability.

In addition to Mahjoubi, Olivetti, and Kirchain, co-authors on the work include MIT postdoc Vineeth Venugopal, Ipek Bensu Manav SM ’21, PhD ’24; and CSHub Deputy Director Hessam AzariJafari.

This research was conducted through the MIT Concrete Sustainability Hub, which is supported by the Concrete Advancement Foundation. This work also received funding from the MIT-IBM Watson AI Lab.

When you’re trying to communicate or understand ideas, words don’t always do the trick. Sometimes the more efficient approach is to do a simple sketch of that concept — for example, diagramming a circuit might help make sense of how the system works.

But what if artificial intelligence could help us explore these visualizations? While these systems are typically proficient at creating realistic paintings and cartoonish drawings, many models fail to capture the essence of sketching: its stroke-by-stroke, iterative process, which helps humans brainstorm and edit how they want to represent their ideas.

A new drawing system from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Stanford University can sketch more like we do. Their method, called “SketchAgent,” uses a multimodal language model — AI systems that train on text and images, like Anthropic’s Claude 3.5 Sonnet — to turn natural language prompts into sketches in a few seconds. For example, it can doodle a house either on its own or through collaboration, drawing with a human or incorporating text-based input to sketch each part separately.

The researchers showed that SketchAgent can create abstract drawings of diverse concepts, like a robot, butterfly, DNA helix, flowchart, and even the Sydney Opera House. One day, the tool could be expanded into an interactive art game that helps teachers and researchers diagram complex concepts or give users a quick drawing lesson.

CSAIL postdoc Yael Vinker, who is the lead author of a paper introducing SketchAgent, notes that the system introduces a more natural way for humans to communicate with AI.

“Not everyone is aware of how much they draw in their daily life. We may draw our thoughts or workshop ideas with sketches,” she says. “Our tool aims to emulate that process, making multimodal language models more useful in helping us visually express ideas.”

SketchAgent teaches these models to draw stroke-by-stroke without training on any data — instead, the researchers developed a “sketching language” in which a sketch is translated into a numbered sequence of strokes on a grid. The system was given an example of how things like a house would be drawn, with each stroke labeled according to what it represented — such as the seventh stroke being a rectangle labeled as a “front door” — to help the model generalize to new concepts.

Vinker wrote the paper alongside three CSAIL affiliates — postdoc Tamar Rott Shaham, undergraduate researcher Alex Zhao, and MIT Professor Antonio Torralba — as well as Stanford University Research Fellow Kristine Zheng and Assistant Professor Judith Ellen Fan. They’ll present their work at the 2025 Conference on Computer Vision and Pattern Recognition (CVPR) this month.

Assessing AI’s sketching abilities

While text-to-image models such as DALL-E 3 can create intriguing drawings, they lack a crucial component of sketching: the spontaneous, creative process where each stroke can impact the overall design. On the other hand, SketchAgent’s drawings are modeled as a sequence of strokes, appearing more natural and fluid, like human sketches.

Prior works have mimicked this process, too, but they trained their models on human-drawn datasets, which are often limited in scale and diversity. SketchAgent uses pre-trained language models instead, which are knowledgeable about many concepts, but don’t know how to sketch. When the researchers taught language models this process, SketchAgent began to sketch diverse concepts it hadn’t explicitly trained on.

Still, Vinker and her colleagues wanted to see if SketchAgent was actively working with humans on the sketching process, or if it was working independently of its drawing partner. The team tested their system in collaboration mode, where a human and a language model work toward drawing a particular concept in tandem. Removing SketchAgent’s contributions revealed that their tool’s strokes were essential to the final drawing. In a drawing of a sailboat, for instance, removing the artificial strokes representing a mast made the overall sketch unrecognizable.

In another experiment, CSAIL and Stanford researchers plugged different multimodal language models into SketchAgent to see which could create the most recognizable sketches. Their default backbone model, Claude 3.5 Sonnet, generated the most human-like vector graphics (essentially text-based files that can be converted into high-resolution images). It outperformed models like GPT-4o and Claude 3 Opus.

“The fact that Claude 3.5 Sonnet outperformed other models like GPT-4o and Claude 3 Opus suggests that this model processes and generates visual-related information differently,” says co-author Tamar Rott Shaham.

She adds that SketchAgent could become a helpful interface for collaborating with AI models beyond standard, text-based communication. “As models advance in understanding and generating other modalities, like sketches, they open up new ways for users to express ideas and receive responses that feel more intuitive and human-like,” says Rott Shaham. “This could significantly enrich interactions, making AI more accessible and versatile.”

While SketchAgent’s drawing prowess is promising, it can’t make professional sketches yet. It renders simple representations of concepts using stick figures and doodles, but struggles to doodle things like logos, sentences, complex creatures like unicorns and cows, and specific human figures.

At times, their model also misunderstood users’ intentions in collaborative drawings, like when SketchAgent drew a bunny with two heads. According to Vinker, this may be because the model breaks down each task into smaller steps (also called “Chain of Thought” reasoning). When working with humans, the model creates a drawing plan, potentially misinterpreting which part of that outline a human is contributing to. The researchers could possibly refine these drawing skills by training on synthetic data from diffusion models.

Additionally, SketchAgent often requires a few rounds of prompting to generate human-like doodles. In the future, the team aims to make it easier to interact and sketch with multimodal language models, including refining their interface. 

Still, the tool suggests AI could draw diverse concepts the way humans do, with step-by-step human-AI collaboration that results in more aligned final designs.

This work was supported, in part, by the U.S. National Science Foundation, a Hoffman-Yee Grant from the Stanford Institute for Human-Centered AI, the Hyundai Motor Co., the U.S. Army Research Laboratory, the Zuckerman STEM Leadership Program, and a Viterbi Fellowship.

Every year, thousands of students take courses that teach them how to deploy artificial intelligence models that can help doctors diagnose disease and determine appropriate treatments. However, many of these courses omit a key element: training students to detect flaws in the training data used to develop the models.

Leo Anthony Celi, a senior research scientist at MIT’s Institute for Medical Engineering and Science, a physician at Beth Israel Deaconess Medical Center, and an associate professor at Harvard Medical School, has documented these shortcomings in a new paper and hopes to persuade course developers to teach students to more thoroughly evaluate their data before incorporating it into their models. Many previous studies have found that models trained mostly on clinical data from white males don’t work well when applied to people from other groups. Here, Celi describes the impact of such bias and how educators might address it in their teachings about AI models.

Q: How does bias get into these datasets, and how can these shortcomings be addressed?

A: Any problems in the data will be baked into any modeling of the data. In the past we have described instruments and devices that don’t work well across individuals. As one example, we found that pulse oximeters overestimate oxygen levels for people of color, because there weren’t enough people of color enrolled in the clinical trials of the devices. We remind our students that medical devices and equipment are optimized on healthy young males. They were never optimized for an 80-year-old woman with heart failure, and yet we use them for those purposes. And the FDA does not require that a device work well on this diverse of a population that we will be using it on. All they need is proof that it works on healthy subjects.

Additionally, the electronic health record system is in no shape to be used as the building blocks of AI. Those records were not designed to be a learning system, and for that reason, you have to be really careful about using electronic health records. The electronic health record system is to be replaced, but that’s not going to happen anytime soon, so we need to be smarter. We need to be more creative about using the data that we have now, no matter how bad they are, in building algorithms.

One promising avenue that we are exploring is the development of a transformer model of numeric electronic health record data, including but not limited to laboratory test results. Modeling the underlying relationship between the laboratory tests, the vital signs and the treatments can mitigate the effect of missing data as a result of social determinants of health and provider implicit biases.

Q: Why is it important for courses in AI to cover the sources of potential bias? What did you find when you analyzed such courses’ content?

A: Our course at MIT started in 2016, and at some point we realized that we were encouraging people to race to build models that are overfitted to some statistical measure of model performance, when in fact the data that we’re using is rife with problems that people are not aware of. At that time, we were wondering: How common is this problem?

Our suspicion was that if you looked at the courses where the syllabus is available online, or the online courses, that none of them even bothers to tell the students that they should be paranoid about the data. And true enough, when we looked at the different online courses, it’s all about building the model. How do you build the model? How do you visualize the data? We found that of 11 courses we reviewed, only five included sections on bias in datasets, and only two contained any significant discussion of bias.

That said, we cannot discount the value of these courses. I’ve heard lots of stories where people self-study based on these online courses, but at the same time, given how influential they are, how impactful they are, we need to really double down on requiring them to teach the right skillsets, as more and more people are drawn to this AI multiverse. It’s important for people to really equip themselves with the agency to be able to work with AI. We’re hoping that this paper will shine a spotlight on this huge gap in the way we teach AI now to our students.

Q: What kind of content should course developers be incorporating?

A: One, giving them a checklist of questions in the beginning. Where did this data came from? Who were the observers? Who were the doctors and nurses who collected the data? And then learn a little bit about the landscape of those institutions. If it’s an ICU database, they need to ask who makes it to the ICU, and who doesn’t make it to the ICU, because that already introduces a sampling selection bias. If all the minority patients don’t even get admitted to the ICU because they cannot reach the ICU in time, then the models are not going to work for them. Truly, to me, 50 percent of the course content should really be understanding the data, if not more, because the modeling itself is easy once you understand the data.

Since 2014, the MIT Critical Data consortium has been organizing datathons (data “hackathons”) around the world. At these gatherings, doctors, nurses, other health care workers, and data scientists get together to comb through databases and try to examine health and disease in the local context. Textbooks and journal papers present diseases based on observations and trials involving a narrow demographic typically from countries with resources for research. 

Our main objective now, what we want to teach them, is critical thinking skills. And the main ingredient for critical thinking is bringing together people with different backgrounds.

You cannot teach critical thinking in a room full of CEOs or in a room full of doctors. The environment is just not there. When we have datathons, we don’t even have to teach them how do you do critical thinking. As soon as you bring the right mix of people — and it’s not just coming from different backgrounds but from different generations — you don’t even have to tell them how to think critically. It just happens. The environment is right for that kind of thinking. So, we now tell our participants and our students, please, please do not start building any model unless you truly understand how the data came about, which patients made it into the database, what devices were used to measure, and are those devices consistently accurate across individuals?

When we have events around the world, we encourage them to look for data sets that are local, so that they are relevant. There’s resistance because they know that they will discover how bad their data sets are. We say that that’s fine. This is how you fix that. If you don’t know how bad they are, you’re going to continue collecting them in a very bad manner and they’re useless. You have to acknowledge that you’re not going to get it right the first time, and that’s perfectly fine. MIMIC (the Medical Information Marked for Intensive Care database built at Beth Israel Deaconess Medical Center) took a decade before we had a decent schema, and we only have a decent schema because people were telling us how bad MIMIC was.

We may not have the answers to all of these questions, but we can evoke something in people that helps them realize that there are so many problems in the data. I’m always thrilled to look at the blog posts from people who attended a datathon, who say that their world has changed. Now they’re more excited about the field because they realize the immense potential, but also the immense risk of harm if they don’t do this correctly.

Scientists at the McGovern Institute for Brain Research at MIT and the Broad Institute of MIT and Harvard have re-engineered a compact RNA-guided enzyme they found in bacteria into an efficient, programmable editor of human DNA. 

The protein they created, called NovaIscB, can be adapted to make precise changes to the genetic code, modulate the activity of specific genes, or carry out other editing tasks. Because its small size simplifies delivery to cells, NovaIscB’s developers say it is a promising candidate for developing gene therapies to treat or prevent disease.

The study was led by Feng Zhang, the James and Patricia Poitras Professor of Neuroscience at MIT who is also an investigator at the McGovern Institute and the Howard Hughes Medical Institute, and a core member of the Broad Institute. Zhang and his team reported their open-access work this month in the journal Nature Biotechnology.

NovaIscB is derived from a bacterial DNA cutter that belongs to a family of proteins called IscBs, which Zhang’s lab discovered in 2021. IscBs are a type of OMEGA system, the evolutionary ancestors to Cas9, which is part of the bacterial CRISPR system that Zhang and others have developed into powerful genome-editing tools. Like Cas9, IscB enzymes cut DNA at sites specified by an RNA guide. By reprogramming that guide, researchers can redirect the enzymes to target sequences of their choosing.

IscBs had caught the team’s attention not only because they share key features of CRISPR’s DNA-cutting Cas9, but also because they are a third of its size. That would be an advantage for potential gene therapies: compact tools are easier to deliver to cells, and with a small enzyme, researchers would have more flexibility to tinker, potentially adding new functionalities without creating tools that were too bulky for clinical use.

From their initial studies of IscBs, researchers in Zhang’s lab knew that some members of the family could cut DNA targets in human cells. None of the bacterial proteins worked well enough to be deployed therapeutically, however: the team would have to modify an IscB to ensure it could edit targets in human cells efficiently without disturbing the rest of the genome.

To begin that engineering process, Soumya Kannan, a graduate student in Zhang’s lab who is now a junior fellow at the Harvard Society of Fellows, and postdoc Shiyou Zhu first searched for an IscB that would make good starting point. They tested nearly 400 different IscB enzymes that can be found in bacteria. Ten were capable of editing DNA in human cells.

Even the most active of those would need to be enhanced to make it a useful genome editing tool. The challenge would be increasing the enzyme’s activity, but only at the sequences specified by its RNA guide. If the enzyme became more active, but indiscriminately so, it would cut DNA in unintended places. “The key is to balance the improvement of both activity and specificity at the same time,” explains Zhu.

Zhu notes that bacterial IscBs are directed to their target sequences by relatively short RNA guides, which makes it difficult to restrict the enzyme’s activity to a specific part of the genome. If an IscB could be engineered to accommodate a longer guide, it would be less likely to act on sequences beyond its intended target.

To optimize IscB for human genome editing, the team leveraged information that graduate student Han Altae-Tran, who is now a postdoc at the University of Washington, had learned about the diversity of bacterial IscBs and how they evolved. For instance, the researchers noted that IscBs that worked in human cells included a segment they called REC, which was absent in other IscBs. They suspected the enzyme might need that segment to interact with the DNA in human cells. When they took a closer look at the region, structural modeling suggested that by slightly expanding part of the protein, REC might also enable IscBs to recognize longer RNA guides.

Based on these observations, the team experimented with swapping in parts of REC domains from different IscBs and Cas9s, evaluating how each change impacted the protein’s function. Guided by their understanding of how IscBs and Cas9s interact with both DNA and their RNA guides, the researchers made additional changes, aiming to optimize both efficiency and specificity.

In the end, they generated a protein they called NovaIscB, which was over 100 times more active in human cells than the IscB they had started with, and that had demonstrated good specificity for its targets.

Kannan and Zhu constructed and screened hundreds of new IscBs before arriving at NovaIscB — and every change they made to the original protein was strategic. Their efforts were guided by their team’s knowledge of IscBs’s natural evolution, as well as predictions of how each alteration would impact the protein’s structure, made using an artificial intelligence tool called AlphaFold2. Compared to traditional methods of introducing random changes into a protein and screening for their effects, this rational engineering approach greatly accelerated the team’s ability to identify a protein with the features they were looking for.

The team demonstrated that NovaIscB is a good scaffold for a variety of genome editing tools. “It biochemically functions very similarly to Cas9, and that makes it easy to port over tools that were already optimized with the Cas9 scaffold,” Kannan says. With different modifications, the researchers used NovaIscB to replace specific letters of the DNA code in human cells and to change the activity of targeted genes.

Importantly, the NovaIscB-based tools are compact enough to be easily packaged inside a single adeno-associated virus (AAV) — the vector most commonly used to safely deliver gene therapy to patients. Because they are bulkier, tools developed using Cas9 can require a more complicated delivery strategy.

