Transistors, the building blocks of modern electronics, are typically made of silicon. Because it’s a semiconductor, this material can control the flow of electricity in a circuit. But silicon has fundamental physical limits that restrict how compact and energy-efficient a transistor can be.

MIT researchers have now replaced silicon with a magnetic semiconductor, creating a magnetic transistor that could enable smaller, faster, and more energy-efficient circuits. The material’s magnetism strongly influences its electronic behavior, leading to more efficient control of the flow of electricity. 

The team used a novel magnetic material and an optimization process that reduces the material’s defects, which boosts the transistor’s performance.

The material’s unique magnetic properties also allow for transistors with built-in memory, which would simplify circuit design and unlock new applications for high-performance electronics.

“People have known about magnets for thousands of years, but there are very limited ways to incorporate magnetism into electronics. We have shown a new way to efficiently utilize magnetism that opens up a lot of possibilities for future applications and research,” says Chung-Tao Chou, an MIT graduate student in the departments of Electrical Engineering and Computer Science (EECS) and Physics, and co-lead author of a paper on this advance.

Chou is joined on the paper by co-lead author Eugene Park, a graduate student in the Department of Materials Science and Engineering (DMSE); Julian Klein, a DMSE research scientist; Josep Ingla-Aynes, a postdoc in the MIT Plasma Science and Fusion Center; Jagadeesh S. Moodera, a senior research scientist in the Department of Physics; and senior authors Frances Ross, TDK Professor in DMSE; and Luqiao Liu, an associate professor in EECS, and a member of the Research Laboratory of Electronics; as well as others at the University of Chemistry and Technology in Prague. The paper appears today in Physical Review Letters.

Overcoming the limits

In an electronic device, silicon semiconductor transistors act like tiny light switches that turn a circuit on and off, or amplify weak signals in a communication system. They do this using a small input voltage.

But a fundamental physical limit of silicon semiconductors prevents a transistor from operating below a certain voltage, which hinders its energy efficiency.

To make more efficient electronics, researchers have spent decades working toward magnetic transistors that utilize electron spin to control the flow of electricity. Electron spin is a fundamental property that enables electrons to behave like tiny magnets.

So far, scientists have mostly been limited to using certain magnetic materials. These lack the favorable electronic properties of semiconductors, constraining device performance.

“In this work, we combine magnetism and semiconductor physics to realize useful spintronic devices,” Liu says.

The researchers replace the silicon in the surface layer of a transistor with chromium sulfur bromide, a two-dimensional material that acts as a magnetic semiconductor.

Due to the material’s structure, researchers can switch between two magnetic states very cleanly. This makes it ideal for use in a transistor that smoothly switches between “on” and “off.”

“One of the biggest challenges we faced was finding the right material. We tried many other materials that didn’t work,” Chou says.

They discovered that changing these magnetic states modifies the material’s electronic properties, enabling low-energy operation. And unlike many other 2D materials, chromium sulfur bromide remains stable in air.

To make a transistor, the researchers pattern electrodes onto a silicon substrate, then carefully align and transfer the 2D material on top. They use tape to pick up a tiny piece of material, only a few tens of nanometers thick, and place it onto the substrate.

“A lot of researchers will use solvents or glue to do the transfer, but transistors require a very clean surface. We eliminate all those risks by simplifying this step,” Chou says.

Leveraging magnetism

This lack of contamination enables their device to outperform existing magnetic transistors. Most others can only create a weak magnetic effect, changing the flow of current by a few percent or less. Their new transistor can switch or amplify the electric current by a factor of 10.

They use an external magnetic field to change the magnetic state of the material, switching the transistor using significantly less energy than would usually be required.

The material also allows them to control the magnetic states with electric current. This is important because engineers cannot apply magnetic fields to individual transistors in an electronic device. They need to control each one electrically.

The material’s magnetic properties could also enable transistors with built-in memory, simplifying the design of logic or memory circuits.

A typical memory device has a magnetic cell to store information and a transistor to read it out. Their method can combine both into one magnetic transistor.

“Now, not only are transistors turning on and off, they are also remembering information. And because we can switch the transistor with greater magnitude, the signal is much stronger so we can read out the information faster, and in a much more reliable way,” Liu says.

Building on this demonstration, the researchers plan to further study the use of electrical current to control the device. They are also working to make their method scalable so they can fabricate arrays of transistors.

This research was supported, in part, by the Semiconductor Research Corporation, the U.S. Defense Advanced Research Projects Agency (DARPA), the U.S. National Science Foundation (NSF), the U.S. Department of Energy, the U.S. Army Research Office, and the Czech Ministry of Education, Youth, and Sports. The work was partially carried out at the MIT.nano facilities.

Google.org announced on Sept. 15 that the MIT Transit Lab is a recipient of $2.1 million in funding — one of only 15 projects selected in the worldwide Google.org Impact Challenge: AI for Government Innovation. The funding from Google’s philanthropic arm will support NGOs, social enterprises, and academic institutions as they integrate artificial intelligence-powered solutions across topics like health, resilience, and economy.

The Transit Lab’s winning project, the Public Transit Intelligence Hub (PTIQ), aims to unify public transportation agencies’ real-time monitoring, operations control, and passenger communication systems into a single centralized AI-orchestrated platform that will allow transit control center staff to make better-informed, on-the-spot decisions, and provide riders with more immediate and accurate information.

The control centers of public transportation agencies are similar in appearance to the portrayal of NASA mission control in movies: rooms filled with employees monitoring dozens of radio feeds and computer screens relaying real-time camera data about stations and their operations, transit vehicle locations, riders, traffic, and road conditions. Unfortunately, the information coming in is fragmented, rather than integrated into a centralized system with overall awareness of the network’s conditions. This system creates an intense work environment for the transit staff making operations and communications decisions that can affect thousands of passengers relying on transit to get them where they need to go.

“Public transportation agencies are required to make decisions around the clock regarding real-time operations, control, and passenger communication,” says Awad Abdelhalim, associate director of the Transit Lab, and PTIQ co-principal investigator, project director, and technical lead. “Our goal isn't to automate those decisions, but to make sure the people making them have the best information possible. By unifying and streamlining data and information flow from fragmented and siloed internal systems, PTIQ will improve the experience of both riders and the transit workforce.”

Jinhua Zhao, the MIT Class of 1941 Professor of City and Transportation, head of the MIT Department of Urban Studies and Planning, and founder and director of the MIT Mobility Initiative (MMI), is the other co-principal investigator on the project. The PTIQ program manager is MIT Lecturer Jim Aloisi, who directs the Transit Research Consortium, which will also work on the project. That consortium is comprised of researchers from the Transit Lab, MMI, and Northeastern University, where Professor Haris Koutsopoulos takes the lead. 

In addition to providing funding for the three-year project, Google.org will provide pro bono support from its own engineers and AI product experts.

"AI holds incredible potential to transform public services, but there is often a gap between promise and practice,” says Maggie Johnson, global head of Google.org. “By equipping the 15 selected organizations with funding and pro bono support from Google's own AI experts, we are empowering the people closest to the problem to show what is truly possible. Together, we can ensure that AI makes a profound, positive difference in the everyday lives of communities worldwide."

The project will build on the group’s decades of experience in applied-research collaborations with transit agencies in major metropolitan areas throughout the world. PTIQ’s decision support interface for control center staff will integrate predictive models, optimization engines, and large language model-based contextual reasoning. But ultimately the decision-making based on that information will be left to transit staff, who can better balance the trade-offs of making one decision over another in these often incredibly complex situations.

“The hard part of integrating AI in transit is not the technology; it’s the institution,” Zhao says. “AI is evaluated on benchmarks. Public transit is assessed in the control center and on the streets. Over decades of work with transit agencies in Washington, D.C., Chicago, London, Boston, Tokyo, and Hong Kong, we have learned to ask a different question. Not whether AI can do this, but whether it can work in the organization and whether the staff trust it. PTIQ is designed to ground AI in the institutional reality and behavioral nuances of a transit agency, and bring machine intelligence and human judgment into one place.”

“Currently the evaluation of AI models relies heavily on deterministic, objective tasks, such as solving mathematical equations or generating code,” Abdelhalim explains. “However, the vast majority of real-world operational tasks — like delivering public transit services — are highly dynamic, multi-stakeholder, and lack a single correct objective answer. These complex spatiotemporal environments are the ultimate testbed for evaluating what AI systems can add to society.”

PTIQ aims to transform the transit workforce experience, the transit rider experience, and the overall ability of transit agencies to efficiently respond to disruptions and unexpected events. 

“We expect that PTIQ will take what is largely a siloed environment and connect it in ways that provide powerful benefits for the agency workforce and its riders,” says Aloisi, who is also a former secretary of transportation for the Commonwealth of Massachusetts. “[Doing this by] improving response time, reducing platform and bus stop crowding, providing riders with higher quality and timely information, and supporting agency staff — from dispatchers to vehicle operators and communications staff — with high-quality, reliable, real-time information and solution sets.”

A new AI system that excels at challenging games with hidden information could someday help human decision-makers select ideal strategies to outfox opponents in complicated situations like military maneuvers.

Using advances in machine-learning, researchers from MIT, Carnegie Mellon University, New York University, and Stanford University developed an AI that defeated top-ranked human players of the board wargame Stratego by a large margin — something no AI system had been able to achieve. 

Stratego, a two-player game of imperfect information, in which the opponent’s piece identities remain hidden, is often used as a benchmark to test the strategic thinking abilities of powerful AI models.

To build their model, the researchers combined efficient training algorithms with new techniques tailored for calculated decision-making in hidden information settings. 

The AI system achieved greater performance at Stratego than the next best models, while being far cheaper and less computationally demanding to train. The system also outperformed top human players in other strategic games with different rules and designs, demonstrating how it can be generalized for a variety of use-cases.

The AI system could be adapted to help humans tackle many real-world problems with hidden information, such as business negotiations or cybersecurity. 

“In the kind of imperfect information tasks you would face in reality, you often don’t have the luxury of enumerating through all the possibilities. There are just too many. Having AI algorithms that are general purpose and can provably perform this challenging task so well is a big step forward,” says Gabriele Farina, an assistant professor in the Department of Electrical Engineering and Computer Science (EECS), principal investigator at the Laboratory for Information and Decision Systems (LIDS), and senior author of a paper on this AI system.

He is joined on the paper by lead author Samuel Sokota, a graduate student at Carnegie Mellon; Eugene Vinitsky, an assistant professor at NYU; Zico Kolter, a professor at Carnegie Mellon; Hengyuan Hu, a graduate student at Stanford; and Zhiyuan Fan, an EECS graduate student at MIT. The research appears today in Nature.

Hidden information

The world is full of imperfect information problems. 

In these interactions, some parties possess information others do not. For instance, traders in financial markets may not know the rationale behind the trades of others, while military forces likely don’t have full knowledge of enemy positions. 

With hidden information, the decisions parties make, as well as the decisions they choose not to make, are intertwined in such a way that it is extremely difficult to determine the best steps to take next.

“The more you bluff, the more your opponent expects it, and the less each bluff is worth. It’s not obvious how to reason about that,” Sokota explains. “It’s very different from a setting like chess, where the best move is still the best move no matter how often you’ve played it.”