Demonstrating NovaIscB’s potential for therapeutic use, Zhang’s team created a tool called OMEGAoff that adds chemical markers to DNA to dial down the activity of specific genes. They programmed OMEGAoff to repress a gene involved in cholesterol regulation, then used AAV to deliver the system to the livers of mice, leading to lasting reductions in cholesterol levels in the animals’ blood.

The team expects that NovaIscB can be used to target genome editing tools to most human genes, and look forward to seeing how other labs deploy the new technology. They also hope others will adopt their evolution-guided approach to rational protein engineering. “Nature has such diversity, and its systems have different advantages and disadvantages,” Zhu says. “By learning about that natural diversity, we can make the systems we are trying to engineer better and better.”

This study was funded, in part, by the K. Lisa Yang and Hock E. Tan Center for Molecular Therapeutics at MIT, Broad Institute Programmable Therapeutics Gift Donors, Pershing Square Foundation, William Ackman, Neri Oxman, the Phillips family, and J. and P. Poitras.

Consuming a high-fat diet can lead to a variety of health problems — not only weight gain but also an increased risk of diabetes and other chronic diseases.

At the cellular level, hundreds of changes take place in response to a high-fat diet. MIT researchers have now mapped out some of those changes, with a focus on metabolic enzyme dysregulation that is associated with weight gain.

Their study, conducted in mice, revealed that hundreds of enzymes involved in sugar, lipid, and protein metabolism are affected by a high-fat diet, and that these disruptions lead to an increase in insulin resistance and an accumulation of damaging molecules called reactive oxygen species. These effects were more pronounced in males than females.

The researchers also showed that most of the damage could be reversed by giving the mice an antioxidant along with their high-fat diet.

“Under metabolic stress conditions, enzymes can be affected to produce a more harmful state than what was initially there,” says Tigist Tamir, a former MIT postdoc. “Then what we’ve shown with the antioxidant study is that you can bring them to a different state that is less dysfunctional.”

Tamir, who is now an assistant professor of biochemistry and biophysics at the University of North Carolina at Chapel Hill School of Medicine, is the lead author of the new study, which appears today in Molecular Cell. Forest White, the Ned C. and Janet C. Rice Professor of Biological Engineering and a member of the Koch Institute for Integrative Cancer Research at MIT, is the senior author of the paper.

Metabolic networks

In previous work, White’s lab has found that a high-fat diet stimulates cells to turn on many of the same signaling pathways that are linked to chronic stress. In the new study, the researchers wanted to explore the role of enzyme phosphorylation in those responses.

Phosphorylation, or the addition of a phosphate group, can turn enzyme activity on or off. This process, which is controlled by enzymes called kinases, gives cells a way to quickly respond to environmental conditions by fine-tuning the activity of existing enzymes within the cell.

Many enzymes involved in metabolism — the conversion of food into the building blocks of key molecules such as proteins, lipids, and nucleic acids — are known to undergo phosphorylation.

The researchers began by analyzing databases of human enzymes that can be phosphorylated, focusing on enzymes involved in metabolism. They found that many of the metabolic enzymes that undergo phosphorylation belong to a class called oxidoreductases, which transfer electrons from one molecule to another. Such enzymes are key to metabolic reactions such as glycolysis — the breakdown of glucose into a smaller molecule known as pyruvate.

Among the hundreds of enzymes the researchers identified are IDH1, which is involved in breaking down sugar to generate energy, and AKR1C1, which is required for metabolizing fatty acids. The researchers also found that many phosphorylated enzymes are important for the management of reactive oxygen species, which are necessary for many cell functions but can be harmful if too many of them accumulate in a cell.

Phosphorylation of these enzymes can lead them to become either more or less active, as they work together to respond to the intake of food. Most of the metabolic enzymes identified in this study are phosphorylated on sites found in regions of the enzyme that are important for binding to the molecules that they act upon or for forming dimers — pairs of proteins that join together to form a functional enzyme.

“Tigist’s work has really shown categorically the importance of phosphorylation in controlling the flux through metabolic networks. It’s fundamental knowledge that emerges from this systemic study that she’s done, and it’s something that is not classically captured in the biochemistry textbooks,” White says.

Out of balance

To explore these effects in an animal model, the researchers compared two groups of mice, one that received a high-fat diet and one that consumed a normal diet. They found that overall, phosphorylation of metabolic enzymes led to a dysfunctional state in which cells were in redox imbalance, meaning that their cells were producing more reactive oxygen species than they could neutralize. These mice also became overweight and developed insulin resistance.

“In the context of continued high fat diet, what we see is a gradual drift away from redox homeostasis towards a more disease-like setting,” White says.

These effects were much more pronounced in male mice than female mice. Female mice were better able to compensate for the high fat diet by activating pathways involved in processing fat and metabolizing it for other uses, the researchers found.

“One of the things we learned is that the overall systemic effect of these phosphorylation events led to, especially in males, an increased imbalance in redox homeostasis. They were expressing a lot more stress and a lot more of the metabolic dysfunction phenotype compared to females,” Tamir says.

The researchers also found that if they gave mice who were on a high-fat diet an antioxidant called BHA, many of these effects were reversed. These mice showed a significant decrease in weight gain and did not become prediabetic, unlike the other mice fed a high-fat diet.

It appears that the antioxidant treatment leads cells back into a more balanced state, with fewer reactive oxygen species, the researchers say. Additionally, metabolic enzymes showed a systemic rewiring and changed state of phosphorylation in those mice.

“They’re experiencing a lot of metabolic dysfunction, but if you co-administer something that counters that, then they have enough reserve to maintain some sort of normalcy,” Tamir says. “The study suggests that there is something biochemically happening in cells to bring them to a different state — not a normal state, just a different state in which now, at the tissue and organism levels, the mice are healthier.”

In her new lab at the University of North Carolina, Tamir now plans to further explore whether antioxidant treatment could be an effective way to prevent or treat obesity-associated metabolic dysfunction, and what the optimal timing of such a treatment would be.

The research was funded in part by the Burroughs Wellcome Fund, the National Cancer Institute, the National Institutes of Health, the Ludwig Center at MIT, and the MIT Center for Precision Cancer Medicine.

A $20 million gift from the Leinweber Foundation, in addition to a $5 million commitment from the MIT School of Science, will support theoretical physics research and education at MIT.

Leinweber Foundation gifts to five institutions, totaling $90 million, will establish the newly renamed MIT Center for Theoretical Physics – A Leinweber Institute within the Department of Physics, affiliated with the Laboratory for Nuclear Science at the School of Science, as well as Leinweber Institutes for Theoretical Physics at three other top research universities: the University of Michigan, the University of California at Berkeley, and the University of Chicago, as well as a Leinweber Forum for Theoretical and Quantum Physics at the Institute for Advanced Study.

“MIT has one of the strongest and broadest theory groups in the world,” says Professor Washington Taylor, the director of the newly funded center and a leading researcher in string theory and its connection to observable particle physics and cosmology.

“This landmark endowment from the Leinweber Foundation will enable us to support the best graduate students and postdoctoral researchers to develop their own independent research programs and to connect with other researchers in the Leinweber Institute network. By pledging to support this network and fundamental curiosity-driven science, Larry Leinweber and his family foundation have made a huge contribution to maintaining a thriving scientific enterprise in the United States in perpetuity.”

The Leinweber Foundation’s investment across five institutions — constituting the largest philanthropic commitment ever for theoretical physics research, according to the Science Philanthropy Alliance, a nonprofit organization that supports philanthropic support for science — will strengthen existing programs at each institution and foster collaboration across the universities. Recipient institutions will work both independently and collaboratively to explore foundational questions in theoretical physics. Each institute will continue to shape its own research focus and programs, while also committing to big-picture cross-institutional convenings around topics of shared interest. Moreover, each institute will have significantly more funding for graduate students and postdocs, including fellowship support for three to eight fully endowed Leinweber Physics Fellows at each institute.

“This gift is a commitment to America’s scientific future,” says Larry Leinweber, founder and president of the Leinweber Foundation. “Theoretical physics may seem abstract to many, but it is the tip of the spear for innovation. It fuels our understanding of how the world works and opens the door to new technologies that can shape society for generations. As someone who has had a lifelong fascination with theoretical physics, I hope this investment not only strengthens U.S. leadership in basic science, but also inspires curiosity, creativity, and groundbreaking discoveries for generations to come.”

The gift to MIT will create a postdoc program that, once fully funded, will initially provide support for up to six postdocs, with two selected per year for a three-year program. In addition, the gift will provide student financial support, including fellowship support, for up to six graduate students per year studying theoretical physics. The goal is to attract the top talent to the MIT Center for Theoretical Physics – A Leinweber Institute and support the ongoing research programs in a more robust way.

A portion of the funding will also provide support for visitors, seminars, and other scholarly activities of current postdocs, faculty, and students in theoretical physics, as well as helping with administrative support.

“Graduate students are the heart of our country’s scientific research programs. Support for their education to become the future leaders of the field is essential for the advancement of the discipline,” says Nergis Mavalvala, dean of the MIT School of Science and the Curtis (1963) and Kathleen Marble Professor of Astrophysics.

The Leinweber Foundation gift is the second significant gift for the center. “We are always grateful to Virgil Elings, whose generous gift helped make possible the space that houses the center,” says Deepto Chakrabarty, head of the Department of Physics. Elings PhD ’66, co-founder of Digital Instruments, which designed and sold scanning probe microscopes, made his gift more than 20 years ago to support a space for theoretical physicists to collaborate.

“Gifts like those from Larry Leinweber and Virgil Elings are critical, especially now in this time of uncertain funding from the federal government for support of fundamental scientific research carried out by our nation’s leading postdocs, research scientists, faculty and students,” adds Mavalvala.

Professor Tracy Slatyer, whose work is motivated by questions of fundamental particle physics — particularly the nature and interactions of dark matter — will be the subsequent director of the MIT Center for Theoretical Physics – A Leinweber Institute beginning this fall. Slatyer will join Mavalvala, Taylor, Chakrabarty, and the entirety of the theoretical physics community for a dedication ceremony planned for the near future.

The Leinweber Foundation was founded in 2015 by software entrepreneur Larry Leinweber, and has worked with the Science Philanthropy Alliance since 2021 to shape its philanthropic strategy. “It’s been a true pleasure to work with Larry and the Leinweber family over the past four years and to see their vision take shape,” says France Córdova, president of the Science Philanthropy Alliance. “Throughout his life, Larry has exemplified curiosity, intellectual openness, and a deep commitment to learning. This gift reflects those values, ensuring that generations of scientists will have the freedom to explore, to question, and to pursue ideas that could change how we understand the universe.”

Batteries are nearing their limits in terms of how much power they can store for a given weight. That’s a serious obstacle for energy innovation and the search for new ways to power airplanes, trains, and ships. Now, researchers at MIT and elsewhere have come up with a solution that could help electrify these transportation systems.

Instead of a battery, the new concept is a kind of fuel cell — which is similar to a battery but can be quickly refueled rather than recharged. In this case, the fuel is liquid sodium metal, an inexpensive and widely available commodity. The other side of the cell is just ordinary air, which serves as a source of oxygen atoms. In between, a layer of solid ceramic material serves as the electrolyte, allowing sodium ions to pass freely through, and a porous air-facing electrode helps the sodium to chemically react with oxygen and produce electricity.

In a series of experiments with a prototype device, the researchers demonstrated that this cell could carry more than three times as much energy per unit of weight as the lithium-ion batteries used in virtually all electric vehicles today. Their findings are being published today in the journal Joule, in a paper by MIT doctoral students Karen Sugano, Sunil Mair, and Saahir Ganti-Agrawal; professor of materials science and engineering Yet-Ming Chiang; and five others.

“We expect people to think that this is a totally crazy idea,” says Chiang, who is the Kyocera Professor of Ceramics. “If they didn’t, I’d be a bit disappointed because if people don’t think something is totally crazy at first, it probably isn’t going to be that revolutionary.”

And this technology does appear to have the potential to be quite revolutionary, he suggests. In particular, for aviation, where weight is especially crucial, such an improvement in energy density could be the breakthrough that finally makes electrically powered flight practical at significant scale.

“The threshold that you really need for realistic electric aviation is about 1,000 watt-hours per kilogram,” Chiang says. Today’s electric vehicle lithium-ion batteries top out at about 300 watt-hours per kilogram — nowhere near what’s needed. Even at 1,000 watt-hours per kilogram, he says, that wouldn’t be enough to enable transcontinental or trans-Atlantic flights.

That’s still beyond reach for any known battery chemistry, but Chiang says that getting to 1,000 watts per kilogram would be an enabling technology for regional electric aviation, which accounts for about 80 percent of domestic flights and 30 percent of the emissions from aviation.

The technology could be an enabler for other sectors as well, including marine and rail transportation. “They all require very high energy density, and they all require low cost,” he says. “And that’s what attracted us to sodium metal.”

A great deal of research has gone into developing lithium-air or sodium-air batteries over the last three decades, but it has been hard to make them fully rechargeable. “People have been aware of the energy density you could get with metal-air batteries for a very long time, and it’s been hugely attractive, but it’s just never been realized in practice,” Chiang says.

By using the same basic electrochemical concept, only making it a fuel cell instead of a battery, the researchers were able to get the advantages of the high energy density in a practical form. Unlike a battery, whose materials are assembled once and sealed in a container, with a fuel cell the energy-carrying materials go in and out.

The team produced two different versions of a lab-scale prototype of the system. In one, called an H cell, two vertical glass tubes are connected by a tube across the middle, which contains a solid ceramic electrolyte material and a porous air electrode. Liquid sodium metal fills the tube on one side, and air flows through the other, providing the oxygen for the electrochemical reaction at the center, which ends up gradually consuming the sodium fuel. The other prototype uses a horizontal design, with a tray of the electrolyte material holding the liquid sodium fuel. The porous air electrode, which facilitates the reaction, is affixed to the bottom of the tray. 

Tests using an air stream with a carefully controlled humidity level produced a level of more than 1,500 watt-hours per kilogram at the level of an individual “stack,” which would translate to over 1,000 watt-hours at the full system level, Chiang says.

The researchers envision that to use this system in an aircraft, fuel packs containing stacks of cells, like racks of food trays in a cafeteria, would be inserted into the fuel cells; the sodium metal inside these packs gets chemically transformed as it provides the power. A stream of its chemical byproduct is given off, and in the case of aircraft this would be emitted out the back, not unlike the exhaust from a jet engine.

But there’s a very big difference: There would be no carbon dioxide emissions. Instead the emissions, consisting of sodium oxide, would actually soak up carbon dioxide from the atmosphere. This compound would quickly combine with moisture in the air to make sodium hydroxide — a material commonly used as a drain cleaner — which readily combines with carbon dioxide to form a solid material, sodium carbonate, which in turn forms sodium bicarbonate, otherwise known as baking soda.

“There’s this natural cascade of reactions that happens when you start with sodium metal,” Chiang says. “It’s all spontaneous. We don’t have to do anything to make it happen, we just have to fly the airplane.”

As an added benefit, if the final product, the sodium bicarbonate, ends up in the ocean, it could help to de-acidify the water, countering another of the damaging effects of greenhouse gases.

Using sodium hydroxide to capture carbon dioxide has been proposed as a way of mitigating carbon emissions, but on its own, it’s not an economic solution because the compound is too expensive. “But here, it’s a byproduct,” Chiang explains, so it’s essentially free, producing environmental benefits at no cost.

Importantly, the new fuel cell is inherently safer than many other batteries, he says. Sodium metal is extremely reactive and must be well-protected. As with lithium batteries, sodium can spontaneously ignite if exposed to moisture. “Whenever you have a very high energy density battery, safety is always a concern, because if there’s a rupture of the membrane that separates the two reactants, you can have a runaway reaction,” Chiang says. But in this fuel cell, one side is just air, “which is dilute and limited. So you don’t have two concentrated reactants right next to each other. If you’re pushing for really, really high energy density, you’d rather have a fuel cell than a battery for safety reasons.”