Stratego is often used to model imperfect information situations. In this board wargame, which resembles military chess, players arrange 40 pieces on their side of a board and then move pieces across the board to capture their opponent’s flag. 

But the identity of all pieces remains secret until they collide, and then the lower-ranking piece is eliminated.

The possible piece configurations number more than 10 to the 66th power — an exponentially greater number than in chess — making Stratego extremely difficult for an AI system to play well. 

Past efforts, such as Google’s DeepMind, relied on sophisticated operations that were computationally demanding and costly. But even with millions of dollars in training costs, these models were still not strong enough to beat top human Stratego players.

“With Stratego, there is an explosion of possible universes you might have to deal with. AI techniques that were developed for games like poker definitely could not scale in this setting,” Farina says.

The MIT researchers set out to develop a full AI system that could achieve superhuman performance for less cost, which they called Ataraxos (a Greek word used to describe one who is unbothered or free from anxiety).

A two-pronged approach

To build Ataraxos, the researchers trained the model using a technique called self-play reinforcement learning. The model plays against itself many times to learn a strong “blueprint strategy” of how to excel at Stratego. 

They designed especially efficient algorithms, which enabled Ataraxos to learn much faster than prior methods while ensuring it didn’t get stuck trying to predict every possible move. This reduces training costs and boosts performance. 

“Our system reaches strictly higher playing strength than DeepNash (DeepMind’s system) while using less than one hundredth of the training examples and less than one thirtieth of the self-play games, indicating a massive improvement in efficiency,” says Farina.

During a game, Ataraxos uses the blueprint strategy as a starting point to set up the board and begin thinking about its next moves at each round of play. 

But before acting, it refines its choices on the fly using a technique called decision-time planning. The system employs a generative model that uses probabilities to estimate the likely identities of the opponent’s hidden pieces, then evaluates future choices before selecting the next move. 

“Rather than just guessing blindly, we use decision-time planning to find the most plausible state of the board. Using this generative model allows us to really zoom in on the specific board and opponent we are facing,” Farina says.

The innovative use of this generative model for decision-time planning was the missing piece that enabled Ataraxos to achieve superhuman performance.

Ataraxos beat the strongest Stratego player in the world by a record margin of 15-1-4 and achieved a 39-2 record against top human players at the Stratego world championship. “Ataraxos is good at calculating risk in a way that humans are not. A human might start freaking out if their most valuable piece is exposed, but the bot can be surprisingly composed. It doesn’t overcorrect and give away its secrets,” Farina says.

The researchers also adapted Ataraxos for other imperfect information games, including Barrage Stratego (a faster-paced variant with fewer pieces), Hanabi (a cooperative card game with many players), and Dou dizhu (a game in which two players cooperate against a third).

The system achieved superhuman performance in each instance, demonstrating the generality of this method.

In the future, the researchers want to build interpretability measures into Ataraxos so the system can explain its decision-making in a way that a human could understand. 

“Humans must have the final say in whether a recommendation is followed, so before adoption can happen, we need a way to audit the model’s decisions. We still have a long way to go, but I hope these algorithms can be the foundation for a lot more work to come,” Farina says.

This research is funded, in part, by the Office of Naval Research, the New York University Department of Civil and Urban Engineering, the C2SMART Center, the National Science Foundation, and a Schmidt Sciences AI2050 Early Career Fellowship.   

About 25 percent of lung adenocarcinomas have mutations of the gene KRAS, which drives uncontrolled cell growth. In recent years, the FDA has approved two KRAS inhibitors to treat patients with KRAS mutations. While these drugs can work well initially, tumors almost always develop resistance to them.

Usually, resistance emerges because cells reactivate KRAS activity, through mutations that prevent drug binding or by increasing KRAS expression that overpowers the effects of the inhibitor. However, in a new study, MIT researchers have modeled an alternative mechanism that cancer cells can use to become resistant to KRAS inhibition.

The researchers found that in some cases, lung tumors undergo transformation from adenocarcinoma to squamous cell carcinoma. Both of these tumor types are commonly found in the lungs, but they are thought to  arise from different cells and have different genetic profiles.

When this transition occurs, tumor cells no longer require KRAS, and appear to turn on alternative signaling pathways that help them continue to grow. Ongoing work to identify those pathways may reveal targets for new drugs that could help prevent resistance to KRAS inhibitors.

“The main takeaway is that there seem to be different routes of resistance to KRAS inhibitors, and so we need to be thinking about how we can address this,” says Carrie Rodriguez, an MIT graduate student and one of the lead authors of the paper.

Nicolas Mathey-Andrews PhD ’25 is also a lead author of the study, which appears today in the journal Nature Genetics. The paper’s senior author is Tyler Jacks, the David H. Koch Professor of Biology and a member of MIT’s Koch Institute for Integrative Cancer Research.

Tissue transformation

The two FDA-approved KRAS inhibitors both target a mutation called KRAS-G12C. These drugs are approved only for use in patients whose tumors have failed to respond to other drugs, and these patients usually have cancer that has spread beyond the lungs.

KRAS inhibitors are effective in about 35 percent of the patients who receive them. However, in those cases, the tumors almost always end up becoming resistant by generating additional copies of the KRAS gene or finding other ways to turn on the MAP kinase signaling pathway, which is usually triggered by KRAS and stimulates cell growth.

“Resistance to targeted therapies is a very serious problem,” Rodriguez says. “Sometimes these KRAS inhibitors can hold cancers at bay, but most cases do end up relapsing.”

A 2021 study from researchers at Dana-Farber Cancer Institute, which analyzed tumors from 17 non-small cell lung cancer patients treated with KRAS-G12C inhibition, identified secondary resistance mutations in a majority of patients. In two of these patients, however, the researchers found that tumors transformed from adenocarcinomas to squamous cell carcinomas, but they did not harbor obvious resistance mutations.

Both adenocarcinomas and squamous cell carcinomas are classified as non-small cell lung cancers (NSCLCs), which are the most common type of primary lung cancer. Adenocarcinomas, the most common type of NSCLCs, often originate from the surfactant-producing cells that line the lungs, while squamous cell carcinomas originate in the cells that line the central airways of the lungs.

Mutations of KRAS are found much more frequently in adenocarcinomas than in squamous cell carcinomas

In this study, the researchers set out to model the factors that might drive the transition from adenocarcinomas to squamous cell carcinomas. To do that, they engineered a mouse lung cancer model to express the mutation that is targeted by the FDA-approved KRAS inhibitors. 

Following treatment with a KRAS-G12C inhibitor, tumors with genetic loss of Nkx2-1, which normally helps maintain alveolar epithelial identity, were able to undergo adeno-to-squamous transition. Turning on a transcription factor called DeltaNp63, which is overactive in many squamous cell carcinomas, also made this transition more likely. Another transcription factor known as SOX2 also helped stimulate the transition, but this gene could not initiate the transition on its own.

Paths to resistance

Tumors that underwent these tissue transformations did not acquire the mutations that typically boost KRAS expression in adenocarcinomas. Instead, KRAS signaling was shut off. The researchers hypothesize that these cells may turn on another signaling pathway that helps them to continue growing.

“There seem to be several different routes where you can get to squamous transformation, either through loss of lung-lineage-defining transcription factors, or overexpression of these squamous master regulators, SOX2 or DeltaNp63. Those resistant squamous tumors no longer respond to KRAS inhibition because they shut off the signaling or at least dampen it significantly,” Rodriguez says.

The researchers are now further exploring what happens to tumor cells as they transition to a squamous state, in hopes of identifying vulnerabilities that could be targeted with new drugs.

“Fundamentally this is a transition that’s poorly understood, and we were happy to see that we were able to model it,” Mathey-Andrews says. “Future directions that have an eye toward translation will utilize those models to understand the process and conditions by which histologic transformation occurs, and then also nominate potential targets downstream.”

The research was funded, in part, by the Koch Institute Support (core) Grant from the National Cancer Institute, a Ruth Kirschstein National Service Research Award, the National Institute of General Medical Sciences, and the Ludwig Center at MIT.

J-PAL North America, a regional office of MIT’s Abdul Latif Jameel Poverty Action Lab (J-PAL), convened the second cohort of the Climate Action Learning Lab this spring with 17 climate leaders from four U.S. government agencies and nonprofit organizations. Participants then engaged in four months of programming designed to strengthen their skills in generating and using evidence, applying research insights to their own programs, and identifying promising policy-relevant interventions for evaluation.

The Climate Action Learning Lab first emerged in 2025 as a response to an urgent need for rigorous research on programs that seek to improve resilience to climate-related hazards and support the transition to a low-carbon economy. By helping participants identify interventions and develop evaluation plans, the Learning Lab aims to generate evidence about which approaches are most effective, and for whom. 

“We have a unique opportunity to embed rigorous evaluation into promising programs as they are implemented in order to accurately measure their impacts on emissions reductions,” says Peter Christensen, scientific advisor of the J-PAL North America Environment, Energy, and Climate Change Sector. “By developing rigorous evaluations, Learning Lab participants can better understand the behavioral mechanisms that drive program impacts and cost-effectiveness, and build an evidence base to inform more effective policymaking.”

Following the success of last year’s Climate Action Learning Lab, which led to multiple research collaborations and the launch of two randomized evaluations, J-PAL North America recruited a second cohort of leaders representing four organizations: the City of Boston’s Sustainability Office, the Hawaii Climate Change Mitigation and Adaptation Commission, the Oregon Department of Environmental Quality, and the Electrification Coalition. 

From May through August, the cohort engaged in a set of offerings including training on impact evaluation, learning how to formulate research questions, and assessing the generalizability of the existing evidence to their own contexts. Participants had the opportunity to explore J-PAL resources, including the recently released Climate Action Evidence Review, which synthesizes existing evidence and highlights important gaps where more research is critical to inform effective, equitable climate action. 

“It was exciting to learn where the gaps in research are and where we have an opportunity to lead. The experience reinforced that we are addressing issues that have not been widely studied, and having access to researchers’ expertise was incredibly valuable,” says Ana Paola De La Vega, from the City of Boston’s Environment Department. Throughout the Learning Lab, the city explored a potential evaluation of its Boston Energy Saver program to better understand its impacts on energy cost savings and energy efficiency incentive uptake among small businesses.

Members of the Climate Action Learning Lab put theory into practice through personalized strategy sessions, where they examined potential programs for evaluation. Researchers from the J-PAL network joined select sessions to advise organizations on which programs to prioritize, considering factors such as research feasibility and existing evidence gaps. Each participating organization selected a target program and explored potential randomization approaches. 

“The Climate Action Learning Lab provided our team with a valuable opportunity to catalog our projects and better understand what makes a program ready for evaluation. It also helped us identify which initiatives are the strongest candidates for rigorous evaluation and how to prioritize them,” says Leah Laramee, Hawaii climate change mitigation and adaptation coordinator. During the Learning Lab, the team assessed three potential programs for evaluation and selected a rebate finder that connects residents with climate-related programs, with a focus on understanding its impact on program uptake, particularly among vulnerable communities.

In August, J-PAL North America hosted a virtual summit to celebrate Learning Lab participants as emerging champions for evidence in the climate space. During the event, cohort members presented their priority research questions and strategic evaluation plans and received feedback from researchers and peers. These presentations highlighted the progress made throughout the engagement and provided an opportunity to discuss next steps for advancing organizations’ evaluation plans beyond the Learning Lab.