While the device so far exists only as a small, single-cell prototype, Chiang says the system should be quite straightforward to scale up to practical sizes for commercialization. Members of the research team have already formed a company, Propel Aero, to develop the technology. The company is currently housed in MIT’s startup incubator, The Engine.

Producing enough sodium metal to enable widespread, full-scale global implementation of this technology should be practical, since the material has been produced at large scale before. When leaded gasoline was the norm, before it was phased out, sodium metal was used to make the tetraethyl lead used as an additive, and it was being produced in the U.S. at a capacity of 200,000 tons a year. “It reminds us that sodium metal was once produced at large scale and safely handled and distributed around the U.S.,” Chiang says.

What’s more, sodium primarily originates from sodium chloride, or salt, so it is abundant, widely distributed around the world, and easily extracted, unlike lithium and other materials used in today’s EV batteries.

The system they envisage would use a refillable cartridge, which would be filled with liquid sodium metal and sealed. When it’s depleted, it would be returned to a refilling station and loaded with fresh sodium. Sodium melts at 98 degrees Celsius, just below the boiling point of water, so it is easy to heat to the melting point to refuel the cartridges.

Initially, the plan is to produce a brick-sized fuel cell that can deliver about 1,000 watt-hours of energy, enough to power a large drone, in order to prove the concept in a practical form that could be used for agriculture, for example. The team hopes to have such a demonstration ready within the next year.

Sugano, who conducted much of the experimental work as part of her doctoral thesis and will now work at the startup, says that a key insight was the importance of moisture in the process. As she tested the device with pure oxygen, and then with air, she found that the amount of humidity in the air was crucial to making the electrochemical reaction efficient. The humid air resulted in the sodium producing its discharge products in liquid rather than solid form, making it much easier for these to be removed by the flow of air through the system. “The key was that we can form this liquid discharge product and remove it easily, as opposed to the solid discharge that would form in dry conditions,” she says.

Ganti-Agrawal notes that the team drew from a variety of different engineering subfields. For example, there has been much research on high-temperature sodium, but none with a system with controlled humidity. “We’re pulling from fuel cell research in terms of designing our electrode, we’re pulling from older high-temperature battery research as well as some nascent sodium-air battery research, and kind of mushing it together,” which led to the “the big bump in performance” the team has achieved, he says.

The research team also included Alden Friesen, an MIT summer intern who attends Desert Mountain High School in Scottsdale, Arizona; Kailash Raman and William Woodford of Form Energy in Somerville, Massachusetts; Shashank Sripad of And Battery Aero in California, and Venkatasubramanian Viswanathan of the University of Michigan. The work was supported by ARPA-E, Breakthrough Energy Ventures, and the National Science Foundation, and used facilities at MIT.nano.

The human brain contains about 86 billion neurons. These cells fire electrical signals that help the brain store memories and send information and commands throughout the brain and the nervous system.

The brain also contains billions of astrocytes — star-shaped cells with many long extensions that allow them to interact with millions of neurons. Although they have long been thought to be mainly supportive cells, recent studies have suggested that astrocytes may play a role in memory storage and other cognitive functions.

MIT researchers have now put forth a new hypothesis for how astrocytes might contribute to memory storage. The architecture suggested by their model would help to explain the brain’s massive storage capacity, which is much greater than would be expected using neurons alone.

“Originally, astrocytes were believed to just clean up around neurons, but there’s no particular reason that evolution did not realize that, because each astrocyte can contact hundreds of thousands of synapses, they could also be used for computation,” says Jean-Jacques Slotine, an MIT professor of mechanical engineering and of brain and cognitive sciences, and an author of the new study.

Dmitry Krotov, a research staff member at the MIT-IBM Watson AI Lab and IBM Research, is the senior author of the open-access paper, which appeared May 23 in the Proceedings of the National Academy of Sciences. Leo Kozachkov PhD ’22 is the paper’s lead author.

Memory capacity

Astrocytes have a variety of support functions in the brain: They clean up debris, provide nutrients to neurons, and help to ensure an adequate blood supply.

Astrocytes also send out many thin tentacles, known as processes, which can each wrap around a single synapse — the junctions where two neurons interact with each other — to create a tripartite (three-part) synapse.

Within the past couple of years, neuroscientists have shown that if the connections between astrocytes and neurons in the hippocampus are disrupted, memory storage and retrieval are impaired.

Unlike neurons, astrocytes can’t fire action potentials, the electrical impulses that carry information throughout the brain. However, they can use calcium signaling to communicate with other astrocytes. Over the past few decades, as the resolution of calcium imaging has improved, researchers have found that calcium signaling also allows astrocytes to coordinate their activity with neurons in the synapses that they associate with.

These studies suggest that astrocytes can detect neural activity, which leads them to alter their own calcium levels. Those changes may trigger astrocytes to release gliotransmitters — signaling molecules similar to neurotransmitters — into the synapse.

“There’s a closed circle between neuron signaling and astrocyte-to-neuron signaling,” Kozachkov says. “The thing that is unknown is precisely what kind of computations the astrocytes can do with the information that they’re sensing from neurons.”

The MIT team set out to model what those connections might be doing and how they might contribute to memory storage. Their model is based on Hopfield networks — a type of neural network that can store and recall patterns.

Hopfield networks, originally developed by John Hopfield and Shun-Ichi Amari in the 1970s and 1980s, are often used to model the brain, but it has been shown that these networks can’t store enough information to account for the vast memory capacity of the human brain. A newer, modified version of a Hopfield network, known as dense associative memory, can store much more information through a higher order of couplings between more than two neurons.

However, it is unclear how the brain could implement these many-neuron couplings at a hypothetical synapse, since conventional synapses only connect two neurons: a presynaptic cell and a postsynaptic cell. This is where astrocytes come into play.

“If you have a network of neurons, which couple in pairs, there’s only a very small amount of information that you can encode in those networks,” Krotov says. “In order to build dense associative memories, you need to couple more than two neurons. Because a single astrocyte can connect to many neurons, and many synapses, it is tempting to hypothesize that there might exist an information transfer between synapses mediated by this biological cell. That was the biggest inspiration for us to look into astrocytes and led us to start thinking about how to build dense associative memories in biology.”

The neuron-astrocyte associative memory model that the researchers developed in their new paper can store significantly more information than a traditional Hopfield network — more than enough to account for the brain’s memory capacity.

Intricate connections

The extensive biological connections between neurons and astrocytes offer support for the idea that this type of model might explain how the brain’s memory storage systems work, the researchers say. They hypothesize that within astrocytes, memories are encoded by gradual changes in the patterns of calcium flow. This information is conveyed to neurons by gliotransmitters released at synapses that astrocyte processes connect to.

“By careful coordination of these two things — the spatial temporal pattern of calcium in the cell and then the signaling back to the neurons — you can get exactly the dynamics you need for this massively increased memory capacity,” Kozachkov says.

One of the key features of the new model is that it treats astrocytes as collections of processes, rather than a single entity. Each of those processes can be considered one computational unit. Because of the high information storage capabilities of dense associative memories, the ratio of the amount of information stored to the number of computational units is very high and grows with the size of the network. This makes the system not only high capacity, but also energy efficient.

“By conceptualizing tripartite synaptic domains — where astrocytes interact dynamically with pre- and postsynaptic neurons — as the brain’s fundamental computational units, the authors argue that each unit can store as many memory patterns as there are neurons in the network. This leads to the striking implication that, in principle, a neuron-astrocyte network could store an arbitrarily large number of patterns, limited only by its size,” says Maurizio De Pitta, an assistant professor of physiology at the Krembil Research Institute at the University of Toronto, who was not involved in the study.

To test whether this model might accurately represent how the brain stores memory, researchers could try to develop ways to precisely manipulate the connections between astrocytes’ processes, then observe how those manipulations affect memory function.

“We hope that one of the consequences of this work could be that experimentalists would consider this idea seriously and perform some experiments testing this hypothesis,” Krotov says.

In addition to offering insight into how the brain may store memory, this model could also provide guidance for researchers working on artificial intelligence. By varying the connectivity of the process-to-process network, researchers could generate a huge range of models that could be explored for different purposes, for instance, creating a continuum between dense associative memories and attention mechanisms in large language models.

“While neuroscience initially inspired key ideas in AI, the last 50 years of neuroscience research have had little influence on the field, and many modern AI algorithms have drifted away from neural analogies,” Slotine says. “In this sense, this work may be one of the first contributions to AI informed by recent neuroscience research.” 

MIT today launched its Initiative for New Manufacturing (INM), an Institute-wide effort to reinfuse U.S. industrial production with leading-edge technologies, bolster crucial U.S. economic sectors, and ignite job creation.

The initiative will encompass advanced research, innovative education programs, and partnership with companies across many sectors, in a bid to help transform manufacturing and elevate its impact.

“We want to work with firms big and small, in cities, small towns and everywhere in between, to help them adopt new approaches for increased productivity,” MIT President Sally A. Kornbluth wrote in a letter to the Institute community this morning. “We want to deliberately design high-quality, human-centered manufacturing jobs that bring new life to communities across the country.”

Kornbluth added: “Helping America build a future of new manufacturing is a perfect job for MIT — and I’m convinced that there is no more important work we can do to meet the moment and serve the nation now.”

The Initiative for New Manufacturing also announced its first six founding industry consortium members: Amgen, Flex, GE Vernova, PTC, Sanofi, and Siemens. Participants in the INM Industry Consortium will support seed projects proposed by MIT researchers, initially in the area of artificial intelligence for manufacturing.

INM joins the ranks of MIT’s other presidential initiatives — including The Climate Project at MIT; MITHIC, which supports the human-centered disciplines; MIT HEALS, centered on the life sciences and health; and MGAIC, the MIT Generative AI Impact Consortium.

“There is tremendous opportunity to bring together a vibrant community working across every scale — from nanotechnology to large-scale manufacturing — and across a wide-range of applications including semiconductors, medical devices, automotive, energy systems, and biotechnology,” says Anantha Chandrakasan, MIT’s chief innovation and strategy officer and dean of engineering, who is part of the initiative’s leadership team. “MIT is uniquely positioned to harness the transformative power of digital tools and AI to shape future of manufacturing. I’m truly excited about what we can build together and the synergies this creates with other cross-cutting initiatives across the Institute.”

The initiative is just the latest MIT-centered effort in recent decades aiming to expand American manufacturing. A faculty research group wrote the 1989 bestseller “Made in America: Regaining the Productive Edge,” advocating for a renewal of manufacturing; another MIT project, called Production in the Innovation Economy, called for expanded manufacturing in the early 2010s. In 2016, MIT also founded The Engine, a venture fund investing in hardware-based “tough tech” start-ups including many with potential to became substantial manufacturing firms.

As developed, the MIT Initiative for New Manufacturing is based around four major themes:

• Reimagining manufacturing technologies and systems: realizing breakthrough technologies and system-level approaches to advance energy production, health care, computing, transportation, consumer products, and more;
• Elevating the productivity and experience of manufacturing: developing and deploying new digitally driven methods and tools to amplify productivity and improve the human experience of manufacturing;
• Scaling new manufacturing: accelerating the scaling of manufacturing companies and transforming supply chains to maximize efficiency and resilience, fostering product innovation and business growth; and
• Transforming the manufacturing base: driving the deployment of a sustainable global manufacturing ecosystem that provides compelling opportunities to workers, with major efforts focused on the U.S.

The initiative has mapped out many concrete activities and programs, which will include an Institute-wide research program on emerging technologies and other major topics; workforce and education programs; and industry engagement and participation. INM also aims to establish new labs for developing manufacturing tools and techniques; a “factory observatory” program which immerses students in manufacturing through visits to production sites; and key “pillars” focusing on areas from semiconductors and biomanufacturing to defense and aviation.

The workforce and education element of INM will include TechAMP, an MIT-created program that works with community colleges to bridge the gap between technicians and engineers; AI-driven teaching tools; professional education; and an effort to expand manufacturing education on campus in collaboration with MIT departments and degree programs.

INM’s leadership team has three faculty co-directors: John Hart, the Class of 1922 Professor and head of the Department of Mechanical Engineering; Suzanne Berger, Institute Professor at MIT and a political scientist who has conducted influential empirical studies of manufacturing; and Chris Love, the Raymond A. and Helen E. St. Laurent Professor of Chemical Engineering. The initiative’s executive director is Julie Diop.

The initiative is in the process of forming a faculty steering committee with representation from across the Institute, as well as an external advisory board. INM stems partly from the work of the Manufacturing@MIT working group, formed in 2022 to assess many of these issues.

The launch of the new initiative was previewed at a daylong MIT symposium on May 7, titled “A Vision for New Manufacturing.” The event, held before a capacity audience in MIT’s Wong Auditorium, featured over 30 speakers from a wide range of manufacturing sectors.

“The rationale for growing and transforming U.S. manufacturing has never been more urgent than it is today,” Berger said at the event. “What we are trying to build at MIT now is not just another research project. … Together, with people in this room and outside this room, we’re trying to change what’s happening in our country.”

“We need to think about the importance of manufacturing again, because it is what brings product ideas to people,” Love told MIT News. “For instance, in biotechnology, new life-saving medicines can’t reach patients without manufacturing. There is a real urgency about this issue for both economic prosperity and creating jobs. We have seen the impact for our country when we have lost our lead in manufacturing in some sectors. Biotechnology, where the U.S. has been the global leader for more than 40 years, offers the potential to promote new robust economies here, but we need to advance our capabilities in biomanufacturing to maintain our advantage in this area.”

Hart adds: “While manufacturing feels very timely today, it is of enduring importance. Manufactured products enable our daily lives and manufacturing is critical to advancing the frontiers of technology and society. Our efforts leading up to launch of the initiative revealed great excitement about manufacturing across MIT, especially from students. Working with industry — from small to large companies, and from young startups to industrial giants — will be instrumental to creating impact and realizing the vision for new manufacturing.”

In her letter to the MIT community today, Kornbluth stressed that the initiative’s goal is to drive transformation by making manufacturing more productive, resilient, and sustainable.

“We want to reimagine manufacturing technologies and systems to advance fields like energy production, health care, computing, transportation, consumer products, and more,” she wrote. “And we want to reach well beyond the shop floor to tackle challenges like how to make supply chains more resilient, and how to inform public policy to foster a broad, healthy manufacturing ecosystem that can drive decades of innovation and growth.”

Editor’s note: A seventh founding member, Autodesk, was announced on May 30.

Where did the moon’s magnetism go? Scientists have puzzled over this question for decades, ever since orbiting spacecraft picked up signs of a high magnetic field in lunar surface rocks. The moon itself has no inherent magnetism today. 

Now, MIT scientists may have solved the mystery. They propose that a combination of an ancient, weak magnetic field and a large, plasma-generating impact may have temporarily created a strong magnetic field, concentrated on the far side of the moon.

In a study appearing today in the journal Science Advances, the researchers show through detailed simulations that an impact, such as from a large asteroid, could have generated a cloud of ionized particles that briefly enveloped the moon. This plasma would have streamed around the moon and concentrated at the opposite location from the initial impact. There, the plasma would have interacted with and momentarily amplified the moon’s weak magnetic field. Any rocks in the region could have recorded signs of the heightened magnetism before the field quickly died away.

This combination of events could explain the presence of highly magnetic rocks detected in a region near the south pole, on the moon’s far side. As it happens, one of the largest impact basins — the Imbrium basin — is located in the exact opposite spot on the near side of the moon. The researchers suspect that whatever made that impact likely released the cloud of plasma that kicked off the scenario in their simulations.