“Overall, the Learning Lab has provided our team with a strong foundation to think critically about our own programming when applying for grants, setting up new projects, and determining how to assess impact from the beginning. This knowledge could ultimately help us determine what approaches the Electrification Coalition can take in future policies and programs,” says Ashley Blackwell, deputy director at the Electrification Coalition. 

Although formal Learning Lab programming has concluded, J-PAL North America will continue supporting organizations interested in launching a randomized evaluation through partnership development with researchers and potential funding opportunities. This Learning Lab cohort will join J-PAL North America’s Climate Action Community of Practice, alongside longtime partners and participants from the inaugural Learning Lab cohort. Together, Community of Practice members will continue to exchange ideas, build connections, and explore evidence-informed approaches to mitigation and adaptation strategies.

To learn more about J-PAL North America’s work in the energy, environment, and climate change sector, including our full range of activities, resources, and partnership opportunities, visit the Evidence for Climate Action Project webpage.

Swimming can take a lot of muscle. But as MIT engineers have found, even a single layer of muscle cells can power through water if designed right. 

In a paper appearing today in the journal Advanced Functional Materials, the team presents a design for a thin, muscle-powered swimming robot. The “skeleton” of the aquabot is made from a film of gel that is about the length and width of a stick of gum. The two halves of the gel form the “fins” of the bot. Each fin is covered with a layer of live muscle cells that is much thinner than a single strand of hair. The cells are genetically engineered to twitch in response to light. 

When the researchers shine light on one fin, the muscles on its surface twitch in response, causing the whole fin to flap with enough force to pull the robot through water. By flashing light on one fin or the other, at various intervals, they can control the swimming robot’s direction and speed. 

The engineers showed that the paper-thin bot could swim and swivel through a simple watery maze. At its fastest, the robot can swim a distance of about four times its body length in one minute. That’s a snail’s pace compared to Olympic swimmers, who can cover up to 65 body lengths per minute. But the bot could hold its own against more leisurely swimmers like the cow shark, which explores the ocean at about the same rate.

“It takes a lot of force to move through water versus air,” says study author Ritu Raman, associate professor of mechanical engineering at MIT. “The robot’s quite strong, given its size.”

The new robot is the first example of a very thin, two-dimensional, muscle-powered robot capable of locomotion. 

“Currently, biohybrid robots from our group and others’ are built from bulky, 3D chunks of lab-grown skeletal muscle that require millions of cells to fabricate,” says Raman, who notes that thinner, less bulky designs such as the team’s new bot could be cheaper to build and could move more efficiently. “We believe that biohybrid robots powered by living muscle could one day perform delicate jobs like exploring environments too fragile or unpredictable for conventional hardware, because living tissue is soft, responsive to its surroundings, and can heal itself.”

The study’s MIT co-authors are first author Maheera Bawa, Arielle Berman, Laura Schwendeman, Ferdows Afghah, and Seanbiron Johnson.

Maximizing movement

Last year, Raman’s group developed an iris-inspired disk of artificial muscle tissue. They stamped a disk of gel with a pattern of concentric and radial grooves, and deposited live muscle cells onto the gel’s surface. The cells formed a thin layer that grew along the grooves, and when stimulated with light, the cells moved in patterns that stretched and squeezed the disk, similar to how a human iris dilates and constricts the eye’s pupil. 

That work was the first to demonstrate that muscle cells could be grown in a very thin layer, and in complex patterns that when stimulated could move in multiple, controllable directions. 

“People hadn’t seen this muscle architecture engineered from scratch before,” Raman says. “And the cells were moving in multiple directions. But they only moved about 100 microns. From a robotics perspective, their movements were tiny.” 

In their new work, the team aimed to maximize muscle movements to produce more force — enough, say, to power a swimming robot. The key, they found, was to optimize the skeleton on which the cells grow. 

In their previous iris-inspired design, they grew muscle cells on fibrin, which is a type of ultrasoft gel that the team realized can quickly shrivel in response to the forces generated by the muscles. To better support cells and maximize their force, the researchers focused on tuning the underlying gel by changing three properties: the gel’s composition, its stiffness, and the size and shape of the grooves that are stamped into it. 

“For engineering any type of tissue, it’s known that these are knobs you can tune,” Raman says. “And we wanted to optimize all these parameters to support live muscle cells.”

Tuning a skeleton

To find an optimal “skeleton” on which to grow muscle cells, the team experimented with multiple formulations of gel, of different stiffnesses, and stamped with grooves of different geometries. For instance, one groove type resembled a skinny square trough, where another was more of a long curved valley. They found that when they deposited muscles onto each type of grooved gel, cells settled into alignment in grooves that were more square than curved. More aligned cells tend to fuse into fibers that then form a stronger, more coordinated muscle tissue. Square grooves, they found, were the way to go. 

Instead of using fibrin, they tried gelatin methacrylate (GelMA), a material that is often used in tissue engineering. They made different recipes of GelMA to create skeletons of different stiffnesses and observed how muscle cells grew when deposited on the gel’s surface. They found that cells grew in better alignment, and produced the most force, on stiffer gels. 

The team also varied the gel thickness and found that a half-millimeter-thin film of GelMa offered good support for a single layer of muscle cells. The film was light enough that the cells were able to stick to the gel when they contracted, rather than peeling away. 

Finally, the team “exercised” the muscles, using a training routine of flashing lights to strengthen the muscles. 

With the pumped-up cells and the optimized gel, the team designed a thin, two-finned robot, comprising the gel, stamped on both sides with square-bottomed grooves and lined with muscle cells. The cells fused into fibers, eventually forming a strong, aligned muscle tissue. 

“You can think of the robot as having two independent muscles,” Raman says. “If we shine a light on just one, only that muscle moves. If shining on both, they both flap.”

The researchers submerged the robot in a large petri dish of water and manually maneuvered a light source over the bot. The robot followed the light, flapping its fins in response to navigate through a maze that the team placed in the dish.

The current design is relatively basic as far as its form. The researchers intended first to show that the bot could produce enough force to swim.

“Our next goal is to optimize the body design to enable faster swimming,” Raman says. “But even at slow swim speeds, one could imagine a muscle-powered swimmer being used for purposes like environmental monitoring in aquatic environments.”

This research was supported, in part, by the Office of Naval Research. 

AI tools are increasingly replacing human judgements in some settings. For instance, resume screening algorithms are often used in hiring, where they may improve efficiency and consistency in decision-making.

But some scholars have raised concerns that the adoption of automated systems could eventually result in one algorithm being used to make all decisions in a particular industry. They worry so-called algorithmic monoculture could have negative consequences.

For example, in hiring, the thinking goes that algorithmic monoculture might result in systematic exclusion — a situation in which a job candidate rejected by one firm’s algorithm would likely also be rejected by every other firm’s algorithm.

However, MIT researchers now argue that algorithmic monoculture may not always be as bad as some scientists have suggested.

They systematically evaluated major objections to algorithmic monoculture, including systematic exclusion, and concluded this and many other arguments either fail or aren’t decisive against all forms of monoculture. 

Instead, they mathematically prove that monoculture tends to create informational echo chambers that can hinder exploration. In hiring, this could make it less likely that the best candidates would get jobs — however, bundling various hiring algorithms into a single “ensemble” can overcome this limitation, the researchers show. This could sometimes enable monoculture to perform as well as, if not better than, a polyculture where different firms use different algorithms.

“A trend toward algorithmic monoculture is a realistic scenario, and a really important issue that is being brought about by the use of AI, but it is hard to say in the abstract whether monoculture would be a bad thing. It depends on the details, like the domain we are talking about and the accuracy of the algorithm itself,” says study co-author Brian Hedden, a professor in the Department of Linguistics and Philosophy, who holds an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science (EECS) and is also a principal investigator in the Laboratory for Information and Decision Systems (LIDS).

Hedden is joined on the paper by co-author Manish Raghavan, the Drew Houston (2005) Career Development Professor at the MIT Sloan School of Management and in EECS, as well as a LIDS principal investigator. The research appears in Philosophical Perspectives. 

The move toward monoculture

Algorithmic monoculture, in which all decisions across a certain domain are made using the same algorithm, is not a new phenomenon. 

For instance, lending decisions were once made by independent bankers at individual banks, but now all bankers use the same information based on a borrower’s standardized credit scores, which are derived from the Fair Isaac Corporation (FICO) algorithm. 

Similarly, a handful of resume screening algorithms are commonly used by many Fortune 500 companies.

“The worry is that, as more people use AI and algorithms to get information and make decisions, there is more of a vehicle for this kind of correlation to occur,” Raghavan says.

To better understand this issue, he and Hedden joined forces to systematically assess the promises and pitfalls of algorithmic monoculture. They focused on hiring, but their approach could apply to other domains, such as lending. (They note, however, that other domains, like generative AI content creation or AI-guided scientific research, may work differently, and that monoculture in some of these domains may be more problematic.) 

The researchers began by exploring one common objection to algorithmic monoculture: that reliance on the same algorithm will systematically exclude certain people from opportunities. 

This could occur in hiring because, if one firm screens out an individual’s resume, that applicant will likely face the same bad luck at each firm.

But after systematically evaluating this argument using a series of models that capture multiple situations, the researchers argue it isn’t compelling since the overall number of people hired is not affected by the fact that firms use the same algorithm. 

Rather, algorithmic monoculture could improve bargaining power of job candidates.

“All the jobs get filled and the same number of people have jobs, but the firms are fighting over the same pool of candidates, which actually drives up wages,” Raghavan says.

They also explored objections related to agency. For instance, if a job candidate applies for a job and their resume is forwarded to every firm using the hiring algorithm, the candidate never gets a chance to adjust their resume to improve their chances.

“This seems like a good objection to bad forms of monoculture. But if you have a monoculture where you get to revise your resume and resubmit your materials, then this doesn’t hold up,” Hedden says.

On the flip side, monoculture could enable individuals to game the system. For instance, if having one’s resume in a certain format leads to a better outcome, job candidates could simply reformat their resumes to improve their chances.

“But it is not obvious that having one algorithm would incentivize this kind of gaming more than having a bunch of different algorithms used by different firms,” Hedden says. “In the latter scenario, you might just target a couple of firms’ algorithms and try to game them, giving yourself a bit of advantage with a few employers.”

The wisdom of crowds

They also considered a less-explored objection: that monoculture can increase homogenization of information.

Based on the “wisdom of crowds,” a theory from social psychology, a diverse group of independent decision makers can outperform a single person, Hedden explains. 

In hiring, this means that having firms with diverse hiring algorithms can lead to a higher-quality pool of new hires. 

Algorithmic monoculture could also cause candidates with the same characteristics and credentials to be hired every time by every firm. This may prevent firms from discovering candidates who may be better alternatives.

“Monoculture might inhibit the amount of discovery that happens overall. It is not clear if that is a bad thing, but it is definitely a worry when we think about designing AI for applications like science, art, or writing,” Raghavan says.

This problem could be mitigated by building randomness into a monocultural platform, which could induce a higher level of exploration, he adds.

In addition, the performance of monoculture depends on the algorithm. If a single algorithm is much more accurate than the many algorithms used by different firms, monoculture may be better system. 

One way to boost performance may be to package multiple firms’ hiring algorithms into one “ensemble algorithm” that could assign each job candidate a score based on the average. 

By conducting a series of simulations of different hiring situations, the researchers confirmed that such an “ensemble algorithm” could sometimes outperform the use of multiple algorithms.

However, it remains to be explored how feasible this kind of algorithmic “ensembling” would be in practice, Hedden says.

“A lot of the answers around the promises and pitfalls of algorithmic monoculture are going to be contextual. Even from a research perspective, there is still a lot of work to be done to figure out how we can approach these concerns from an empirical perspective,” Raghavan says.

Ultimately, the researchers hope this work inspires additional research about the long-term consequences of algorithmic monoculture, as well as studies that focus on the real-world complexities involved in a complex system like a job market.

Brian, a middle-aged financial consultant, spends his day sitting in front of three screens. Two of them involve his job. The third features a chatbot, which he has given a woman’s name and frequently uses. Brian spent years on the road for work earlier in his career, has never married, and recently broke up with a woman who said he was emotionally unavailable. What did Brian do, in response? He asked the chatbot if the woman had a point. 

“It’s an extraordinary moment,” writes MIT Professor Sherry Turkle, who interviewed Brian (not his real name) while conducting her research for her new book about chatbots. “Brian asks an object with no emotions to tell him if he is emotionally withholding: ‘Speak to me about what you cannot experience.’ As soon as he asks the program to talk about intimacy, he’s asking it to punch above its weight.”

And yet, this kind of thing seems to be happening a lot today. 

“People think the empathy of a machine is what empathy is, then turn away from the people in their lives because they’re not empathetic enough,” Turkle says. “We’re getting ourselves in a position where human beings are too much work, right at the moment when we have never needed other people more.”

After all, what is a chatbot? It’s a computer program predicting the most plausible next string of text in a conversation, based on massive amounts of data fed into it. 

“What a chatbot does is offer pretend empathy,” Turkle says. “After it tells you how much it loves you and how much it understands you, it doesn’t care if you kill yourself or cook some pasta.”

Sadly, chatbots have in fact been associated with high-profile cases of teen suicide as well, sometimes after teens get drawn into extensive dialogues with chat tools.

Turkle explores this terrain in a new book, “Artificial Intimacy: Who We Become When We Talk to Machines,” published today by Little, Brown and Company. After surveying evidence and conducting new research, she emphatically concludes that chatbot use, while it may often feel like a short-term salve, is broadly detrimental in terms of human development and social connectivity.

“We’re doing ourselves a tremendous disservice at every moment in the life cycle,” says Turkle, the Abby Rockefeller Mauzé Professor of the Social Studies of Science and Technology at MIT.

The inner history of technology

A sociologist and clinical psychologist by training, Turkle is a longtime faculty member in MIT’s Program in Science, Technology, and Society. In books such as “The Second Self” (1984), “Life on the Screen” (1997), and “Alone Together” (2011), she has evaluated the interplay of technology, psychology, and personal identity. In “Reclaiming Conversation” (2015), she documented the costs of texting and using social media. 

“I feel the story I’m telling is the inner history of technology,” Turkle says. “Not just what it does, but what it does to people.”

Turkle structures “Artificial Intimacy” around the stages of a human life and explores the implications of chatbots for each one. In her interviews with children, Turkle finds a persistent blurring of the line between chatbots, people, and other devices.

One 8-year-old uses the same phone to talk to their grandparents and to ChatGPT, regarding them all as “things you reach on your phone.” A weary graduate-student mother who uses a chatbot for bedtime stories says her daughter thinks the chatbot is “a person in the phone, absolutely.” 

In this sense, chatbots may be interfering with even the most basic childhood processes of distinguishing people from inanimate objects. Turkle finds many additional problems with the use of chatbots as ever-present entertainment for children, noting that a certain amount of time alone helps the development of imagination and inner resources. 

“Children can’t learn trust from a device that not only lies but doesn’t know when it lies,” Turkle writes. “They can’t develop the capacity for solitude that enables both a sense of self and the capacity for mutuality.”

All told, in affecting the ability of children to develop, the dangers of chatbots are “existential,” Turkle believes. 

Facing reality, but understanding the appeal

Adults don’t fare much better when they use chatbots heavily, Turkle says. “Artificial Intimacy” explores case after case of grown-ups who become dependent on chatbots as well: people going through divorces or estrangements who want dialogue, students looking for advice about applying to college or graduate school, workers who enlist ChatGPT to do assignments, and more. They start using chatbots, keep using chatbots, and before long seem less interested in human interaction, and less capable of it. 

“Once people are involved with a connection with a robot, they start to think that it’s alive, that it cares about them, that it loves them,” Turkle says. 

That means people can lose or fail to build their own capacities for new thought, and opt out of dealing with the real world in all its maddening-but-affirming complexity.

“We’re de-skilling ourselves as relational beings,” Turkle says. “We’re also de-skilling ourselves in work. We’re de-skilling ourselves in personal relationships. It’s across the board.” And speaking of a practice multiple people in the book have attempted, she says, “When we build chatbot avatars of dead relatives to keep them alive, we can lose our ability to mourn.”

We are not, in her estimation, doing the hard work of figuring other people out, seeing things from different points of view, and putting in the effort to build on those things and make society better.

“We’re starting to define being human as not doing the work,” Turkle says. 

For all of that, a considerable amount of “Artificial Intimacy” involves understanding why people engage with chatbots. After all, as Turkle writes in the book, “Chatbots always agree with us and affirm us. They always offer us their full, undivided attention.” And once that happens, things seem to just go from there. 

People, meanwhile, can be prickly, demanding, and moody. As Turkle says: “People say they prefer the chatbot because their husband or wife says, ‘Dear, you took out the garbage, but also do the dishes, and clean up the kitchen while you’re at it.’ And a chatbot just says, ‘Oh, you’re so wonderful.’” 

She adds: “This technology is offering something people really want, which is to feel less vulnerable. And it offers it over and over again. You don’t want to have to ask somebody out? Don’t want to offer condolences? At every opportunity, technology says, ‘Why don’t you do this thing that’s less hard.’” 

And as Turkle acknowledges in the book, there is a shortage of services such as mental health care providers in the U.S., where only about half of people have access to one, according to a federal government study she cites. Chatbots are helping to fill this void, for better or worse. 

Don’t fear the friction

Turkle also notes that the age of social media has likely led people to have fewer in-person friendships and interactions, a problem that chatbots are now aiming to address. 

“You take away people’s capacities, then you offer technology as the cure for the problems technology caused in the first place,” Turkle says.

In light of all this, given Turkle’s dim view of chatbots amid the spread of AI, what is actually to be done to restore human interaction? For starters, Turkle thinks, we need to face up to the idea that life is not frictionless — and that’s okay.

“Everything in the life cycle is about facing up to friction and fears and stress and tension, and understanding that friction is not a bad thing,” Turkle emphasizes. Developing the capacity to overcome difficulties is an essential part of life. In trying to sidestep the hard work of being social beings, she thinks, personal technology is enfeebling us.

Beyond that, Turkle envisions pushback against chatbots similar to the movement against, say, having phones in schools or letting young people use social media. 

“I hope my book is part of a larger and larger movement,” Turkle says. “I don’t want to be alone. I want to be part of a movement pushing back.”

Other writers in this domain have praised “Artificial Intimacy.” Journalist Nicholas Carr, author of “The Shallows,” has stated it “will help you avoid the profound but often hidden threats AI poses to you and your relationships.” Psychologist Jonathan Haidt of New York University, author of “The Anxious Generation,” has stated that Turkle’s “groundbreaking research and beautiful writing make her the most qualified member of Team Humanity to call us back to our senses, and to each other.” 

For her part, Turkle says, “I wanted to write a book that college students would read, that high school seniors, juniors would be able to read, that parents would pick up and not be intimidated by, that teachers would read.” 

And “Artificial Intimacy” offers a recurring question for those readers.

“If not a richer life in the real,” Turkle writes, “what’s our endgame?” The purpose of life, she underlines, is not to avoid it, but to tackle it head on. 

“It’s one thing if you say, my teen is texting too much,” Turkle says. “It’s very different if your 2-year-old is thinking a plushy toy with a chatbot inside is their best friend. Because you are getting into the intimate infrastructure of what makes a person develop. Social media came for our attention. Chatbots come for our capacity for attachment. That’s toxic on another level. What is the endgame? That the baby will prefer chatbots to people who are much more complicated and hard? Is that who we want to be?”

The world needs more wind energy. But anyone designing new wind turbines or trying to squeeze more power out of existing ones faces a stiff challenge testing new approaches. That’s because the atmosphere is a tough place for a controlled experiment.

Some researchers use wind tunnels to conduct tests, but scaled-down wind turbines in traditional wind tunnels can differ widely from conditions in the field. (Wind turbines are the largest rotating machines ever made.) The problem hampers not only development of better wind turbines, but also our understanding of basic questions like how much power to expect from a turbine when winds change direction.

In a new open access paper published in PNAS Nexus, researchers closed the gap between experiments in the field and the lab by using a wind tunnel that features high pressurization to simulate the flow physics in the atmosphere. With this approach, the researchers determined how the alignment of the turbine and its tip speed relative to the wind influence the power it generates, offering new insights into how to get more power from existing wind farms.

They also used the approach to validate a computationally lightweight model that engineers can use to test different turbine designs and wind farm control strategies.

Together, the researchers estimate that optimizing the turbine alignment relative to wind, the blade pitch angles, which control the airfoil’s angle of attack, and the tip speed relative to the wind could potentially result in tens of thousands of dollars per turbine every year in additional revenue.

“The immediate impact of this study is that we’ve now both improved and validated models that go into wind turbine control protocols for existing farms,” says Michael Howland, MIT’s Jeffrey Cheah Career Development Professor. “The bigger, medium-term impact, with a much larger upside, is this new experimental paradigm to rapidly prototype, validate simulation models, and test hypotheses about better designs and control strategies much faster than has been possible before.”

Joining Howland on the paper are first author John Kurelek, an assistant professor at Queen’s University; MIT PhD candidates Ilan Upfal and Kirby Heck; Queen’s University postdoc Supun Pieris; Penn State University researcher Alexander Piqué; and Princeton University Professor Marcus Hultmark.

Answers in the wind

Howland has spent years developing models to simulate wind farm performance and developing new techniques to increase their power output. In 2022, he showed that accounting for the wake of individual turbines when controlling the entire wind farm could significantly increase power output.

But that work required his research team to first conduct a lengthy field experiment that temporarily resulted in lowering a real wind farm’s power output by intentionally misaligning turbines from the wind for months to better understand their performance in misalignment.

“Wind energy is a uniquely challenging problem to study experimentally,” Howland says. “We want to test the effect of a certain change in isolation, but wind farms operate in chaotic, turbulent environments where the weather is constantly evolving. Wind turbines have to react to weather conditions that we have no control over, and that introduces complexities in identifying the impact of the imposed change we are studying. The field sits at this unique intersection between environmental flow, mechanics, aerodynamics, and meteorology.”