“There are large parts of lunar magnetism that are still unexplained,” says lead author Isaac Narrett, a graduate student in the MIT Department of Earth, Atmospheric and Planetary Sciences (EAPS). “But the majority of the strong magnetic fields that are measured by orbiting spacecraft can be explained by this process — especially on the far side of the moon.”

Narrett’s co-authors include Rona Oran and Benjamin Weiss at MIT, along with Katarina Miljkovic at Curtin University, Yuxi Chen and Gábor Tóth at the University of Michigan at Ann Arbor, and Elias Mansbach PhD ’24 at Cambridge University. Nuno Loureiro, professor of nuclear science and engineering at MIT, also contributed insights and advice.

Beyond the sun

Scientists have known for decades that the moon holds remnants of a strong magnetic field. Samples from the surface of the moon, returned by astronauts on NASA’s Apollo missions of the 1960s and 70s, as well as global measurements of the moon taken remotely by orbiting spacecraft, show signs of remnant magnetism in surface rocks, especially on the far side of the moon.

The typical explanation for surface magnetism is a global magnetic field, generated by an internal “dynamo,” or a core of molten, churning material. The Earth today generates a magnetic field through a dynamo process, and it’s thought that the moon once may have done the same, though its much smaller core would have produced a much weaker magnetic field that may not explain the highly magnetized rocks observed, particularly on the moon’s far side.

An alternative hypothesis that scientists have tested from time to time involves a giant impact that generated plasma, which in turn amplified any weak magnetic field. In 2020, Oran and Weiss tested this hypothesis with simulations of a giant impact on the moon, in combination with the solar-generated magnetic field, which is weak as it stretches out to the Earth and moon.

In simulations, they tested whether an impact to the moon could amplify such a solar field, enough to explain the highly magnetic measurements of surface rocks. It turned out that it wasn’t, and their results seemed to rule out plasma-induced impacts as playing a role in the moon’s missing magnetism.

A spike and a jitter

But in their new study, the researchers took a different tack. Instead of accounting for the sun’s magnetic field, they assumed that the moon once hosted a dynamo that produced a magnetic field of its own, albeit a weak one. Given the size of its core, they estimated that such a field would have been about 1 microtesla, or 50 times weaker than the Earth’s field today.

From this starting point, the researchers simulated a large impact to the moon’s surface, similar to what would have created the Imbrium basin, on the moon’s near side. Using impact simulations from Katarina Miljkovic, the team then simulated the cloud of plasma that such an impact would have generated as the force of the impact vaporized the surface material. They adapted a second code, developed by collaborators at the University of Michigan, to simulate how the resulting plasma would flow and interact with the moon’s weak magnetic field.

These simulations showed that as a plasma cloud arose from the impact, some of it would have expanded into space, while the rest would stream around the moon and concentrate on the opposite side. There, the plasma would have compressed and briefly amplified the moon’s weak magnetic field. This entire process, from the moment the magnetic field was amplified to the time that it decays back to baseline, would have been incredibly fast — somewhere around 40 minutes, Narrett says.

Would this brief window have been enough for surrounding rocks to record the momentary magnetic spike? The researchers say, yes, with some help from another, impact-related effect.

They found that an Imbrium-scale impact would have sent a pressure wave through the moon, similar to a seismic shock. These waves would have converged to the other side, where the shock would have “jittered” the surrounding rocks, briefly unsettling the rocks’ electrons — the subatomic particles that naturally orient their spins to any external magnetic field. The researchers suspect the rocks were shocked just as the impact’s plasma amplified the moon’s magnetic field. As the rocks’ electrons settled back, they assumed a new orientation, in line with the momentary high magnetic field.

“It’s as if you throw a 52-card deck in the air, in a magnetic field, and each card has a compass needle,” Weiss says. “When the cards settle back to the ground, they do so in a new orientation. That’s essentially the magnetization process.”

The researchers say this combination of a dynamo plus a large impact, coupled with the impact’s shockwave, is enough to explain the moon’s highly magnetized surface rocks — particularly on the far side. One way to know for sure is to directly sample the rocks for signs of shock, and high magnetism. This could be a possibility, as the rocks lie on the far side, near the lunar south pole, where missions such as NASA’s Artemis program plan to explore.

“For several decades, there’s been sort of a conundrum over the moon’s magnetism — is it from impacts or is it from a dynamo?” Oran says. “And here we’re saying, it’s a little bit of both. And it’s a testable hypothesis, which is nice.”

The team’s simulations were carried out using the MIT SuperCloud. This research was supported, in part, by NASA. 

A team of international astronomers led by Richard Teague, the Kerr-McGee Career Development Professor in the Department of Earth, Atmospheric and Planetary Sciences (EAPS) has gathered the most sensitive and detailed observations of 15 protoplanetary disks to date, giving the astronomy community a new look at the mechanisms of early planetary formation.

“The new approaches we’ve developed to gather this data and images are like switching from reading glasses to high-powered binoculars — they reveal a whole new level of detail in these planet-forming systems,” says Teague.

Their open-access findings were published in a special collection of 17 papers in the Astrophysical Journal of Letters, with several more coming out this summer. The report sheds light on a breadth of questions, including ways to calculate the mass of a disk by measuring its gravitational influence and extracting rotational velocity profiles to a precision of meters per second.

Protoplanetary disks are a collection of dust and gas around young stars, from which planets form. Observing the dust in these disks is easier because it is brighter, but the information that can be gleaned from dust alone is only a snapshot of what is going on. Teague’s research focus has shifted attention to the gas in these systems, as they can tell us more about the dynamics in a disk, including properties such as gravity, velocity, and mass.

To achieve the resolution necessary to study gas, the exoALMA program spent five years coordinating longer observation windows on the Atacama Large Millimeter/submillimeter Array (ALMA) in Chile. As a result, the international team of astronomers, many of whom are early-career scientists, were able to collect some of the most detailed images ever taken of protoplanetary disks.

“The impressive thing about the data is that it’s so good, the community is developing new tools to extract signatures from planets,” says Marcelo Barraza-Alfaro, a postdoc in the Planet Formation Lab and a member of the exoALMA project. Several new techniques to improve and calibrate the images taken were developed to maximize the higher resolution and sensitivity that was used.

As a result, “we are seeing new things that require us to modify our understanding of what’s going on in protoplanetary disks,” he says.

One of the papers with the largest EAPS influence explores planetary formation through vortices. It has been known for some time that the simple model of formation often proposed, where dust grains clump together and “snowball” into a planetary core, is not enough. One possible way to help is through vortices, or localized perturbations in the gas that pull dust into the center. Here, they are more likely to clump, the way soap bubbles collect in a draining tub.

“We can see the concentration of dust in different regions, but we cannot see how it is moving,” says Lisa Wölfer, another postdoc in the Planet Formation Lab at MIT and first author on the paper. While astronomers can see that the dust has gathered, there isn’t enough information to rule out how it got to that point.

“Only through the dynamics in the gas can we actually confirm that it’s a vortex, and not something else, creating the structure,” she says.

During the data collection period, Teague, Wölfer, and Barraza-Alfaro developed simple models of protoplanetary disks to compare to their observations. When they got the data back, however, the models couldn’t explain what they were seeing.

“We saw the data and nothing worked anymore. It was way too complicated,” says Teague. “Before, everyone thought they were not dynamic. That’s completely not the case.”

The team was forced to reevaluate their models and work with more complex ones incorporating more motion in the gas, which take more time and resources to run. But early results look promising.

“We see that the patterns look very similar; we think this is the best test case to study further with more observations,” says Wölfer.

The new data, which have been made public, come at a fortuitous time: ALMA will be going dark for a period in the next few years while it undergoes upgrades. During this time, astronomers can continue the monumental process of sifting through all the data.

“It’s going to just keep on producing results for years and years to come,” says Teague.

Separating crude oil into products such as gasoline, diesel, and heating oil is an energy-intensive process that accounts for about 6 percent of the world’s CO₂ emissions. Most of that energy goes into the heat needed to separate the components by their boiling point.

In an advance that could dramatically reduce the amount of energy needed for crude oil fractionation, MIT engineers have developed a membrane that filters the components of crude oil by their molecular size.

“This is a whole new way of envisioning a separation process. Instead of boiling mixtures to purify them, why not separate components based on shape and size? The key innovation is that the filters we developed can separate very small molecules at an atomistic length scale,” says Zachary P. Smith, an associate professor of chemical engineering at MIT and the senior author of the new study.

The new filtration membrane can efficiently separate heavy and light components from oil, and it is resistant to the swelling that tends to occur with other types of oil separation membranes. The membrane is a thin film that can be manufactured using a technique that is already widely used in industrial processes, potentially allowing it to be scaled up for widespread use.

Taehoon Lee, a former MIT postdoc who is now an assistant professor at Sungkyunkwan University in South Korea, is the lead author of the paper, which appears today in Science.

Oil fractionation

Conventional heat-driven processes for fractionating crude oil make up about 1 percent of global energy use, and it has been estimated that using membranes for crude oil separation could reduce the amount of energy needed by about 90 percent. For this to succeed, a separation membrane needs to allow hydrocarbons to pass through quickly, and to selectively filter compounds of different sizes.

Until now, most efforts to develop a filtration membrane for hydrocarbons have focused on polymers of intrinsic microporosity (PIMs), including one known as PIM-1. Although this porous material allows the fast transport of hydrocarbons, it tends to excessively absorb some of the organic compounds as they pass through the membrane, leading the film to swell, which impairs its size-sieving ability.

To come up with a better alternative, the MIT team decided to try modifying polymers that are used for reverse osmosis water desalination. Since their adoption in the 1970s, reverse osmosis membranes have reduced the energy consumption of desalination by about 90 percent — a remarkable industrial success story.

The most commonly used membrane for water desalination is a polyamide that is manufactured using a method known as interfacial polymerization. During this process, a thin polymer film forms at the interface between water and an organic solvent such as hexane. Water and hexane do not normally mix, but at the interface between them, a small amount of the compounds dissolved in them can react with each other.

In this case, a hydrophilic monomer called MPD, which is dissolved in water, reacts with a hydrophobic monomer called TMC, which is dissolved in hexane. The two monomers are joined together by a connection known as an amide bond, forming a polyamide thin film (named MPD-TMC) at the water-hexane interface.

While highly effective for water desalination, MPD-TMC doesn’t have the right pore sizes and swelling resistance that would allow it to separate hydrocarbons.

To adapt the material to separate the hydrocarbons found in crude oil, the researchers first modified the film by changing the bond that connects the monomers from an amide bond to an imine bond. This bond is more rigid and hydrophobic, which allows hydrocarbons to quickly move through the membrane without causing noticeable swelling of the film compared to the polyamide counterpart.

“The polyimine material has porosity that forms at the interface, and because of the cross-linking chemistry that we have added in, you now have something that doesn’t swell,” Smith says. “You make it in the oil phase, react it at the water interface, and with the crosslinks, it’s now immobilized. And so those pores, even when they’re exposed to hydrocarbons, no longer swell like other materials.”

The researchers also introduced a monomer called triptycene. This shape-persistent, molecularly selective molecule further helps the resultant polyimines to form pores that are the right size for hydrocarbons to fit through.

This approach represents “an important step toward reducing industrial energy consumption,” says Andrew Livingston, a professor of chemical engineering at Queen Mary University of London, who was not involved in the study.

“This work takes the workhorse technology of the membrane desalination industry, interfacial polymerization, and creates a new way to apply it to organic systems such as hydrocarbon feedstocks, which currently consume large chunks of global energy,” Livingston says. “The imaginative approach using an interfacial catalyst coupled to hydrophobic monomers leads to membranes with high permeance and excellent selectivity, and the work shows how these can be used in relevant separations.”

Efficient separation

When the researchers used the new membrane to filter a mixture of toluene and triisopropylbenzene (TIPB) as a benchmark for evaluating separation performance, it was able to achieve a concentration of toluene 20 times greater than its concentration in the original mixture. They also tested the membrane with an industrially relevant mixture consisting of naphtha, kerosene, and diesel, and found that it could efficiently separate the heavier and lighter compounds by their molecular size.

If adapted for industrial use, a series of these filters could be used to generate a higher concentration of the desired products at each step, the researchers say.

“You can imagine that with a membrane like this, you could have an initial stage that replaces a crude oil fractionation column. You could partition heavy and light molecules and then you could use different membranes in a cascade to purify complex mixtures to isolate the chemicals that you need,” Smith says.

Interfacial polymerization is already widely used to create membranes for water desalination, and the researchers believe it should be possible to adapt those processes to mass produce the films they designed in this study.

“The main advantage of interfacial polymerization is it’s already a well-established method to prepare membranes for water purification, so you can imagine just adopting these chemistries into existing scale of manufacturing lines,” Lee says.

The research was funded, in part, by ExxonMobil through the MIT Energy Initiative. 

Magnets and superconductors go together like oil and water — or so scientists have thought. But a new finding by MIT physicists is challenging this century-old assumption.

In a paper appearing today in the journal Nature, the physicists report that they have discovered a “chiral superconductor” — a material that conducts electricity without resistance, and also, paradoxically, is intrinsically magnetic. What’s more, they observed this exotic superconductivity in a surprisingly ordinary material: graphite, the primary material in pencil lead.

Graphite is made from many layers of graphene — atomically thin, lattice-like sheets of carbon atoms — that are stacked together and can easily flake off when pressure is applied, as when pressing down to write on a piece of paper. A single flake of graphite can contain several million sheets of graphene, which are normally stacked such that every other layer aligns. But every so often, graphite contains tiny pockets where graphene is stacked in a different pattern, resembling a staircase of offset layers.

The MIT team has found that when four or five sheets of graphene are stacked in this “rhombohedral” configuration, the resulting structure can exhibit exceptional electronic properties that are not seen in graphite as a whole.

In their new study, the physicists isolated microscopic flakes of rhombohedral graphene from graphite, and subjected the flakes to a battery of electrical tests. They found that when the flakes are cooled to 300 millikelvins (about -273 degrees Celsius), the material turns into a superconductor, meaning that any electrical current passing through the material can flow through without resistance.

They also found that when they swept an external magnetic field up and down, the flakes could be switched between two different superconducting states, just like a magnet. This suggests that the superconductor has some internal, intrinsic magnetism. Such switching behavior is absent in other superconductors.

“The general lore is that superconductors do not like magnetic fields,” says Long Ju, assistant professor of physics at MIT. “But we believe this is the first observation of a superconductor that behaves as a magnet with such direct and simple evidence. And that’s quite a bizarre thing because it is against people’s general impression on superconductivity and magnetism.”

Ju is senior author of the study, which includes MIT co-authors Tonghang Han, Zhengguang Lu, Zach Hadjri, Lihan Shi, Zhenghan Wu, Wei Xu, Yuxuan Yao, Jixiang Yang, Junseok Seo, Shenyong Ye, Muyang Zhou, and Liang Fu, along with collaborators from Florida State University, the University of Basel in Switzerland, and the National Institute for Materials Science in Japan.

Graphene twist

In everyday conductive materials, electrons flow through in a chaotic scramble, whizzing by each other, and pinging off the material’s atomic latticework. Each time an electron scatters off an atom, it has, in essence, met some resistance, and loses some energy as a result, normally in the form of heat. In contrast, when certain materials are cooled to ultracold temperatures, they can become superconducting, meaning that the material can allow electrons to pair up, in what physicists term “Cooper pairs.” Rather than scattering away, these electron pairs glide through a material without resistance. With a superconductor, then, no energy is lost in translation.