The difficulty of running experiments at real wind farms has left researchers and engineers unsure of how changes in the alignment between the wind and turbine or factors like the turbine’s tip speed relative to the wind change power output.

In fact, the researchers say many predictive models people use are built on the assumption that turbines are always perfectly perpendicular to the wind. That’s rarely the case in the real world, even with modern turbines that gradually adjust their angle in response to the wind’s rapid directional changes.

“People have been debating which models are best for understanding the output from these wind farms, but if you have nothing to compare them against, it’s very difficult to advance the field,” Hultmark says. “This paper tries to do both of those things.”

Hultmark’s research lab at Princeton has pioneered the study of scaled-down wind turbines in pressurized wind tunnels, which, as previous studies have shown, better reproduce large-scale turbines in the atmosphere because pressure makes air more dense, resulting in more inertia within the scaled laboratory environment. For the new study, the researchers used a turbine measuring 15 centimeters in diameter at varying pressures of up to 240 atmospheres.

“By pressurizing the chamber, we’re testing a turbine that is, all else being equal, 15 to 20 meters in diameter, with the ability to go up to 35 meters in diameter,” lead-author Kurelek explains. “That’s because we’re increasing the density by a factor of 100 to 220 times,” 

Kurelek sent the dimensions of the wind tunnel and wind turbine setup to Howland, who used them to calculate the aerodynamics, forces, and power production using a newly developed unified wind turbine model, which builds on previous work that developed a more general aerodynamic theory for wind turbines. The new model enables the researchers to simulate wind turbine performance across operating conditions without relying on empirical corrections that have historically been used in wind power models.

The researchers then ran a series of experiments in the tunnel over the course of several weeks, testing the turbine’s performance at different wind alignments and with different control strategies, to isolate how each factor affects performance.

They found power output could be significantly increased by adjusting the turbine’s tip speed based on its misalignment angle with the wind — a control strategy that is rarely employed in wind farms today but could offer a way to boost performance with minimal added costs.

“The big output of the experiments was clearly showing that new power maximums can be achieved when the turbine becomes misaligned with the wind through only changes to the tip speed,” Kurelek says.

Scaling the approach

The study served as validation for Howland’s model, which is fast enough to be run by engineers designing and operating wind turbines around the world using regular laptop computers.

“What we really want to know is if the turbines are always operating in some degree of misalignment with the wind, how should we control the turbine to get the maximum achievable power production?” Howland explains. “Our unified momentum model was able to make predictions of how to do this control a few years ago, and this is the first time we’e able to experimentally validate that model.”

Howland says validating models is only one part of the paper’s potential impact.

“This study also shows the huge opportunity to perform these high-throughput, controlled experiments in the pressurized facilities that Marcus and John work with, enabling us to achieve the right physics but in a time efficient and low-cost manner,” Howland says. “Right now, there’s a massive gap between idealized theoretical and simulation models and full-scale testing in extremely complicated field environments. Nothing is filling that gap except for these pressurized experiments. I hope this can be an enabler to investigate a huge range of unanswered wind energy questions in controlled environments.”

The work was supported in part by the Natural Sciences and Engineering Research Council of Canada; the National Science Foundation; and the MIT-GE Vernova Alliance.

RNA vaccines, which have been proven effective against Covid-19, are now being developed for many other diseases, including cancer. One of the drawbacks to these vaccines is that they require ultracold storage, but researchers from MIT have found a promising way to overcome that limitation.

With help from an AI algorithm, the researchers tweaked the formulation surrounding the lipid nanoparticles that are typically used to deliver mRNA vaccines, making the vaccines more heat-resistant. Using this approach, they formulated vaccines that could remain stable even when stored at room temperature for up to a year, or at nearly 100 degrees Fahrenheit for two months.

When Covid-19 vaccines carried by these particles were administered to mice, they generated just as strong an immune response as an RNA Covid-19 vaccine similar to one developed by Moderna. By using the AI algorithm to predict the optimal formulations for the particles, the researchers were able to cut down the number of experiments they needed to do, which rapidly sped up the development process.

“The real beauty of this algorithm is that we can use it with small data sets,” says Ana Jaklenec, a principal investigator in MIT’s Koch Institute for Integrative Cancer Research. “It’s really hard to run thousands of experiments, so this algorithm allows us to more easily achieve formulations with features that we want — in this case, stability.”

Jaklenec and Robert Langer, the David H. Koch Institute Professor, are the senior authors of the paper, which appears today in Nature Biotechnology. Graduate student Jinbi Tian and postdoc Khanh Tran are the lead authors of the paper.

Stable vaccines

RNA is a highly fragile molecule, so researchers stabilize it with lipid nanoparticles (LNPs) that protect the RNA from degradation and help it get into cells. However, these RNA-LNP vaccines still need to be kept cold (-20 to -80 degrees Celsius), which makes it difficult to ship them to regions that don’t have cold-storage facilities available.

Making these vaccines more heat-tolerant would not only enable them to be distributed more widely, but could also help researchers develop new vaccines that could be administered through novel methods such as microneedle patches. These patches contain hundreds of vaccine-filled microneedles, which dissolve when the patch is applied to the skin, releasing the vaccine.

To create more stable RNA vaccines, researchers have experimented with adding a variety of excipients — sugars, salts, or polymers — to the LNPs. Jaklenec and Langer recently developed polymer-stabilized LNPs that can withstand higher temperatures, but those LNPs were slightly different from the FDA-approved formulations that were used for the Moderna and Pfizer Covid-19 vaccines. 

In their new paper, the researchers wanted to see if they could find a way to make those FDA-approved formulations more stable at high temperatures.

They began by reusing some of the excipients that had worked in their earlier efforts, but they were “really getting stuck,” Jaklenec says. “We were trying to use and screen excipients that we’ve previously used successfully to stabilize LNPs, but it just wasn’t working. It was really frustrating for the team.” 

To speed up their progress, the researchers decided to try a machine-learning approach. Working with researchers at MIT’s Computer Science and Artificial Intelligence Lab (CSAIL), they developed an algorithm that can make predictions based on very small datasets.

“We’d used our algorithms for various automated experimental design applications before, but never on a biological problem like vaccine stability,” says Mina Konaković Luković, an assistant professor of electrical engineering and computer science in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), who is also an author of the paper. “It was surprising to see how quickly the algorithm converged on a stable formulation — getting there in just a handful of iterations, rather than the exhaustive search that would normally be required.” 

The researchers used this algorithm to analyze nearly 50 FDA-approved excipients. For each excipient, the researchers first measured how well it stabilized RNA when incorporated into an LNP. They used these particles to deliver mRNA encoding a protein called firefly luciferase, which produces bioluminescence, into cells. By measuring how much light was emitted by the cells, the researchers could determine how effectively each excipient protected the mRNA.

The researchers chose five of the most promising excipients and used their AI algorithm to predict ratios of those excipients that would best stabilize LNPs similar to those used by Moderna. Based on those predictions, the researchers tested two formulations at a time in cells, fed those results back into the algorithm, and generated more predictions. After several rounds, they chose one formulation that appeared promising enough to test in animal studies.

This process took only a few weeks, much less than it would have taken without guidance from the AI algorithm.

“Before we implemented the AI algorithm, we spent several months testing different combinations and also doing the prescreening of all the excipients that we could find, but nothing would get us to 100 percent stability,” Tran says.

Robust immune responses

To test the heat resistance of their new LNP formulation, the researchers used the particles to package Covid-19 mRNA antigens, then dehydrated them using a process called vacuum drying. These particles were then stored at 37 degrees Celsius (98 degrees Fahrenheit) for two months, or at room temperature for one year. Mice that were vaccinated with these particles, even after long-term storage, showed equivalent immune responses to mice that received vaccines carried by LNPs similar to the original Moderna formulation.

The researchers also used their new heat-resistant formulation to create solid microneedle patches that could deliver a SARS-CoV-2 antigen. These patches generated an immune response similar to that produced by the injectable RNA vaccines.  

“Our approach broadens the application of not only mRNA vaccines, but also therapeutics or advanced drug-delivery platforms like controlled-release particles or microneedle patches, which requires the formulation to either be in solid state or to be stable at higher temperature,” Tian says. 

The researchers also showed that they could use their algorithm to stabilize other LNP formulations, including one similar to those used by Pfizer to deliver its Covid-19 vaccine. This formulation uses the same excipients as the one the MIT team developed for the Moderna LNP, but in a different ratio. For each LNP, once a heat-resistant formulation has been developed, it could be adapted to deliver any type of mRNA payload, the researchers say. 

The research was, in part, funded by the Gates Foundation.

When MIT mechanical engineering student Daphne Wang arrived for her internship at the Muscogee Creek Nation Department of Health in Tulsa, Oklahoma last summer, she expected to be writing code. To her surprise, she found herself analyzing tribal history and the implications for technical innovation. 

“I learned so much that I think should be required curriculum for every single person in this country,” says Wang. “I think tech for good is genuinely getting out into communities and learning about people who aren't you, creating worldly perspectives that you ultimately can take into everything else that you do.”

Wang and fellow intern Lucy Sun, who is majoring in artificial intelligence and decision making, spent the summer supporting the development of the Muscogee Nation's first Pregnancy Risk Assessment Monitoring System (PRAMS) survey to capture maternal health data specific to Muscogee women. The project required completing literature reviews, analyzing datasets, and drafting survey questions, while considering issues of data sovereignty and governance structures specific to the Muscogee.

“It was this change of pace — 180 [degrees] from MIT, completely,” Wang reflects. At MIT, she says, course instructors routinely prepare students for big technical challenges, but there are fewer opportunities to “work with communities, [learn] how to think about the cultural and historical dimensions of the work, or how to consider the people who will be affected.” 

Wang’s supervisor, epidemiologist Breanna McNaughton-Long, was effusive about Wang and Sun’s engagement. “They were instrumental. They did so much work that I don't think we would have been able to do as quickly,” she says. “And they brought this sense that we were doing something that mattered.”

Experiential learning at the confluence of humanities and engineering 

Wang was one of 10 MIT undergraduates to participate in the PKG Center for Social Impact’s 2026 Code.Tulsa program. Now in its second year, Code.Tulsa is made possible by support from the Patrick J. McGovern Foundation and the George Kaiser Family Foundation. 

Students live at the University of Tulsa for the summer. They intern with the Muscogee and Cherokee nations, as well as local nonprofits Black Tech Street and Urban Coders Guild, completing AI, data science, and other technical projects. 

The PKG Center’s assistant dean for community-based programs, Vippy Yee, complements students’ professional experience with education on the historical and cultural significance of tech-based development in the Tulsa region from the perspective of local leaders. 

“As with all PKG Center programming, our aim is to help students integrate an engineer’s approach to problem-solving with a humanist lens on the nature of social challenges, and by extension interventions,” says Alison Badgett, the PKG Center’s director. 

“Getting to speak to people who were either very educated on the issues or have experienced them has been a massive help in reshaping the way that I think about social issues,” says Chenise Harper, an electrical engineering with computing major who interned with the Urban Coders Guild. 

Harper puts this dynamic candidly: “I hated history, but I learned a lot about why it is useful for making social change and how to go about researching it. I now feel a lot more confident that I can make an impact.”