Since superconductivity was first observed in 1911, physicists have shown many times over that zero electrical resistance is a hallmark of a superconductor. Another defining property was first observed in 1933, when the physicist Walther Meissner discovered that a superconductor will expel an external magnetic field. This “Meissner effect” is due in part to a superconductor’s electron pairs, which collectively act to push away any magnetic field.

Physicists have assumed that all superconducting materials should exhibit both zero electrical resistance, and a natural magnetic repulsion. Indeed, these two properties are what could enable Maglev, or “magnetic levitation” trains, whereby a superconducting rail repels and therefore levitates a magnetized car.

Ju and his colleagues had no reason to question this assumption as they carried out their experiments at MIT. In the last few years, the team has been exploring the electrical properties of pentalayer rhombohedral graphene. The researchers have observed surprising properties in the five-layer, staircase-like graphene structure, most recently that it enables electrons to split into fractions of themselves. This phenomenon occurs when the pentalayer structure is placed atop a sheet of hexagonal boron nitride (a material similar to graphene), and slightly offset by a specific angle, or twist. 

Curious as to how electron fractions might change with changing conditions, the researchers followed up their initial discovery with similar tests, this time by misaligning the graphene and hexagonal boron nitride structures. To their surprise, they found that when they misaligned the two materials and sent an electrical current through, at temperatures less than 300 millikelvins, they measured zero resistance. It seemed that the phenomenon of electron fractions disappeared, and what emerged instead was superconductivity.

The researchers went a step further to see how this new superconducting state would respond to an external magnetic field. They applied a magnet to the material, along with a voltage, and measured the electrical current coming out of the material. As they dialed the magnetic field from negative to positive (similar to a north and south polarity) and back again, they observed that the material maintained its superconducting, zero-resistance state, except in two instances, once at either magnetic polarity. In these instances, the resistance briefly spiked, before switching back to zero, and returning to a superconducting state.

“If this were a conventional superconductor, it would just remain at zero resistance, until the magnetic field reaches a critical point, where superconductivity would be killed,” Zach Hadjri, a first-year student in the group, says. “Instead, this material seems to switch between two superconducting states, like a magnet that starts out pointing upward, and can flip downwards when you apply a magnetic field. So it looks like this is a superconductor that also acts like a magnet. Which doesn’t make any sense!”

“One of a kind”

As counterintuitive as the discovery may seem, the team observed the same phenomenon in six similar samples. They suspect that the unique configuration of rhombohedral graphene is the key. The material has a very simple arrangement of carbon atoms. When cooled to ultracold temperatures, the thermal fluctuation is minimized, allowing any electrons flowing through the material to slow down, sense each other, and interact.

Such quantum interactions can lead electrons to pair up and superconduct. These interactions can also encourage electrons to coordinate. Namely, electrons can collectively occupy one of two opposite momentum states, or “valleys.” When all electrons are in one valley, they effectively spin in one direction, versus the opposite direction. In conventional superconductors, electrons can occupy either valley, and any pair of electrons is typically made from electrons of opposite valleys that cancel each other out. The pair overall then, has zero momentum, and does not spin.

In the team’s material structure, however, they suspect that all electrons interact such that they share the same valley, or momentum state. When electrons then pair up, the superconducting pair overall has a “non-zero” momentum, and spinning, that, along with many other pairs, can amount to an internal, superconducting magnetism.

“You can think of the two electrons in a pair spinning clockwise, or counterclockwise, which corresponds to a magnet pointing up, or down,” Tonghang Han, a fifth-year student in the group, explains. “So we think this is the first observation of a superconductor that behaves as a magnet due to the electrons’ orbital motion, which is known as a chiral superconductor. It’s one of a kind. It is also a candidate for a topological superconductor which could enable robust quantum computation.”

“Everything we’ve discovered in this material has been completely out of the blue,” says Zhengguang Lu, a former postdoc in the group and now an assistant professor at Florida State University. “But because this is a simple system, we think we have a good chance of understanding what is going on, and could demonstrate some very profound and deep physics principles.”

“It is truly remarkable that such an exotic chiral superconductor emerges from such simple ingredients,” adds Liang Fu, professor of physics at MIT. “Superconductivity in rhombodedral graphene will surely have a lot to offer.”     

The part of the research carried out at MIT was supported by the U.S. Department of Energy and a MathWorks Fellowship. This research was carried out, in part, using facilities at MIT.nano.

On Dec. 21, 2022, just as peak holiday season travel was getting underway, Southwest Airlines went through a cascading series of failures in their scheduling, initially triggered by severe winter weather in the Denver area. But the problems spread through their network, and over the course of the next 10 days the crisis ended up stranding over 2 million passengers and causing losses of $750 million for the airline.

How did a localized weather system end up triggering such a widespread failure? Researchers at MIT have examined this widely reported failure as an example of cases where systems that work smoothly most of the time suddenly break down and cause a domino effect of failures. They have now developed a computational system for using the combination of sparse data about a rare failure event, in combination with much more extensive data on normal operations, to work backwards and try to pinpoint the root causes of the failure, and hopefully be able to find ways to adjust the systems to prevent such failures in the future.

The findings were presented at the International Conference on Learning Representations (ICLR), which was held in Singapore from April 24-28 by MIT doctoral student Charles Dawson, professor of aeronautics and astronautics Chuchu Fan, and colleagues from Harvard University and the University of Michigan.

“The motivation behind this work is that it’s really frustrating when we have to interact with these complicated systems, where it’s really hard to understand what’s going on behind the scenes that’s creating these issues or failures that we’re observing,” says Dawson.

The new work builds on previous research from Fan’s lab, where they looked at problems involving hypothetical failure prediction problems, she says, such as with groups of robots working together on a task, or complex systems such as the power grid, looking for ways to predict how such systems may fail. “The goal of this project,” Fan says, “was really to turn that into a diagnostic tool that we could use on real-world systems.”

The idea was to provide a way that someone could “give us data from a time when this real-world system had an issue or a failure,” Dawson says, “and we can try to diagnose the root causes, and provide a little bit of a look behind the curtain at this complexity.”

The intent is for the methods they developed “to work for a pretty general class of cyber-physical problems,” he says. These are problems in which “you have an automated decision-making component interacting with the messiness of the real world,” he explains. There are available tools for testing software systems that operate on their own, but the complexity arises when that software has to interact with physical entities going about their activities in a real physical setting, whether it be the scheduling of aircraft, the movements of autonomous vehicles, the interactions of a team of robots, or the control of the inputs and outputs on an electric grid. In such systems, what often happens, he says, is that “the software might make a decision that looks OK at first, but then it has all these domino, knock-on effects that make things messier and much more uncertain.”

One key difference, though, is that in systems like teams of robots, unlike the scheduling of airplanes, “we have access to a model in the robotics world,” says Fan, who is a principal investigator in MIT’s Laboratory for Information and Decision Systems (LIDS). “We do have some good understanding of the physics behind the robotics, and we do have ways of creating a model” that represents their activities with reasonable accuracy. But airline scheduling involves processes and systems that are proprietary business information, and so the researchers had to find ways to infer what was behind the decisions, using only the relatively sparse publicly available information, which essentially consisted of just the actual arrival and departure times of each plane.

“We have grabbed all this flight data, but there is this entire system of the scheduling system behind it, and we don’t know how the system is working,” Fan says. And the amount of data relating to the actual failure is just several day’s worth, compared to years of data on normal flight operations.

The impact of the weather events in Denver during the week of Southwest’s scheduling crisis clearly showed up in the flight data, just from the longer-than-normal turnaround times between landing and takeoff at the Denver airport. But the way that impact cascaded though the system was less obvious, and required more analysis. The key turned out to have to do with the concept of reserve aircraft.

Airlines typically keep some planes in reserve at various airports, so that if problems are found with one plane that is scheduled for a flight, another plane can be quickly substituted. Southwest uses only a single type of plane, so they are all interchangeable, making such substitutions easier. But most airlines operate on a hub-and-spoke system, with a few designated hub airports where most of those reserve aircraft may be kept, whereas Southwest does not use hubs, so their reserve planes are more scattered throughout their network. And the way those planes were deployed turned out to play a major role in the unfolding crisis.

“The challenge is that there’s no public data available in terms of where the aircraft are stationed throughout the Southwest network,” Dawson says. “What we’re able to find using our method is, by looking at the public data on arrivals, departures, and delays, we can use our method to back out what the hidden parameters of those aircraft reserves could have been, to explain the observations that we were seeing.”

What they found was that the way the reserves were deployed was a “leading indicator” of the problems that cascaded in a nationwide crisis. Some parts of the network that were affected directly by the weather were able to recover quickly and get back on schedule. “But when we looked at other areas in the network, we saw that these reserves were just not available, and things just kept getting worse.”

For example, the data showed that Denver’s reserves were rapidly dwindling because of the weather delays, but then “it also allowed us to trace this failure from Denver to Las Vegas,” he says. While there was no severe weather there, “our method was still showing us a steady decline in the number of aircraft that were able to serve flights out of Las Vegas.”

He says that “what we found was that there were these circulations of aircraft within the Southwest network, where an aircraft might start the day in California and then fly to Denver, and then end the day in Las Vegas.” What happened in the case of this storm was that the cycle got interrupted. As a result, “this one storm in Denver breaks the cycle, and suddenly the reserves in Las Vegas, which is not affected by the weather, start to deteriorate.”

In the end, Southwest was forced to take a drastic measure to resolve the problem: They had to do a “hard reset” of their entire system, canceling all flights and flying empty aircraft around the country to rebalance their reserves.

Working with experts in air transportation systems, the researchers developed a model of how the scheduling system is supposed to work. Then, “what our method does is, we’re essentially trying to run the model backwards.” Looking at the observed outcomes, the model allows them to work back to see what kinds of initial conditions could have produced those outcomes.

While the data on the actual failures were sparse, the extensive data on typical operations helped in teaching the computational model “what is feasible, what is possible, what’s the realm of physical possibility here,” Dawson says. “That gives us the domain knowledge to then say, in this extreme event, given the space of what’s possible, what’s the most likely explanation” for the failure.

This could lead to a real-time monitoring system, he says, where data on normal operations are constantly compared to the current data, and determining what the trend looks like. “Are we trending toward normal, or are we trending toward extreme events?” Seeing signs of impending issues could allow for preemptive measures, such as redeploying reserve aircraft in advance to areas of anticipated problems.

Work on developing such systems is ongoing in her lab, Fan says. In the meantime, they have produced an open-source tool for analyzing failure systems, called CalNF, which is available for anyone to use. Meanwhile Dawson, who earned his doctorate last year, is working as a postdoc to apply the methods developed in this work to understanding failures in power networks.

The research team also included Max Li from the University of Michigan and Van Tran from Harvard University. The work was supported by NASA, the Air Force Office of Scientific Research, and the MIT-DSTA program.

Global warming will likely hinder our future ability to control ground-level ozone, a harmful air pollutant that is a primary component of smog, according to a new MIT study.

The results could help scientists and policymakers develop more effective strategies for improving both air quality and human health. Ground-level ozone causes a host of detrimental health impacts, from asthma to heart disease, and contributes to thousands of premature deaths each year.

The researchers’ modeling approach reveals that, as the Earth warms due to climate change, ground-level ozone will become less sensitive to reductions in nitrogen oxide emissions in eastern North America and Western Europe. In other words, it will take greater nitrogen oxide emission reductions to get the same air quality benefits.

However, the study also shows that the opposite would be true in northeast Asia, where cutting emissions would have a greater impact on reducing ground-level ozone in the future. 

The researchers combined a climate model that simulates meteorological factors, such as temperature and wind speeds, with a chemical transport model that estimates the movement and composition of chemicals in the atmosphere.

By generating a range of possible future outcomes, the researchers’ ensemble approach better captures inherent climate variability, allowing them to paint a fuller picture than many previous studies.

“Future air quality planning should consider how climate change affects the chemistry of air pollution. We may need steeper cuts in nitrogen oxide emissions to achieve the same air quality goals,” says Emmie Le Roy, a graduate student in the MIT Department of Earth, Atmospheric and Planetary Sciences (EAPS) and lead author of a paper on this study.

Her co-authors include Anthony Y.H. Wong, a postdoc in the MIT Center for Sustainability Science and Strategy; Sebastian D. Eastham, principal research scientist in the MIT Center for Sustainability Science and Strategy; Arlene Fiore, the Peter H. Stone and Paola Malanotte Stone Professor of EAPS; and senior author Noelle Selin, a professor in the Institute for Data, Systems, and Society (IDSS) and EAPS. The research appears today in Environmental Science and Technology.

Controlling ozone

Ground-level ozone differs from the stratospheric ozone layer that protects the Earth from harmful UV radiation. It is a respiratory irritant that is harmful to the health of humans, animals, and plants.

Controlling ground-level ozone is particularly challenging because it is a secondary pollutant, formed in the atmosphere by complex reactions involving nitrogen oxides and volatile organic compounds in the presence of sunlight.

“That is why you tend to have higher ozone days when it is warm and sunny,” Le Roy explains.

Regulators typically try to reduce ground-level ozone by cutting nitrogen oxide emissions from industrial processes. But it is difficult to predict the effects of those policies because ground-level ozone interacts with nitrogen oxide and volatile organic compounds in nonlinear ways.

Depending on the chemical environment, reducing nitrogen oxide emissions could cause ground-level ozone to increase instead.

“Past research has focused on the role of emissions in forming ozone, but the influence of meteorology is a really important part of Emmie’s work,” Selin says.

To conduct their study, the researchers combined a global atmospheric chemistry model with a climate model that simulate future meteorology.

They used the climate model to generate meteorological inputs for each future year in their study, simulating factors such as likely temperature and wind speeds, in a way that captures the inherent variability of a region’s climate.

Then they fed those inputs to the atmospheric chemistry model, which calculates how the chemical composition of the atmosphere would change because of meteorology and emissions.

The researchers focused on Eastern North America, Western Europe, and Northeast China, since those regions have historically high levels of the precursor chemicals that form ozone and well-established monitoring networks to provide data.

They chose to model two future scenarios, one with high warming and one with low warming, over a 16-year period between 2080 and 2095. They compared them to a historical scenario capturing 2000 to 2015 to see the effects of a 10 percent reduction in nitrogen oxide emissions.

Capturing climate variability

“The biggest challenge is that the climate naturally varies from year to year. So, if you want to isolate the effects of climate change, you need to simulate enough years to see past that natural variability,” Le Roy says.

They could overcome that challenge due to recent advances in atmospheric chemistry modeling and by taking advantage of parallel computing to simulate multiple years at the same time. They simulated five 16-year realizations, resulting in 80 model years for each scenario.

The researchers found that eastern North America and Western Europe are especially sensitive to increases in nitrogen oxide emissions from the soil, which are natural emissions driven by increases in temperature.

Due to that sensitivity, as the Earth warms and more nitrogen oxide from soil enters the atmosphere, reducing nitrogen oxide emissions from human activities will have less of an impact on ground-level ozone.

“This shows how important it is to improve our representation of the biosphere in these models to better understand how climate change may impact air quality,” Le Roy says.

On the other hand, since industrial processes in northeast Asia cause more ozone per unit of nitrogen oxide emitted, cutting emissions there would cause greater reductions in ground-level ozone in future warming scenarios.

“But I wouldn’t say that is a good thing because it means that, overall, there are higher levels of ozone,” Le Roy adds.

Running detailed meteorology simulations, rather than relying on annual average weather data, gave the researchers a more complete picture of the potential effects on human health.

“Average climate isn’t the only thing that matters. One high ozone day, which might be a statistical anomaly, could mean we don’t meet our air quality target and have negative human health impacts that we should care about,” Le Roy says.

In the future, the researchers want to continue exploring the intersection of meteorology and air quality. They also want to expand their modeling approach to consider other climate change factors with high variability, like wildfires or biomass burning.

“We’ve shown that it is important for air quality scientists to consider the full range of climate variability, even if it is hard to do in your models, because it really does affect the answer that you get,” says Selin.

This work is funded, in part, by the MIT Praecis Presidential Fellowship, the J.H. and E.V. Wade Fellowship, and the MIT Martin Family Society of Fellows for Sustainability.

Humans naturally learn by making connections between sight and sound. For instance, we can watch someone playing the cello and recognize that the cellist’s movements are generating the music we hear.

A new approach developed by researchers from MIT and elsewhere improves an AI model’s ability to learn in this same fashion. This could be useful in applications such as journalism and film production, where the model could help with curating multimodal content through automatic video and audio retrieval.

In the longer term, this work could be used to improve a robot’s ability to understand real-world environments, where auditory and visual information are often closely connected.

Improving upon prior work from their group, the researchers created a method that helps machine-learning models align corresponding audio and visual data from video clips without the need for human labels.

They adjusted how their original model is trained so it learns a finer-grained correspondence between a particular video frame and the audio that occurs in that moment. The researchers also made some architectural tweaks that help the system balance two distinct learning objectives, which improves performance.

Taken together, these relatively simple improvements boost the accuracy of their approach in video retrieval tasks and in classifying the action in audiovisual scenes. For instance, the new method could automatically and precisely match the sound of a door slamming with the visual of it closing in a video clip.

“We are building AI systems that can process the world like humans do, in terms of having both audio and visual information coming in at once and being able to seamlessly process both modalities. Looking forward, if we can integrate this audio-visual technology into some of the tools we use on a daily basis, like large language models, it could open up a lot of new applications,” says Andrew Rouditchenko, an MIT graduate student and co-author of a paper on this research.

He is joined on the paper by lead author Edson Araujo, a graduate student at Goethe University in Germany; Yuan Gong, a former MIT postdoc; Saurabhchand Bhati, a current MIT postdoc; Samuel Thomas, Brian Kingsbury, and Leonid Karlinsky of IBM Research; Rogerio Feris, principal scientist and manager at the MIT-IBM Watson AI Lab; James Glass, senior research scientist and head of the Spoken Language Systems Group in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Hilde Kuehne, professor of computer science at Goethe University and an affiliated professor at the MIT-IBM Watson AI Lab. The work will be presented at the Conference on Computer Vision and Pattern Recognition.

Syncing up

This work builds upon a machine-learning method the researchers developed a few years ago, which provided an efficient way to train a multimodal model to simultaneously process audio and visual data without the need for human labels.

The researchers feed this model, called CAV-MAE, unlabeled video clips and it encodes the visual and audio data separately into representations called tokens. Using the natural audio from the recording, the model automatically learns to map corresponding pairs of audio and visual tokens close together within its internal representation space.

They found that using two learning objectives balances the model’s learning process, which enables CAV-MAE to understand the corresponding audio and visual data while improving its ability to recover video clips that match user queries.

But CAV-MAE treats audio and visual samples as one unit, so a 10-second video clip and the sound of a door slamming are mapped together, even if that audio event happens in just one second of the video.

In their improved model, called CAV-MAE Sync, the researchers split the audio into smaller windows before the model computes its representations of the data, so it generates separate representations that correspond to each smaller window of audio.

During training, the model learns to associate one video frame with the audio that occurs during just that frame.

“By doing that, the model learns a finer-grained correspondence, which helps with performance later when we aggregate this information,” Araujo says.

They also incorporated architectural improvements that help the model balance its two learning objectives.

Adding “wiggle room”

The model incorporates a contrastive objective, where it learns to associate similar audio and visual data, and a reconstruction objective which aims to recover specific audio and visual data based on user queries.

In CAV-MAE Sync, the researchers introduced two new types of data representations, or tokens, to improve the model’s learning ability.

They include dedicated “global tokens” that help with the contrastive learning objective and dedicated “register tokens” that help the model focus on important details for the reconstruction objective.

“Essentially, we add a bit more wiggle room to the model so it can perform each of these two tasks, contrastive and reconstructive, a bit more independently. That benefitted overall performance,” Araujo adds.

While the researchers had some intuition these enhancements would improve the performance of CAV-MAE Sync, it took a careful combination of strategies to shift the model in the direction they wanted it to go.

“Because we have multiple modalities, we need a good model for both modalities by themselves, but we also need to get them to fuse together and collaborate,” Rouditchenko says.

In the end, their enhancements improved the model’s ability to retrieve videos based on an audio query and predict the class of an audio-visual scene, like a dog barking or an instrument playing.

Its results were more accurate than their prior work, and it also performed better than more complex, state-of-the-art methods that require larger amounts of training data.

“Sometimes, very simple ideas or little patterns you see in the data have big value when applied on top of a model you are working on,” Araujo says.

In the future, the researchers want to incorporate new models that generate better data representations into CAV-MAE Sync, which could improve performance. They also want to enable their system to handle text data, which would be an important step toward generating an audiovisual large language model.

This work is funded, in part, by the German Federal Ministry of Education and Research and the MIT-IBM Watson AI Lab.

We’ve all felt the sting of guilt when fruit and vegetables go bad before we could eat them. Now, researchers from MIT and the Singapore-MIT Alliance for Research and Technology (SMART) have shown they can extend the shelf life of harvested plants by injecting them with melatonin using biodegradable microneedles.

That’s a big deal because the problem of food waste goes way beyond our salads. More than 30 percent of the world’s food is lost after it’s harvested — enough to feed more than 1 billion people. Refrigeration is the most common way to preserve foods, but it requires energy and infrastructure that many regions of the world can’t afford or lack access to.

The researchers believe their system could offer an alternative or complement to refrigeration. Central to their approach are patches of silk microneedles. The microneedles can get through the tough, waxy skin of plants without causing a stress response, and deliver precise amounts of melatonin into plants’ inner tissues.

“This is the first time that we’ve been able to apply these microneedles to extend the shelf life of a fresh-cut crop,” says Benedetto Marelli, the study’s senior author, associate professor of civil and environmental engineering at MIT, and the director of the Wild Cards mission of the MIT Climate Project. “We thought we could use this technology to deliver something that could regulate or control the plant’s post-harvest physiology. Eventually, we looked at hormones, and melatonin is already used by plants to regulate such functions. The food we waste could feed about 1.6 billion people. Even in the U.S., this approach could one day expand access to healthy foods.”

For the study, which appears today in Nano Letters, Marelli and researchers from SMART applied small patches of the microneedles containing melatonin to the base of the leafy vegetable pak choy. After application, the researchers found the melatonin was able to extend the vegetables’ shelf life by four days at room temperature and 10 days when refrigerated, which could allow more crops to reach consumers before they’re wasted.

“Post-harvest waste is a huge issue. This problem is extremely important in emerging markets around Africa and Southeast Asia, where many crops are produced but can't be maintained in the journey from farms to markets,” says Sarojam Rajani, co-senior author of the study and a senior principal investigator at the Temasek Life Sciences Laboratory in Singapore.

Plant destressors

For years, Marelli’s lab has been exploring the use of silk microneedles for things like delivering nutrients to crops and monitoring plant health. Microneedles made from silk fibroin protein are nontoxic and biodegradable, and Marelli’s previous work has described ways of manufacturing them at scale.

To test microneedle’s ability to extend the shelf life of food, the researchers wanted to study their ability to deliver a hormone known to affect the senescence process. Aside from helping humans sleep, melatonin is also a natural hormone in many plants that helps them regulate growth and aging.

“The dose of melatonin we’re delivering is so low that it’s fully metabolized by the crops, so it would not significantly increase the amount of melatonin normally present in the food; we would not ingest more melatonin than usual,” Marelli says. “We chose pak choy because it's a very important crop in Asia, and also because pak choy is very perishable.”

Pak choy is typically harvested by cutting the leafy plant from the root system, exposing the shoot base that provides easy access to vascular bundles which distribute water and nutrients to the rest of the plant. To begin their study, the researchers first used their microneedles to inject a fluorescent dye into the base to confirm that vasculature could spread the dye throughout the plant.

The researchers then compared the shelf life of regular pak choy plants and plants that had been sprayed with or dipped into melatonin, finding no difference.

With their baseline shelf life established, the researchers applied small patches of the melatonin-filled microneedles to the bottom of pak choy plants by hand. They then stored the treated plants, along with controls, in plastic boxes both at room temperature and under refrigeration.

The team evaluated the plants by monitoring their weight, visual appearance, and concentration of chlorophyll, a green pigment that decreases as plants age.

At room temperature, the leaves of the untreated control group began yellowing within two or three days. By the fourth day, the yellowing accelerated to the point that the plants likely could not be sold. Plants treated with the melatonin-loaded silk microneedles, in contrast, remained green on day five, and the yellowing process was significantly delayed. The weight loss and chlorophyll reduction of treated plants also slowed significantly at room temperature. Overall, the researchers estimated the microneedle-treated plants retained their saleable value until the eighth day.

“We clearly saw we could enhance the shelf life of pak choy without the cold chain,” Marelli says.

In refrigerated conditions of about 40 degrees Fahrenheit, plant yellowing was delayed by about five days on average, with treated plants remaining relatively green until day 25.

“Spectrophotometric analysis of the plants indicated the treated plants had higher antioxidant activity, while gene analysis showed the melatonin set off a protective chain reaction inside the plants, preserving chlorophyll and adjusting hormones to slow senescence,” says Monika Jangir, co-first author and former postdoc at the Temasek Life Sciences Laboratory.

“We studied melatonin’s effects and saw it improves the stress response of the plant after it’s been cut, so it’s basically decreasing the stress that plant’s experience, and that extends its shelf life,” says Yangyang Han, co-first author and research scientist at the Disruptive and Sustainable Technologies for Agricultural Precision (DiSTAP) interdisciplinary research group at SMART.

Toward postharvest preservation

While the microneedles could make it possible to minimize waste when compared to other application methods like spraying or dipping crops, the researchers say more work is needed to deploy microneedles at scale. For instance, although the researchers applied the microneedle patches by hand in this experiment, the patches could be applied using tractors, autonomous drones, and other farming equipment in the future.

“For this to be widely adopted, we’d need to reach a performance versus cost threshold to justify its use,” Marelli explains. “This method would need to become cheap enough to be used by farmers regularly.”

Moving forward, the research team plans to study the effects of a variety of hormones on different crops using its microneedle delivery technology. The team believes the technique should work with all kinds of produce.

“We’re going to continue to analyze how we can increase the impact this can have on the value and quality of crops,” Marelli says. “For example, could this let us modulate the nutritional values of the crop, how it’s shaped, its texture, etc.? We're also going to continue looking into scaling up the technology so this can be used in the field.”

The work was supported by the Singapore-MIT Alliance for Research and Technology (SMART) and the National Research Foundation of Singapore.

One of the most profound open questions in modern physics is: “Is gravity quantum?” 

The other fundamental forces — electromagnetic, weak, and strong — have all been successfully described, but no complete and consistent quantum theory of gravity yet exists.  

“Theoretical physicists have proposed many possible scenarios, from gravity being inherently classical to fully quantum, but the debate remains unresolved because we’ve never had a clear way to test gravity’s quantum nature in the lab,” says Dongchel Shin, a PhD candidate in the MIT Department of Mechanical Engineering (MechE). “The key to answering this lies in preparing mechanical systems that are massive enough to feel gravity, yet quiet enough — quantum enough — to reveal how gravity interacts with them.”

Shin, who is also a MathWorks Fellow, researches quantum and precision metrology platforms that probe fundamental physics and are designed to pave the way for future industrial technology. He is the lead author of a new paper that demonstrates laser cooling of a centimeter-long torsional oscillator. The open-access paper, “Active laser cooling of a centimeter-scale torsional oscillator,” was recently published in the journal Optica. 

Lasers have been routinely employed to cool down atomic gases since the 1980s, and have been used in the linear motion of nanoscale mechanical oscillators since around 2010. The new paper presents the first time this technique has been extended to torsional oscillators, which are key to a worldwide effort to study gravity using these systems.

“Torsion pendulums have been classical tools for gravity research since [Henry] Cavendish’s famous experiment in 1798. They’ve been used to measure Newton’s gravitational constant, G, test the inverse-square law, and search for new gravitational phenomena,” explains Shin.

By using lasers to remove nearly all thermal motion from atoms, in recent decades scientists have created ultracold atomic gases at micro- and nanokelvin temperatures. These systems now power the world’s most precise clocks — optical lattice clocks — with timekeeping precision so high that they would gain or lose less than a second over the age of the universe.

“Historically, these two technologies developed separately — one in gravitational physics, the other in atomic and optical physics,” says Shin. “In our work, we bring them together. By applying laser cooling techniques originally developed for atoms to a centimeter-scale torsional oscillator, we try to bridge the classical and quantum worlds. This hybrid platform enables a new class of experiments — ones that could finally let us test whether gravity needs to be described by quantum theory.”

The new paper demonstrates laser cooling of a centimeter-scale torsional oscillator from room temperature to a temperature of 10 millikelvins (1/1,000th of a kelvin) using a mirrored optical lever.

“An optical lever is a simple but powerful measurement technique: You shine a laser onto a mirror, and even a tiny tilt of the mirror causes the reflected beam to shift noticeably on a detector. This magnifies small angular motions into easily measurable signals,” explains Shin, noting that while the premise is simple, the team faced challenges in practice. “The laser beam itself can jitter slightly due to air currents, vibrations, or imperfections in the optics. These jitters can falsely appear as motion of the mirror, limiting our ability to measure true physical signals.”

To overcome this, the team used the mirrored optical lever approach, which employs a second, mirrored version of the laser beam to cancel out the unwanted jitter.

“One beam interacts with the torsional oscillator, while the other reflects off a corner-cube mirror, reversing any jitter without picking up the oscillator’s motion,” Shin says. “When the two beams are combined at the detector, the real signal from the oscillator is preserved, and the false motion from [the] laser jitter is canceled.”

This approach reduced noise by a factor of a thousand, which allowed the researchers to detect motion with extreme precision, nearly 10 times better than the oscillator’s own quantum zero-point fluctuations. “That level of sensitivity made it possible for us to cool the system down to just 10 milli-kelvins using laser light,” Shin says.

Shin says this work is just the beginning. “While we’ve achieved quantum-limited precision below the zero-point motion of the oscillator, reaching the actual quantum ground state remains our next goal,” he says. “To do that, we’ll need to further strengthen the optical interaction — using an optical cavity that amplifies angular signals, or optical trapping strategies. These improvements could open the door to experiments where two such oscillators interact only through gravity, allowing us to directly test whether gravity is quantum or not.”

The paper’s other authors from the Department of Mechanical Engineering include Vivishek Sudhir, assistant professor of mechanical engineering and the Class of 1957 Career Development Professor, and PhD candidate Dylan Fife. Additional authors are Tina Heyward and Rajesh Menon of the Department of Electrical and Computer Engineering at the University of Utah. Shin and Fife are both members of Sudhir’s lab, the Quantum and Precision Measurements Group.

Shin says one thing he’s come to appreciate through this work is the breadth of the challenge the team is tackling. “Studying quantum aspects of gravity experimentally doesn’t just require deep understanding of physics — relativity, quantum mechanics — but also demands hands-on expertise in system design, nanofabrication, optics, control, and electronics,” he says.

“Having a background in mechanical engineering, which spans both the theoretical and practical aspects of physical systems, gave me the right perspective to navigate and contribute meaningfully across these diverse domains,” says Shin. “It’s been incredibly rewarding to see how this broad training can help tackle one of the most fundamental questions in science.”