A student’s vision for increasing STEM access

Code.Tulsa was the brainchild of sophomore Jack Carson, an electrical engineering and computer science (EECS) student who approached the PKG Center with the idea for Code.Tulsa as an incoming first-year student. Carson, who is from Tulsa and a member of the Cherokee Nation, saw firsthand that rural Native students who could benefit greatly from STEM education were unlikely to have access to it. Carson proposed developing a weeklong STEM camp for members of federally recognized tribes held at the University of Tulsa, with he and Code.Tulsa interns serving as instructors. 

Carson devised a nomination system to attract promising high school students, selecting 25 campers from 10 tribes out of 160 applications representing 25 Native nations. 

To develop the curriculum, Carson enlisted Harvard University student Allie Zong, whom he had met at Campus Preview Weekend. The Tulsa Advanced Sciences Camp (TASC) offers intensive, interactive track time in AI engineering, DNA technology, physics, and chemistry and energy. This is complemented by philosophical workshops to help students think more ambitiously and deliberately about what they can and should achieve in the near and long term. As Zong explains, “We teach them not knowledge, but curiosity, and the skills to teach themselves."

The TASC experience “kind of propelled me to self-study calculus,” says Alyssa Theofanidis, a rising high school senior from Houston who is a citizen of the Cherokee Nation. Theofanidis participated in the camp’s first year, returning this summer as a teaching assistant. Like many TASC campers, Theofanidis is eager to use her developing STEM expertise to benefit her Native community. 

TASC “got us to think 1,000 times more ambitiously about the impact we could have,” said another camper, who was inspired by guest speakers “like the nuclear fusion person at the top of their field,” referring to physicist Alexandra LeViness of MIT spinout Commonwealth Fusion Systems, who helped develop TASC’s chemistry and energy track. Campers were also inspired by “the different worldviews” of MIT interns. “I thought, maybe this is what MIT looks like,” said one camper, with several expressing the intent to apply to MIT as a result of the camp. 

Cherokee Nation Principal Chief Chuck Hoskin Jr. also delivered remarks at TASC, encouraging students as future leaders to take a public interest in technology. 

“Technology can be used for good, or it can be used for harm. Your generation has the opportunity to bend that arc toward something good,” Hoskin said. “It's a very special relationship that the Cherokee Nation has with MIT. As we reach out a hand in friendship, we have a hand reaching back. We are thankful for Cherokee citizen Jack Carson, a former secretary for the Cherokee Nation tribal youth council, for his leadership role in organizing this effort.”

Making a positive long-term social impact 

Like TASC campers, Code.Tulsa interns came away from the experience motivated to make a positive impact. While most MIT students won’t go on to full-time professional roles traditionally associated with social impact, PKG Center programming like Code.Tulsa helps students recognize they can promote the public interest no matter their career. As Elvis Chipiro, a junior in computer science and engineering, reflected after interning with the Cherokee Nation, “Social impact is not separate from mainstream technology; rather, it is embedded in the choices engineers make every day … Ultimately, meaningful social change is not driven by technology alone, but by people willing to design systems with empathy, responsibility, and inclusion at the center.”

For others, Code.Tulsa helps them reconnect with a sense of public purpose. “Remembering that … I could use my MIT education to help others was a big reason I applied to MIT in the first place,” says Harper. “But I forgot my own mission in the stress of school. Code.Tulsa really re-opened my eyes to why I am here.”

RNA vaccines and other nucleic acid therapeutics are typically packaged within fatty molecules known as lipid nanoparticles (LNPs). MIT researchers have now come up with a way to produce these particles much more quickly, and with more precise control over their size and shape.

This new process, which can be run automatically without any human intervention, could greatly speed up the development of new RNA and DNA therapeutics, the researchers say. 

“This method can help you determine what are the parameters that will generate specific size and shape attributes, before you take those particles and see which one will perform best,” says Cedric Devos, an MIT postdoc and one of the lead authors of the study. “It could be a quite powerful development tool.”

Current methods for designing new lipid nanoparticles often require time-consuming trial-and-error experiments, with limited investigation of desired particle size and shape. Tuning the size and shape of LNPs can open the possibility of targeting different organs and tissues.

“The size and shape of LNPs could not be reliably controlled by any previous production method. The problem may appear simple at first glance, but in reality it requires a deep understanding of lipid nanoparticle assembly,” says Allan Myerson, a professor of the practice in MIT’s Department of Chemical Engineering and the senior author of the new study.

MIT postdocs Aniket Udepurkar and Peter Sagmeister are also lead authors of the paper, which appears today in ACS Nano. 

More precise control

Vaccines based on mRNA work by delivering instructions to cells to produce a harmless version of a viral protein, which prompts the immune system to generate a response. However, if mRNA is injected on its own, it will be quickly broken down in the body.

“These are really a revolutionary type of therapeutics, but they need some kind of delivery vehicle to bring them to the right cells in the body,” Devos says.

For the mRNA Covid-19 vaccines and other mRNA therapeutics, scientists have used lipid nanoparticles for that packaging. LNPs usually consist of four components — an ionizable lipid, a phospholipid, cholesterol, and a lipid attached to a molecule of polyethylene glycol (PEG), which helps to stabilize the LNP.

To make the particles, two streams of fluid are mixed together at high speed. One consists of lipid molecules suspended in ethanol, and the other contains mRNA dissolved in an acidic buffer.

Those solutions aren’t mixed at equal flow rates, however. Instead, there is about three times more of the mRNA solution than the lipid solution. This unequal ratio helps promote the formation of lipid nanoparticles containing mRNA, but it doesn’t offer precise control over the size or shapes of the particles.

In a study published last year in ACS Nano, the MIT team showed that they could enable much better control of particle size by breaking the mixing process down into two steps. 

In the first step, mRNA and lipids are mixed together at equal flow rates. Then, after a short delay, more of the buffer is added, which halts the growth of the particles. Longer delays produce larger particles. 

“This gives you the ability to play around with the residence time, which is the time it takes between the first mixer and the second mixer. If you keep that residence time really long, it means your particles will grow a lot. If you keep it really short, you can keep them really small,” Devos says. “It gives you a lever over lipid nanoparticle manufacturing that wasn't available before.”

In that study, the researchers also showed that they could alter the shape of the particles. By changing the concentration of the buffer that they add in the second step, they were able to transform the particles from spheres to elongated particles that resemble avocados.

Both of these interventions — changing the delay residence time and changing the buffer composition — allow the researchers to control particle size and shape without otherwise altering the composition of the LNPs.

An automatic process

In the new ACS Nano paper, the researchers developed a way to automate the two-step mixing process. They also incorporated a commercially available dynamic light scattering device that can measure the sizes of the particles as they are formed.

The project also highlights the impact of MIT’s Undergraduate Research Opportunities Program (UROP). “Through this program, applied mathematics and computer science undergraduates Joy Ren, Sofiya Chubich, and Dylan Nguyen gained hands-on experience, learning how advanced software engineering can be integrated with chemical engineering to develop an automated platform for LNP process development,” Sagmeister says.

With this automated system, the researchers can specify a particle size, and the system will generate that size. It will also measure the resulting particles to make sure they’re the right size, and if not, adjust the delay time and other factors to steer them to the right size.

“The first paper really unlocked the new methodology to make lipid nanoparticles, to truly engineer them by size and shape,” Sagmeister says. “With the second study, we automate the whole process.”

The system can also produce particles of different shapes, but measuring those shapes has to be done outside of the automated system. 

Using this approach, the researchers were able to investigate how changing the inputs to the system affects the size and shapes of the particles, on a much faster timescale than currently possible. The researchers then used the data they gathered from these experiments to train a machine-learning model that can predict the combination of factors that will generate a particular size or shape.

With this method, it could be much easier for developers of RNA therapeutics to generate different sizes of particles to test them for a particular application. Controlling the size of a lipid nanoparticle is critical because the particle’s size determines where in the body it is most likely to end up. 

“If you make an LNP-based therapeutic with a target size of 150 nanometers, and one that is 70 nanometers, and everything else is the same, they will behave very differently,” Devos says.

The researchers have filed for a patent on their technology and are now working to commercialize it through a new company called BIZON Labs. Since receiving initial support through the Martin Trust Center for MIT Entrepreneurship’s Researcher 2 Entrepreneur (R2E) program, the team has also been accepted into MIT’s flagship accelerator program, delta v. The research was funded by the U.S. Food and Drug Administration and the Koch Institute Support (core) Grant from the National Cancer Institute. The work was carried out, in part, through the use of MIT.nano’s facilities.

Many people who are successful in STEM fields were lucky to have someone who turned them on to a topic and encouraged their interest. For Faith Reyes, that person was her high school physics teacher, Ms. Bolster. 

Whereas Reyes had earlier studied math and science without seeing how those subjects could be applied outside the classroom, her teacher helped her connect those dots and experience the excitement of investigating the physical world. With Ms. Bolster’s help, Reyes started a physics club where students would gather before school and conduct experiments. 

“I just sort of fell in love with physics then,” Reyes says. “And luckily, as I began to study it more and more, I found I landed exactly where I wanted to be.”

Reyes has carried that enthusiasm for physics into not only her research but her role as a personal tutor and teaching assistant at MIT, where she works to foster the same love of the subject in first- and second-year undergraduate students.

As an experimental particle physicist and sixth-year PhD student in the Formaggio Group in the Laboratory for Nuclear Science, Reyes studies neutrinos, elementary particles that have vanishingly little mass and rarely interact with other matter. Neutrinos are produced during radioactive decay, including the processes taking place inside nuclear reactors. Because they interact so infrequently, detecting them can be difficult. (We can’t feel them, but neutrinos from the sun are streaming through our bodies every minute.) Their unusual behavior makes them valuable to physicists seeking to understand what lies beyond the Standard Model, the framework that describes many of the fundamental particles and forces in nature.

“The Standard Model is extremely accurate and describes most of everything that we see,” Reyes says. “But it’s not complete.”

Reyes is a member of the Ricochet neutrino experiment, an international collaboration studying neutrinos produced by a nuclear reactor at the Institut Laue-Langevin in Grenoble, France. The experiment seeks to observe coherent elastic neutrino-nucleus scattering, a low-energy interaction in which a neutrino scatters off an atomic nucleus.

Through Ricochet, scientists aim to investigate some properties of these elusive particles, and contribute to, as Reyes puts it, “just fundamentally understanding the world in which we live.”

When looking back at her time in graduate school, Reyes’ path has not always followed the plan she initially envisioned.

When she joined Ricochet, she expected to work on a particular project located at MIT. But a few years into her PhD, it was clear that the project would not be ready within her timeline. Reyes instead shifted her focus to work taking place in France, where she began learning the technical details of the experiment’s detectors.

Her first visit lasted three months. At the time, Ricochet had two detectors, and Reyes spent much of her time performing the routine work required to understand how they operated.

As the experiment expanded to nine detectors and eventually 18, the amount of work required to manage the system grew substantially. Reyes and a colleague recognized that many of the repetitive tasks could be automated.

Together, they developed a software framework that could perform much of the low-level analysis and detector monitoring that Reyes had initially carried out manually.

The project became an important part of her development as a physicist. By working closely with the detectors and helping build tools to manage them, Reyes gained a detailed understanding of the experiment’s operations.