Removing carbon dioxide from the atmosphere efficiently is often seen as a crucial need for combatting climate change, but systems for removing carbon dioxide suffer from a tradeoff. Chemical compounds that efficiently remove CO₂ from the air do not easily release it once captured, and compounds that release CO₂ efficiently are not very efficient at capturing it. Optimizing one part of the cycle tends to make the other part worse.

Now, using nanoscale filtering membranes, researchers at MIT have added a simple intermediate step that facilitates both parts of the cycle. The new approach could improve the efficiency of electrochemical carbon dioxide capture and release by six times and cut costs by at least 20 percent, they say.

The new findings are reported today in the journal ACS Energy Letters, in a paper by MIT doctoral students Simon Rufer, Tal Joseph, and Zara Aamer, and professor of mechanical engineering Kripa Varanasi.

“We need to think about scale from the get-go when it comes to carbon capture, as making a meaningful impact requires processing gigatons of CO₂,” says Varanasi. “Having this mindset helps us pinpoint critical bottlenecks and design innovative solutions with real potential for impact. That’s the driving force behind our work.”

Many carbon-capture systems work using chemicals called hydroxides, which readily combine with carbon dioxide to form carbonate. That carbonate is fed into an electrochemical cell, where the carbonate reacts with an acid to form water and release carbon dioxide. The process can take ordinary air with only about 400 parts per million of carbon dioxide and generate a stream of 100 percent pure carbon dioxide, which can then be used to make fuels or other products.

Both the capture and release steps operate in the same water-based solution, but the first step needs a solution with a high concentration of hydroxide ions, and the second step needs one high in carbonate ions. “You can see how these two steps are at odds,” says Varanasi. “These two systems are circulating the same sorbent back and forth. They’re operating on the exact same liquid. But because they need two different types of liquids to operate optimally, it’s impossible to operate both systems at their most efficient points.”

The team’s solution was to decouple the two parts of the system and introduce a third part in between. Essentially, after the hydroxide in the first step has been mostly chemically converted to carbonate, special nanofiltration membranes then separate ions in the solution based on their charge. Carbonate ions have a charge of 2, while hydroxide ions have a charge of 1. “The nanofiltration is able to separate these two pretty well,” Rufer says.

Once separated, the hydroxide ions are fed back to the absorption side of the system, while the carbonates are sent ahead to the electrochemical release stage. That way, both ends of the system can operate at their more efficient ranges. Varanasi explains that in the electrochemical release step, protons are being added to the carbonate to cause the conversion to carbon dioxide and water, but if hydroxide ions are also present, the protons will react with those ions instead, producing just water.

“If you don’t separate these hydroxides and carbonates,” Rufer says, “the way the system fails is you’ll add protons to hydroxide instead of carbonate, and so you’ll just be making water rather than extracting carbon dioxide. That’s where the efficiency is lost. Using nanofiltration to prevent this was something that we aren’t aware of anyone proposing before.”

Testing showed that the nanofiltration could separate the carbonate from the hydroxide solution with about 95 percent efficiency, validating the concept under realistic conditions, Rufer says. The next step was to assess how much of an effect this would have on the overall efficiency and economics of the process. They created a techno-economic model, incorporating electrochemical efficiency, voltage, absorption rate, capital costs, nanofiltration efficiency, and other factors.

The analysis showed that present systems cost at least $600 per ton of carbon dioxide captured, while with the nanofiltration component added, that drops to about $450 a ton. What’s more, the new system is much more stable, continuing to operate at high efficiency even under variations in the ion concentrations in the solution. “In the old system without nanofiltration, you’re sort of operating on a knife’s edge,” Rufer says; if the concentration varies even slightly in one direction or the other, efficiency drops off drastically. “But with our nanofiltration system, it kind of acts as a buffer where it becomes a lot more forgiving. You have a much broader operational regime, and you can achieve significantly lower costs.”

He adds that this approach could apply not only to the direct air capture systems they studied specifically, but also to point-source systems — which are attached directly to the emissions sources such as power plant emissions — or to the next stage of the process, converting captured carbon dioxide into useful products such as fuel or chemical feedstocks.  Those conversion processes, he says, “are also bottlenecked in this carbonate and hydroxide tradeoff.”

In addition, this technology could lead to safer alternative chemistries for carbon capture, Varanasi says. “A lot of these absorbents can at times be toxic, or damaging to the environment. By using a system like ours, you can improve the reaction rate, so you can choose chemistries that might not have the best absorption rate initially but can be improved to enable safety.”

Varanasi adds that “the really nice thing about this is we’ve been able to do this with what’s commercially available,” and with a system that can easily be retrofitted to existing carbon-capture installations. If the costs can be further brought down to about $200 a ton, it could be viable for widespread adoption. With ongoing work, he says, “we’re confident that we’ll have something that can become economically viable” and that will ultimately produce valuable, saleable products.

Rufer notes that even today, “people are buying carbon credits at a cost of over $500 per ton. So, at this cost we’re projecting, it is already commercially viable in that there are some buyers who are willing to pay that price.” But by bringing the price down further, that should increase the number of buyers who would consider buying the credit, he says. “It’s just a question of how widespread we can make it.” Recognizing this growing market demand, Varanasi says, “Our goal is to provide industry scalable, cost-effective, and reliable technologies and systems that enable them to directly meet their decarbonization targets.”

The research was supported by Shell International Exploration and Production Inc. through the MIT Energy Initiative, and the U.S. National Science Foundation, and made use of the facilities at MIT.nano.

Measuring the density of a cell can reveal a great deal about the cell’s state. As cells proliferate, differentiate, or undergo cell death, they may gain or lose water and other molecules, which is revealed by changes in density.

Tracking these tiny changes in cells’ physical state is difficult to do at a large scale, especially with single-cell resolution, but a team of MIT researchers has now found a way to measure cell density quickly and accurately — measuring up to 30,000 cells in a single hour.

The researchers also showed that density changes could be used to make valuable predictions, including whether immune cells such as T cells have become activated to kill tumors, or whether tumor cells are susceptible to a specific drug.

“These predictions are all based on looking at very small changes in the physical properties of cells, which can tell you how they’re going to respond,” says Scott Manalis, the David H. Koch Professor of Engineering in the departments of Biological Engineering and Mechanical Engineering, and a member of the Koch Institute for Integrative Cancer Research.

Manalis is the senior author of the new study, which appears today in Nature Biomedical Engineering. The paper’s lead author is MIT Research Scientist Weida (Richard) Wu.

Measuring density

As cells enter new states, their molecular contents, including lipids, proteins, and nucleic acids, can become more or less crowded. Measuring the density of a cell offers an indirect view of this crowding.

The new density measurement technique reported in this study builds on work that Manalis’ lab has done over the past two decades on technologies for making measurements of cells and tiny particles. In 2007, his lab developed a microfluidic device known as a suspended microchannel resonator (SMR), which consists of a microchannel across a tiny silicon cantilever that vibrates at a specific frequency. As a cell passes through the channel, the frequency of the vibration changes slightly, and the magnitude of that change can be used to calculate the cell’s mass.

In 2011, the researchers adapted the technique to measure the density of cells. To achieve that, cells are sent through the device twice, suspended in two liquids of different densities. A cell’s buoyant mass (its mass as it floats in fluid) depends on its absolute mass and volume, so by measuring two different buoyant masses for a cell, its mass, volume, and density can be calculated.

That technique works well, but swapping fluids and flowing cells through each one is time-consuming, so it can only be used to measure a few hundred cells at a time.

To create a faster, more streamlined system, the researchers combined their SMR device with a fluorescent microscope, which enables measurements of cell volume. The microscope is positioned at the entrance to the resonator, and cells flow through the device while floating in a fluorescent dye that can’t be absorbed by cells. When cells pass by the microscope, the dip in the fluorescent signal can be used to determine the volume of the cell.

After that volume measurement is taken, the cells flow into the resonator, which measures their mass. This process, which allows for rapid calculation of density, can be used to measure up to 30,000 cells in an hour.

“Instead of trying to flow the cells back and forth at least twice through the cantilever to get cell density, we wanted to try to create a method to do a streamlined measurement, so the cells only need to pass through the cantilever once,” Wu says. “From a cell’s mass and volume, we can then derive its density, without compromising the throughput or the precision.”

Evaluating T cells

The researchers used their new technique to track what happens to the density of T cells after they are activated by signaling molecules.

As T cells transition from a quiescent state to an active state, they gain new molecules, as well as water, the researchers found. From their pre-activation state to the first day of activation, the densities of the cells dropped from an average of 1.08 grams per milliliter to 1.06 grams per milliliter. This means that the cells are becoming less crowded, as they gain water faster than they gain other molecules.

“This is suggesting that cell density is very likely reflecting an increase in cellular water content as the cells transit from a quiescent, non-proliferative state to a high-growth state,” Wu says. “These data are pointing to the notion that cell density is an interesting biomarker that is changing during T-cell activation and may have functional relevance to how well the T cells could proliferate.”

Travera, a clinical-stage company co-founded by Manalis, is working on using the SMR mass measurements to predict whether individual cancer patients’ T cells will respond to drugs meant to stimulate a strong anti-tumor immune response. The company has also begun using the density measurement technique, and preliminary studies have found that using mass and density measurements together gives a much more accurate prediction that using either one alone.

“Both mass and density are revealing something about the overall fitness of the immune cells,” Manalis says.

Using physical measurements of cells to monitor their immune activation “is very exciting and may offer a new way of evaluating and measuring changes in immune cells in circulation,” says Genevieve Boland, an associate professor of surgery at Harvard Medical School and vice chair of research for the Integrated Department of Surgery at Mass General Brigham, who was not involved in the study.

“This is a complementary, but very different, method than those currently used for immune assessments in cancer and other diseases, potentially offering a novel tool to assist in clinical decision-making regarding the need for and the choice of a specific cancer therapy, allow monitoring of response to therapy, and/or in early detection of side effects of immune-based therapies,” she says.

Making predictions

Another potential application for this approach is predicting how tumor cells will respond to different types of cancer drugs. In previous work, Manalis has shown that tracking changes in cell mass after treatment can predict whether a tumor cell is undergoing drug-induced apoptosis. In the new study, he found that density could also reveal these responses.

In those experiments, the researchers treated pancreatic cancer cells with one of two different drugs — one that the cells are susceptible to, and one they are resistant to. They found that density changes after treatment accurately reflected the cells’ known responses to treatment.

“We capture something about the cells that is highly predictive within the first couple of days after they get taken out from the tumor,” Wu says. “Cell density is a rapid biomarker to predict in vivo drug response in a very timely manner.”

Manalis’ lab is now working on using measurements of cell mass and density as a way to evaluate the fitness of cells used to synthesize complex proteins such as therapeutic antibodies.

“As cells are producing these proteins, we can learn from these markers of cell fitness and metabolic state to try to make predictions about how well these cells can produce these proteins, and hopefully in the future also guide design and control strategies to even further improve the yield of these complex proteins,” Wu says.

The research was funded by the Paul G. Allen Frontiers Group, the Virginia and Daniel K. Ludwig Fund for Cancer Research, the MIT Center for Precision Cancer Medicine, the Stand up to Cancer Convergence Program, Bristol Myers Squibb, and the Koch Institute Support (core) Grant from the National Cancer Institute.

By combining information from many large datasets, MIT researchers have identified several new potential targets for treating or preventing Alzheimer’s disease.

The study revealed genes and cellular pathways that haven’t been linked to Alzheimer’s before, including one involved in DNA repair. Identifying new drug targets is critical because many of the Alzheimer’s drugs that have been developed to this point haven’t been as successful as hoped.

Working with researchers at Harvard Medical School, the team used data from humans and fruit flies to identify cellular pathways linked to neurodegeneration. This allowed them to identify additional pathways that may be contributing to the development of Alzheimer’s.

“All the evidence that we have indicates that there are many different pathways involved in the progression of Alzheimer’s. It is multifactorial, and that may be why it’s been so hard to develop effective drugs,” says Ernest Fraenkel, the Grover M. Hermann Professor in Health Sciences and Technology in MIT’s Department of Biological Engineering and the senior author of the study. “We will need some kind of combination of treatments that hit different parts of this disease.”

Matthew Leventhal PhD ’25 is the lead author of the paper, which appears today in Nature Communications.

Alternative pathways

Over the past few decades, many studies have suggested that Alzheimer’s disease is caused by the buildup of amyloid plaques in the brain, which triggers a cascade of events that leads to neurodegeneration.

A handful of drugs have been developed to block or break down these plaques, but these drugs usually do not have a dramatic effect on disease progression. In hopes of identifying new drug targets, many scientists are now working on uncovering other mechanisms that might contribute to the development of Alzheimer’s.

“One possibility is that maybe there’s more than one cause of Alzheimer’s, and that even in a single person, there could be multiple contributing factors,” Fraenkel says. “So, even if the amyloid hypothesis is correct — and there are some people who don’t think it is — you need to know what those other factors are. And then if you can hit all the causes of the disease, you have a better chance of blocking and maybe even reversing some losses.”

To try to identify some of those other factors, Fraenkel’s lab teamed up with Mel Feany, a professor of pathology at Harvard Medical School and a geneticist specializing in fruit fly genetics.

Using fruit flies as a model, Feany and others in her lab did a screen in which they knocked out nearly every conserved gene expressed in fly neurons. Then, they measured whether each of these gene knockdowns had any effect on the age at which the flies develop neurodegeneration. This allowed them to identify about 200 genes that accelerate neurodegeneration.

Some of these were already linked to neurodegeneration, including genes for the amyloid precursor protein and for proteins called presenillins, which play a role in the formation of amyloid proteins.

The researchers then analyzed this data using network algorithms that Fraenkel’s lab has been developing over the past several years. These are algorithms that can identify connections between genes that may be involved in the same cellular pathways and functions.

In this case, the aim was to try to link the genes identified in the fruit fly screen with specific processes and cellular pathways that might contribute to neurodegeneration. To do that, the researchers combined the fruit fly data with several other datasets, including genomic data from postmortem tissue of Alzheimer’s patients.

The first stage of their analysis revealed that many of the genes identified in the fruit fly study also decline as humans age, suggesting that they may be involved in neurodegeneration in humans.

Network analysis

In the next phase of their study, the researchers incorporated additional data relevant to Alzheimer’s disease, including eQTL (expression quantitative trait locus) data — ­a measure of how different gene variants affect the expression levels of certain proteins.

Using their network optimization algorithms on this data, the researchers identified pathways that link genes to their potential role in Alzheimer’s development. The team chose two of those pathways to focus on in the new study.

The first is a pathway, not previously linked to Alzheimer’s disease, related to RNA modification. The network suggested that when one of two of the genes in this pathway — MEPCE and HNRNPA2B1 — are missing, neurons become more vulnerable to the Tau tangles that form in the brains of Alzheimer’s patients. The researchers confirmed this effect by knocking down those genes in studies of fruit flies and in human neurons derived from induced pluripotent stem cells (IPSCs).

The second pathway reported in this study is involved in DNA damage repair. This network includes two genes called NOTCH1 and CSNK2A1, which have been linked to Alzheimer’s before, but not in the context of DNA repair. Both genes are most well-known for their roles in regulating cell growth.

In this study, the researchers found evidence that when these genes are missing, DNA damage builds up in cells, through two different DNA-damaging pathways. Buildup of unrepaired DNA has previously been shown to lead to neurodegeneration.

Now that these targets have been identified, the researchers hope to collaborate with other labs to help explore whether drugs that target them could improve neuron health. Fraenkel and other researchers are working on using IPSCs from Alzheimer’s patients to generate neurons that could be used to evaluate such drugs.

“The search for Alzheimer’s drugs will get dramatically accelerated when there are very good, robust experimental systems,” he says. “We’re coming to a point where a couple of really innovative systems are coming together. One is better experimental models based on IPSCs, and the other one is computational models that allow us to integrate huge amounts of data. When those two mature at the same time, which is what we’re about to see, then I think we’ll have some breakthroughs.”