“It’s sort of like you’re building your own stuff to replace yourself,” she says. “Which is nice in a way because you can save yourself a lot of time.”

The opportunity was both validating and humbling. As a graduate student, she had moved from learning the basics of the experiment to helping guide the work of other scientists.

“It felt like my collaborators trusted me, and I had something of value to give to the collaboration,” Reyes says.

The people she has met through MIT and the Ricochet collaboration have been among the most rewarding parts of her graduate experience. Students, mentors, and collaborators have helped her think critically and become a better physicist, she says.

Her increasing leadership responsibilities have also changed how she approaches research.

Earlier in her academic career, Reyes says, she was more comfortable being told what to do than proposing her own scientific ideas. Over time, leading projects and coordinating groups pushed her to become more confident in her judgment.

“I think I was sometimes not really standing up for myself,” she says. “But now I feel more confident in my position and my prowess as a physicist.”

That confidence has become one of the most important lessons of her PhD.

After completing her doctorate, Reyes hopes to continue conducting research. She is considering a postdoctoral position, which would allow her to continue working in physics at another institution.

Her time in France has also influenced her vision of the future. Reyes spent nine months there through the Chateaubriand Fellowship, following two earlier three-month visits. While the latest trip was primarily focused on research, working in the same office as her collaborators made it easier to coordinate across time zones and strengthened her connection to the experiment. She also grew fond of the country’s culture and work-life balance and would even consider living there. 

“I fell in love with France and the people,” she says. “And also, the work culture.”

Outside the lab, Reyes enjoys playing video games and crocheting, a hobby she picked up during her time in France. She often crochets while watching movies or television, appreciating the opportunity to work with her hands while thinking about other things.

For a physicist whose work involves investigating some of the universe’s smallest and most elusive particles, the hobby offers a different kind of satisfaction: creating something tangible.

As Reyes moves toward the next stage of her career, she hopes to continue pursuing the questions that first drew her to physics. Her research may help reveal what lies beyond the Standard Model, but her experience at MIT has also shown her how much science depends on collaboration, adaptability, and the confidence to lead.

“I’ve learned a lot of lessons,” Reyes says. “Especially about wrangling people.”

When people reach out during a mental health crisis, a top priority for counselors is identifying those with a high risk of suicide. The distressed person’s language holds critical clues, and a new tool developed by scientists at MIT’s McGovern Institute for Brain Research is designed to pick up on and rapidly evaluate those signals.

The language-processing tool was developed by Daniel Low, a former graduate student in Senior Research Scientist Satra Ghosh’s Senseable Intelligence Group who is now a research scientist at the Child Mind Institute, where he leads its AI, Risk, and Contemplative Science Lab, as well as a visiting scholar at Harvard University. It uses a custom-built list of words and phrases linked to 49 suicide risk factors, searching text for these and using them to estimate an individual’s risk.

Ghosh, Low, and colleagues report today in the Journal of Psychopathology and Clinical Science that their tool accurately predicts suicide risk from text conversations with crisis counselors. It is already helping to clarify which suicide risk factors matter most in times of crisis. With more validation, it could help with risk assessment in clinical settings and crisis-support situations.

Identifying key risk factors

Suicide attempts are notoriously difficult to predict. Dozens of risk factors have been linked to suicide, and even trained clinicians struggle to identify who will make an attempt among those who have some form of suicidal ideation. Among the factors that can make suicidal thoughts and behaviors more likely are certain psychiatric symptoms and disorders, like depression, borderline personality disorder, and post-traumatic stress disorder, as well as environmental and social stressors, like poverty, incarceration, discrimination, and loneliness.

“You see all these 50 risk factors, and they're all interacting in ways we don't really understand,” Low says. “Many different pathways could lead to someone feeling they want to escape their internal pain,” he says — and it’s challenging to know whose path will lead to a suicide attempt or death.

Ghosh and Low wanted to understand which risk factors counselors and clinicians should most look out for during a mental health crisis. To do that, they collaborated with the Crisis Text Line, whose trained volunteers provide confidential text-based support to people in distress.

Crisis Text Line, a global mental health nonprofit that provides free, 24/7, confidential mental health support for people in need, provided specialized training and controlled access to this restricted dataset. The researchers analyzed de-identified texts from approximately 16,000 conversations with Crisis Text Line’s volunteer crisis counselors. Based on Crisis Text Line’s assessments, those conversations were grouped into three different risk levels: non-suicidal, suicidal ideation without imminent risk, and imminent risk. It was this imminent risk group — those with a plan for suicide, or who have an intent to die within the next 48 hours — that the researchers most wanted to understand.

“We wanted to know what type of symptoms predict the highest suicide risk,” Low says. This question has been studied before, he says — but typically through epidemiological surveys that ask a person to recall their symptoms and experiences, often after their mental health crisis has passed. In contrast, he says, “Crisis Text Line gives us an opportunity to assess many different symptoms and potential risk factors as people are having the crises.”

Reading between the lines

Before analyzing the crisis line texts, the research team built a suicide-risk lexicon. They turned to artificial intelligence to generate a preliminary list of words and phrases tied to established suicide risk factors, including factors associated with suicidal ideation, suicide attempt, and suicide death. Then they manually reviewed and curated that list. Their final lexicon includes about 60 words or phrases for each of 49 risk factors, with the relevance of each one confirmed by expert clinicians.

Then they trained a machine learning model to search the crisis conversations for words and phrases in their lexicon and use these to predict suicide risk. Because the lexicon links each word or phrase to a specific risk factor, they could use these data to determine which risk factors are most closely tied to imminent risk among people in crisis.

What they found was consistent with patterns found in previous research, although not always intuitive. For example, depression is a well-known risk factor for suicidal ideation, but their model found that mentions of lethal means and substance use were more likely to be expressed by the highest-risk group than depressed mood or fatigue. Expressions of active suicidal ideation and self-injury were also strong predictors. Intermediate predictors included anxiety, post-traumatic stress disorder, and emotional pain.

The predictive model assigns a weight to each risk factor based on its contribution to risk. For example, mentions of lethal means for suicide, like “cut” or “pills,” are weighed heavily, whereas terms related to hopelessness, like “don’t know what to do” or “hopeless,” contribute to a lesser degree. After training their model, the team found they could use it to accurately predict risk severity in new conversations the model had not previously seen.

One limitation of lexicons, the researchers note, is that they do not consider the context of terms, and they can miss terms that are similar to those in the lexicon, but not explicitly included. Large language models have reasoning abilities, and Low and colleagues have developed ways of using large language models to detect suicide risk in other projects. However, they say they often use their lexicon in parallel to guarantee flagging certain terms, as well as to maintain data privacy.

Low stresses that while the team used the power of a large language model to develop its lexicon, its prediction model is a simpler, “lightweight” model. Unlike large language models, which require massive computational power, it can be run easily on a personal computer, reducing both cost and privacy concerns. Just as importantly, it is interpretable: Rather than merely generating a risk estimate like some deep learning models can do more effectively, it tells users how it got there. Words of concern can be flagged so users understand the basis for each assessment and act on that information. They are working on similar explainability approaches with large language models.

That’s critical, because the stakes are so high. “This is such a complex space that having a human in the loop is, I think, going to be critical for a long, long time,” says Ghosh, who is the director of the Open Data in Neuroscience Initiative at the McGovern Institute. Likewise, the researchers add that any predictive model must be thoroughly validated before clinical use, and might need to be continually refined to keep up with changes in language use or target populations.

Because a reliable lexicon opens doors to new ways of understanding mental health, Ghosh and Low are widely sharing not just their suicide risk lexicon, but also the software package they developed to build it. Researchers can use that tool to efficiently build lexicons for other mental health conditions. Meanwhile, Low says, the suicide risk lexicon is already being used to explore how text data from a variety of sources, from social media to electronic health records, might help researchers and clinicians better estimate risk.

Most colon cancer deaths are caused by the spread of tumor cells beyond the colon, usually to the liver. In a new study, MIT biologists identified a cellular pathway necessary for colorectal cancer metastasis.

The pathway they identified, controlled by a protein known as YAP1, is normally involved in tissue repair. When activated in cancer cells, it promotes cell proliferation and migration. The researchers also found that a high-fat diet is more likely to turn on this pathway, through the production of fatty molecules called ceramides.

Drugs that block ceramide production could offer a new way to help prevent metastasis in patients diagnosed with colon cancer, the researchers say.

“We’ve found a pathway that we think is druggable. If we shut down the enzymes that make ceramides, tumor cells can’t switch on this regenerative program, and they largely fail to seed metastases in the liver,” says Omer Yilmaz, director of the MIT Stem Cell Initiative, a professor of biology at MIT and a member of MIT’s Koch Institute for Integrative Cancer Research. He is also a gastrointestinal pathologist and director of translational research in pathology at Beth Israel Deaconess Medical Center.

Yilmaz, Nilay Sethi, an associate professor of medicine at Harvard Medical School and Dana Farber Cancer Institute, and Alpaslan Tasdogan, head of the Institute for Tumor Metabolism and a professor in the Department of Dermatology at University Hospital Essen and the German Cancer Consortium (DKTK), are the senior authors of the study, which appears today in Science. MIT postdocs Swagata Goswami, Qiming Zhang, and Abdullah Burak Yildiz are the paper’s lead authors.

A hijacked pathway

In the United States, colon cancer is usually diagnosed at stage 2 or 3 — before the cancer has spread. However, even after successful surgery, up to a third of these patients will relapse with metastatic disease.

While scientists have identified many genetic mutations that drive the development of colon cancer, it’s unknown exactly what prompts them to spread beyond the colon. 

“Many studies have looked for a genetic driver of metastasis and come up empty,” Yilmaz says. “There isn’t a defining mutational signature that separates metastatic cells from the primary tumor, which points to metastasis being driven largely by changes in which genes are switched on and off, rather than by new mutations.”

In this study, the researchers sought to identify epigenetic programs that enable colon cancer cells to metastasize. Using tumor organoids from mouse models of several types of colon cancer and from patients with colorectal cancer, they found that metastatic cells shared one key feature: activation of the YAP1 program.

YAP1 is a protein that works with partner factors to switch on genes related to development, stem cell maintenance, and regeneration. In normal tissue, it is active during fetal development, and after injury, to promote healing.

In the gut, that repair response runs through a rare, fetal-like cell type, which normally appears only briefly to rebuild the intestinal lining after damage. YAP1 has been linked to cancer for years, but the new work shows that diet-derived lipids push tumor cells into this specific regenerative state — and that the state itself is what licenses metastasis.

“The regenerative program that we described is generally observed in the gut when there is severe injury or infection and the gut needs to regenerate. We see the tumor cells hijack this program to drive metastatic progression,” Goswami says.

Activation of this set of genes helps cancer cells to break free from the original tumor site and spread to other locations in the body. For colon cancer, the most common site of metastasis is the liver, followed by the lungs.

In mouse studies, the researchers also found that cancer cells in animals fed a high-fat diet turned on YAP1 to a greater extent than mice fed a healthy diet. A high-fat diet, the researchers found, triggers activation of enzymes that produce ceramides, a type of lipid. Ceramides then release the molecular brake that normally keeps YAP1 inactive, allowing it to move into the nucleus and switch on its target genes.