The research was funded by the National Institutes of Health.

The ocean is teeming with life. But unless you get up close, much of the marine world can easily remain unseen. That’s because water itself can act as an effective cloak: Light that shines through the ocean can bend, scatter, and quickly fade as it travels through the dense medium of water and reflects off the persistent haze of ocean particles. This makes it extremely challenging to capture the true color of objects in the ocean without imaging them at close range.

Now a team from MIT and the Woods Hole Oceanographic Institution (WHOI) has developed an image-analysis tool that cuts through the ocean’s optical effects and generates images of underwater environments that look as if the water had been drained away, revealing an ocean scene’s true colors. The team paired the color-correcting tool with a computational model that converts images of a scene into a three-dimensional underwater “world,” that can then be explored virtually.

The researchers have dubbed the new tool “SeaSplat,” in reference to both its underwater application and a method known as 3D gaussian splatting (3DGS), which takes images of a scene and stitches them together to generate a complete, three-dimensional representation that can be viewed in detail, from any perspective.

“With SeaSplat, it can model explicitly what the water is doing, and as a result it can in some ways remove the water, and produces better 3D models of an underwater scene,” says MIT graduate student Daniel Yang.

The researchers applied SeaSplat to images of the sea floor taken by divers and underwater vehicles, in various locations including the U.S. Virgin Islands. The method generated 3D “worlds” from the images that were truer and more vivid and varied in color, compared to previous methods.

The team says SeaSplat could help marine biologists monitor the health of certain ocean communities. For instance, as an underwater robot explores and takes pictures of a coral reef, SeaSplat would simultaneously process the images and render a true-color, 3D representation, that scientists could then virtually “fly” through, at their own pace and path, to inspect the underwater scene, for instance for signs of coral bleaching.

“Bleaching looks white from close up, but could appear blue and hazy from far away, and you might not be able to detect it,” says Yogesh Girdhar, an associate scientist at WHOI. “Coral bleaching, and different coral species, could be easier to detect with SeaSplat imagery, to get the true colors in the ocean.”

Girdhar and Yang will present a paper detailing SeaSplat at the IEEE International Conference on Robotics and Automation (ICRA). Their study co-author is John Leonard, professor of mechanical engineering at MIT.

Aquatic optics

In the ocean, the color and clarity of objects is distorted by the effects of light traveling through water. In recent years, researchers have developed color-correcting tools that aim to reproduce the true colors in the ocean. These efforts involved adapting tools that were developed originally for environments out of water, for instance to reveal the true color of features in foggy conditions. One recent work accurately reproduces true colors in the ocean, with an algorithm named “Sea-Thru,” though this method requires a huge amount of computational power, which makes its use in producing 3D scene models challenging.

In parallel, others have made advances in 3D gaussian splatting, with tools that seamlessly stitch images of a scene together, and intelligently fill in any gaps to create a whole, 3D version of the scene. These 3D worlds enable “novel view synthesis,” meaning that someone can view the generated 3D scene, not just from the perspective of the original images, but from any angle and distance.

But 3DGS has only successfully been applied to environments out of water. Efforts to adapt 3D reconstruction to underwater imagery have been hampered, mainly by two optical underwater effects: backscatter and attenuation. Backscatter occurs when light reflects off of tiny particles in the ocean, creating a veil-like haze. Attenuation is the phenomenon by which light of certain wavelengths attenuates, or fades with distance. In the ocean, for instance, red objects appear to fade more than blue objects when viewed from farther away.

Out of water, the color of objects appears more or less the same regardless of the angle or distance from which they are viewed. In water, however, color can quickly change and fade depending on one’s perspective. When 3DGS methods attempt to stitch underwater images into a cohesive 3D whole, they are unable to resolve objects due to aquatic backscatter and attenuation effects that distort the color of objects at different angles.

“One dream of underwater robotic vision that we have is: Imagine if you could remove all the water in the ocean. What would you see?” Leonard says.

A model swim

In their new work, Yang and his colleagues developed a color-correcting algorithm that accounts for the optical effects of backscatter and attenuation. The algorithm determines the degree to which every pixel in an image must have been distorted by backscatter and attenuation effects, and then essentially takes away those aquatic effects, and computes what the pixel’s true color must be.

Yang then worked the color-correcting algorithm into a 3D gaussian splatting model to create SeaSplat, which can quickly analyze underwater images of a scene and generate a true-color, 3D virtual version of the same scene that can be explored in detail from any angle and distance.

The team applied SeaSplat to multiple underwater scenes, including images taken in the Red Sea, in the Carribean off the coast of Curaçao, and the Pacific Ocean, near Panama. These images, which the team took from a pre-existing dataset, represent a range of ocean locations and water conditions. They also tested SeaSplat on images taken by a remote-controlled underwater robot in the U.S. Virgin Islands.

From the images of each ocean scene, SeaSplat generated a true-color 3D world that the researchers were able to virtually explore, for instance zooming in and out of a scene and viewing certain features from different perspectives. Even when viewing from different angles and distances, they found objects in every scene retained their true color, rather than fading as they would if viewed through the actual ocean.

“Once it generates a 3D model, a scientist can just ‘swim’ through the model as though they are scuba-diving, and look at things in high detail, with real color,” Yang says.

For now, the method requires hefty computing resources in the form of a desktop computer that would be too bulky to carry aboard an underwater robot. Still, SeaSplat could work for tethered operations, where a vehicle, tied to a ship, can explore and take images that can be sent up to a ship’s computer.

“This is the first approach that can very quickly build high-quality 3D models with accurate colors, underwater, and it can create them and render them fast,” Girdhar says. “That will help to quantify biodiversity, and assess the health of coral reef and other marine communities.”

This work was supported, in part, by the Investment in Science Fund at WHOI, and by the U.S. National Science Foundation.

A protein located in the wrong part of a cell can contribute to several diseases, such as Alzheimer’s, cystic fibrosis, and cancer. But there are about 70,000 different proteins and protein variants in a single human cell, and since scientists can typically only test for a handful in one experiment, it is extremely costly and time-consuming to identify proteins’ locations manually.

A new generation of computational techniques seeks to streamline the process using machine-learning models that often leverage datasets containing thousands of proteins and their locations, measured across multiple cell lines. One of the largest such datasets is the Human Protein Atlas, which catalogs the subcellular behavior of over 13,000 proteins in more than 40 cell lines. But as enormous as it is, the Human Protein Atlas has only explored about 0.25 percent of all possible pairings of all proteins and cell lines within the database.

Now, researchers from MIT, Harvard University, and the Broad Institute of MIT and Harvard have developed a new computational approach that can efficiently explore the remaining uncharted space. Their method can predict the location of any protein in any human cell line, even when both protein and cell have never been tested before.

Their technique goes one step further than many AI-based methods by localizing a protein at the single-cell level, rather than as an averaged estimate across all the cells of a specific type. This single-cell localization could pinpoint a protein’s location in a specific cancer cell after treatment, for instance.

The researchers combined a protein language model with a special type of computer vision model to capture rich details about a protein and cell. In the end, the user receives an image of a cell with a highlighted portion indicating the model’s prediction of where the protein is located. Since a protein’s localization is indicative of its functional status, this technique could help researchers and clinicians more efficiently diagnose diseases or identify drug targets, while also enabling biologists to better understand how complex biological processes are related to protein localization.

“You could do these protein-localization experiments on a computer without having to touch any lab bench, hopefully saving yourself months of effort. While you would still need to verify the prediction, this technique could act like an initial screening of what to test for experimentally,” says Yitong Tseo, a graduate student in MIT’s Computational and Systems Biology program and co-lead author of a paper on this research.

Tseo is joined on the paper by co-lead author Xinyi Zhang, a graduate student in the Department of Electrical Engineering and Computer Science (EECS) and the Eric and Wendy Schmidt Center at the Broad Institute; Yunhao Bai of the Broad Institute; and senior authors Fei Chen, an assistant professor at Harvard and a member of the Broad Institute, and Caroline Uhler, the Andrew and Erna Viterbi Professor of Engineering in EECS and the MIT Institute for Data, Systems, and Society (IDSS), who is also director of the Eric and Wendy Schmidt Center and a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS). The research appears today in Nature Methods.

Collaborating models

Many existing protein prediction models can only make predictions based on the protein and cell data on which they were trained or are unable to pinpoint a protein’s location within a single cell.

To overcome these limitations, the researchers created a two-part method for prediction of unseen proteins’ subcellular location, called PUPS.

The first part utilizes a protein sequence model to capture the localization-determining properties of a protein and its 3D structure based on the chain of  amino acids that forms it.

The second part incorporates an image inpainting model, which is designed to fill in missing parts of an image. This computer vision model looks at three stained images of a cell to gather information about the state of that cell, such as its type, individual features, and whether it is under stress.

PUPS joins the representations created by each model to predict where the protein is located within a single cell, using an image decoder to output a highlighted image that shows the predicted location.

“Different cells within a cell line exhibit different characteristics, and our model is able to understand that nuance,” Tseo says.

A user inputs the sequence of amino acids that form the protein and three cell stain images — one for the nucleus, one for the microtubules, and one for the endoplasmic reticulum. Then PUPS does the rest.

A deeper understanding

The researchers employed a few tricks during the training process to teach PUPS how to combine information from each model in such a way that it can make an educated guess on the protein’s location, even if it hasn’t seen that protein before.

For instance, they assign the model a secondary task during training: to explicitly name the compartment of localization, like the cell nucleus. This is done alongside the primary inpainting task to help the model learn more effectively.

A good analogy might be a teacher who asks their students to draw all the parts of a flower in addition to writing their names. This extra step was found to help the model improve its general understanding of the possible cell compartments.

In addition, the fact that PUPS is trained on proteins and cell lines at the same time helps it develop a deeper understanding of where in a cell image proteins tend to localize.

PUPS can even understand, on its own, how different parts of a protein’s sequence contribute separately to its overall localization.

“Most other methods usually require you to have a stain of the protein first, so you’ve already seen it in your training data. Our approach is unique in that it can generalize across proteins and cell lines at the same time,” Zhang says.

Because PUPS can generalize to unseen proteins, it can capture changes in localization driven by unique protein mutations that aren’t included in the Human Protein Atlas.

The researchers verified that PUPS could predict the subcellular location of new proteins in unseen cell lines by conducting lab experiments and comparing the results. In addition, when compared to a baseline AI method, PUPS exhibited on average less prediction error across the proteins they tested.

In the future, the researchers want to enhance PUPS so the model can understand protein-protein interactions and make localization predictions for multiple proteins within a cell. In the longer term, they want to enable PUPS to make predictions in terms of living human tissue, rather than cultured cells.

This research is funded by the Eric and Wendy Schmidt Center at the Broad Institute, the National Institutes of Health, the National Science Foundation, the Burroughs Welcome Fund, the Searle Scholars Foundation, the Harvard Stem Cell Institute, the Merkin Institute, the Office of Naval Research, and the Department of Energy.

Around the world, 20 percent of children are not fully immunized, leading to 1.5 million child deaths each year from diseases that are preventable by vaccination. About half of those underimmunized children received at least one vaccine dose but did not complete the vaccination series, while the rest received no vaccines at all.

To make it easier for children to receive all of their vaccines, MIT researchers are working to develop microparticles that can release their payload weeks or months after being injected. This could lead to vaccines that can be given just once, with several doses that would be released at different time points.

In a study appearing today in the journal Advanced Materials, the researchers showed that they could use these particles to deliver two doses of diphtheria vaccine — one released immediately, and the second two weeks later. Mice that received this vaccine generated as many antibodies as mice that received two separate doses two weeks apart.

The researchers now hope to extend those intervals, which could make the particles useful for delivering childhood vaccines that are given as several doses over a few months, such as the polio vaccine.

“The long-term goal of this work is to develop vaccines that make immunization more accessible — especially for children living in areas where it’s difficult to reach health care facilities. This includes rural regions of the United States as well as parts of the developing world where infrastructure and medical clinics are limited,” says Ana Jaklenec, a principal investigator at MIT’s Koch Institute for Integrative Cancer Research.

Jaklenec and Robert Langer, the David H. Koch Institute Professor at MIT, are the senior authors of the study. Linzixuan (Rhoda) Zhang, an MIT graduate student who recently completed her PhD in chemical engineering, is the paper’s lead author.

Self-boosting vaccines

In recent years, Jaklenec, Langer, and their colleagues have been working on vaccine delivery particles made from a polymer called PLGA. In 2018, they showed they could use these types of particles to deliver two doses of the polio vaccine, which were released about 25 days apart.

One drawback to PLGA is that as the particles slowly break down in the body, the immediate environment can become acidic, which may damage the vaccine contained within the particles.

The MIT team is now working on ways to overcome that issue in PLGA particles and is also exploring alternative materials that would create a less acidic environment. In the new study, led by Zhang, the researchers decided to focus on another type of polymer, known as polyanhydride.

“The goal of this work was to advance the field by exploring new strategies to address key challenges, particularly those related to pH sensitivity and antigen degradation,” Jaklenec says.

Polyanhydrides, biodegradable polymers that Langer developed for drug delivery more than 40 years ago, are very hydrophobic. This means that as the polymers gradually erode inside the body, the breakdown products hardly dissolve in water and generate a much less acidic environment.

Polyanhydrides usually consist of chains of two different monomers that can be assembled in a huge number of possible combinations. For this study, the researchers created a library of 23 polymers, which differed from each other based on the chemical structures of the monomer building blocks and the ratio of the two monomers that went into the final product.

The researchers evaluated these polymers based on their ability to withstand temperatures of at least 104 degrees Fahrenheit (40 degrees Celsius, or slightly above body temperature) and whether they could remain stable throughout the process required to form them into microparticles.

To make the particles, the researchers developed a process called stamped assembly of polymer layers, or SEAL. First, they use silicon molds to form cup-shaped particles that can be filled with the vaccine antigen. Then, a cap made from the same polymer is applied and sealed using heat. Polymers that proved too brittle or didn’t seal completely were eliminated from the pool, leaving six top candidates.

The researchers used those polymers to design particles that would deliver diphtheria vaccine two weeks after injection, and gave them to mice along with vaccine that was released immediately. Four weeks after the initial injection, those mice showed comparable levels of antibodies to mice that received two doses two weeks apart.

Extended release

As part of their study, the researchers also developed a machine-learning model to help them explore the factors that determine how long it takes the particles to degrade once in the body. These factors include the type of monomers that go into the material, the ratio of the monomers, the molecular weight of the polymer, and the loading capacity or how much vaccine can go into the particle.

Using this model, the researchers were able to rapidly evaluate nearly 500 possible particles and predict their release time. They tested several of these particles in controlled buffers and showed that the model’s predictions were accurate.

In future work, this model could also help researchers to develop materials that would release their payload after longer intervals — months or even years. This could make them useful for delivering many childhood vaccines, which require multiple doses over several years.

“If we want to extend this to longer time points, let’s say over a month or even further, we definitely have some ways to do this, such as increasing the molecular weight or the hydrophobicity of the polymer. We can also potentially do some cross-linking. Those are further changes to the chemistry of the polymer to slow down the release kinetics or to extend the retention time of the particle,” Zhang says.

The researchers now hope to explore using these delivery particles for other types of vaccines. The particles could also prove useful for delivering other types of drugs that are sensitive to acidity and need to be given in multiple doses, they say.

“This technology has broad potential for single-injection vaccines, but it could also be adapted to deliver small molecules or other biologics that require durability or multiple doses. Additionally, it can accommodate drugs with pH sensitivities,” Jaklenec says.

The research was funded, in part, by the Koch Institute Support (core) Grant from the National Cancer Institute.