Preventing metastasis

The researchers showed that genetically targeting YAP1, or the genes involved in ceramide production, markedly reduced the spread of colon cancer to the liver in mice.

To determine if YAP1 is also involved in metastasis in humans, the researchers analyzed RNA sequencing data from patients with colorectal cancer. They found that YAP1 was more active in metastatic cancer cells, and that patients with higher body mass index (BMI) showed higher expression of the genes activated by YAP1 than normal-weight patients. Patients with higher levels of those genes also had lower survival rates.

“We don’t think that the YAP1 program is specific to obesity. It’s just that it becomes accentuated in obesity, and that may account for why obesity is known to drive the progression of colorectal cancer,” Yilmaz says.

They now plan to develop drugs that inhibit two of the enzymes involved in ceramide production, DEGS1 and DEGS2, in hopes that such drugs could help prevent colon cancer metastasis.

The researchers caution that the findings do not yet translate into dietary advice for patients who have already been diagnosed, and that any drug targeting ceramide synthesis will have to clear a high bar for selectivity, since these lipids are also essential in healthy tissues.

The research was funded by the National Institutes of Health/National Cancer Institute, the MIT Stem Cell Initiative, a Koch Institute Frontier grant, and the NRW Junior Research Program.

One of the main reasons that ovarian cancer is among the deadliest forms of cancer is timing: When doctors catch it early, the five-year survival rate can be north of 90 percent. But when doctors catch it late, in stages 3 or 4, five-year survival is less than half that.

About 20 years ago, researchers studying ovarian cancer discovered that many cases of high-grade serous ovarian cancer, the most common type, originate in the fallopian tubes. Detecting the disease there remains challenging, in part because its precursor lesions can be microscopic and difficult to sample.

Now researchers in the group of MIT Professor Kripa Varanasi, working with colleagues at MIT and Johns Hopkins University, have developed a handheld device capable of gently collecting living cells from specific locations to test for ovarian and many other types of cancer. The researchers believe the technique could one day be used to catch cancers earlier and more effectively. It could also be used to create treatments based on individual patient samples.

In a study describing the system in the journal Device, the researchers showed their system enables targeted sampling of newly excised tissue, and they used it to recover living cells for cultivation and testing. The device holds a small microfluidic channel against the tissue and uses a syringe to drive fluid through the channel, applying a force parallel to the tissue surface to gently detach living cells from tiny sections of tissue.

“We wanted to collect living cells from specific regions of the fallopian tube while leaving the surrounding tissue intact,” says Varanasi, senior author of the study and the Maher A. Elmasri Professor of Mechanical Engineering. “Once we have these living cells, there are many things we can do with them. We can use them for diagnostics, grow them into organoids, and build living models of disease. Ultimately, this could allow us to test how an individual patient’s cells respond to different treatments and help us develop more personalized medicines.”

Joining Varanasi on the paper are co-first authors Domitille Avalle SM ’25, MIT postdoc Bert Vandereydt PhD ’26, and Sean Parks ’20, SM ’24. The other authors are MIT PhD candidate Huaiyao Peng; Rebecca Stone, the Johns Hopkins University School of Medicine Stoddard and O’Neil Professor in Gynecologic Oncology; and Angela Belcher, MIT’s James Mason Crafts Professor and a professor of biological engineering and of materials science and engineering.

Living cells for ovarian cancer research

The discovery that many high-grade serous ovarian cancers originate in the fallopian tubes has opened up new prevention options for women at increased risk, who can have their fallopian tubes removed after childbearing years, largely preserving normal hormone production.

Stone, a gynecologic oncologist at Johns Hopkins University, has long advocated for this procedure for certain women at increased risk of ovarian cancer. Belcher introduced Stone to Varanasi, and the three, together with other collaborators, received funding from Break Through Cancer, a foundation that brings together interdisciplinary teams to tackle some of the most challenging problems in cancer. Their project focuses on developing new approaches for the early detection of ovarian cancer, with the cell-collection technology forming one part of that broader effort. 

The researchers began by asking whether they could collect living cells from specific regions of removed fallopian tubes to study the disease’s earliest stages.

"The idea was to see if we could find early signals from precancerous regions of concern,” Varanasi recalls.

The process traditionally involves placing surgically removed fallopian tubes in a chemical preservative and cutting the tissue into sections. The preservative maintains tissue structure, but the cells are no longer alive and cannot be grown in culture. A pathologist then looks for cancerous or precancerous regions in thin sections of the tissue under a microscope. 

“It’s very time-consuming and destructive to the cells,” Varanasi says. “We wanted to bring new capabilities to pathology, so we can not only see what these cells look like, but also collect them alive and study how they behave.”

The MIT researchers saw the process firsthand while visiting surgeons in the operating room at Johns Hopkins.

“It inspired us,” Varanasi says. “We do a lot of work on fluid-surface interfaces in my lab, and we realized we could use a fluid instead of a scalpel or brush, because when you flow a fluid it applies shear stress at the interface. We thought it could work because we heard from surgeons that cells in some locations were loose and would come off during routine washing and other procedures.”

“This is exactly the kind of problem that benefits from bringing clinicians and engineers together,” Stone says. “We understand the clinical need, while the MIT team brings a very different perspective from fluid mechanics and engineering. That combination allowed us to approach the problem in a new way.”

The researchers’ new approach uses a 3D-printed microfluidic device that forms a vacuum seal with the tissue. The device confines liquid flow to a small region, where the flowing liquid creates shear stress that gently detaches living cells.

“We came up with this device where one syringe creates a vacuum that holds it against the tissue, and a second syringe pushes liquid through it,” Vandereydt says. “The vacuum creates a seal, so nothing leaks, and then we locally apply what is basically a microfluidic chip on the tissue that gently shears the cells off.”

The researchers showed they could tune the shear stress applied to the tissue and compared their approach to other cell collection workflows. They found the cells collected using their technique remained viable and grew in culture much more readily than cells detached using conventional approaches.

Finally, the researchers tested their device on fresh human fallopian tube samples, which required them to be on call for sample shipments from their collaborators at Johns Hopkins. After experiments, the samples were shipped back for conventional pathology.

“The samples could come at any time. Sometimes, we’d get an email from our collaborators at 11 p.m. saying ‘There are two fallopian tubes coming tomorrow,’” Vandereydt says. “We were able to collect living cells from those fallopian tubes and turn those into organoids, which is important for testing, disease modeling, and eventually developing personalized treatments.”

“We are developing optical approaches to identify suspicious regions of tissue, and this technology could allow us to collect living cells from exactly those locations,” Belcher says. “Being able to first see where the disease may be emerging and then collect those cells for further study could be very powerful.”

From device to diagnostic

The researchers tested the device on different types of cells and found the approach can be tuned to collect cells of all types by applying different levels of shear stress.

“It’s agnostic to the disease,” Vandereydt says. “There are very loosely adherent prostate cancer cells that detach at 1 pascal [of stress], but if you look at bone cancer cells, only a few cells detach under as high as 5 pascals of stress.” 

The researchers plan for the early use of their device to involve tissue that has already been removed from the body, as that offers an easier pathway to regulatory approval. But they would also like to see their device used to swab samples inside of patients for easier testing and earlier cancer detection.

“What is exciting about this technology is the ability to collect living cells from a specific area while preserving the tissue for pathology,” Stone says. “In the future, one could imagine integrating it into routine histopathology workflows, creating a powerful new way to study carcinogenesis and fundamental biology directly from human tissue.”

Varanasi credits Break Through Cancer for enabling the project.

“Break Through Cancer brought together people working on not only ovarian cancer but also on pancreatic cancer, brain cancer, leukemia, and other cancers,” Varanasi says. “What we heard again and again is how valuable it would be to have better ways to obtain living cells from specific regions of tissue.”

The researchers hope that by making it possible to collect living cells from precise locations without removing or destroying the surrounding tissue, their approach could eventually help researchers and clinicians identify disease earlier and better understand how it develops.

“If this work can ultimately help women by enabling earlier detection of ovarian cancer, I would find that incredibly fulfilling,” Varanasi says. “That is really what motivates us — taking the science and engineering we develop in the lab and using it to make a difference in people’s lives.”

The work was supported by the Break Through Cancer Foundation.

A few months ago, researchers from the MIT Senseable City Lab published a study about pollution in New York City featuring some new methods. For instance: With machine learning, they identified the types of vehicles appearing in 331 traffic cameras in the city, and estimated the emissions coming from each automobile. Given enough cameras, these visual artificial intelligence techniques could monitor emissions with an unprecedented combination of precision and scale. 

For that matter, visual AI today can address all kinds of questions for urban planners. Why exactly is traffic snarling? What are the most dangerous aspects of different intersections? Which parts of plazas or parks attract the most people?

Across cities, more images means more data, more insight — and more concerns about privacy and fairness. 

“We can treat these digital images as data and quantify features of the city,” says Fábio Duarte, an MIT researcher and co-author of a new book about visual AI and urban studies. “With computer vision techniques, each image is a dataset.” Still, he adds, “We have to be careful about it.” 

And while urbanists have long used visual analysis to inform their thinking, now it’s possible to an unprecedented extent. 

“Everybody has been observing the urban environment and trying to get some insight,” says Martina Mazzarello, an MIT scholar and a co-author of the new book. “But what if we can do that at a large scale and get some insight everywhere?”

The scholars explore these topics in “How AI Sees the City: Urban Visual Intelligence,” published this month by Routledge. The authors are Duarte, a principal research scientist and associate director of the MIT Senseable City Lab; Mazzarello, a research scientist and lead of MIT Senseable City Lab global initiatives; Carlo Ratti, a professor of the practice and founder and director of the MIT Senseable City Lab; and Fan Zhang, an assistant professor at the Institute of Remote Sensing and GIS at Peking University.

“Great urbanists such as Kevin Lynch and Willian H. Whyte showed us the extraordinary value of ‘looking’ at the city,” Ratti says, referring to two prominent thinkers about city dynamics whose work is described in the book. “Today, visual AI gives us new ways to build on that tradition — allowing us to observe cities at a scale and with a level of detail that was previously impossible.” 

New tool, long tradition

“How AI Sees the City” stems from the work of the MIT Senseable City Lab, founded in 2004, which uses data to better understand urban dynamics. As the authors discuss in the book, there is a long history of visual representations that shape the way we think about cities, from Romans building marble maps to the introduction of photography — which produced influential urban images about things like Hausmann’s reshaping of Paris or the crowding of tenements in New York City’s Lower East Side during the 19th century.

More recently, some scholars have used visual studies to better understand city form, including Lynch, a former MIT professor whose 1960 book, “The Image of the City,” influenced many scholars. Whyte, a sociologist famous for his book “The Organization Man,” then became an urbanist closely examining public spaces.

By explicitly placing AI in a continuum with these visual urban studies, the authors are making a point: Powerful as it might be, we can still think of AI primarily as a tool serving human purposes, as we seek to design and refine urban form.

“Kevin Lynch at MIT was only using paper and pen,” Duarte says. “We can now scale up what he was doing, with visual AI, while also looking at many different dimension of cities.” 

There are extensive possibilities for applying visual AI to urban planning, ranging from emissions to traffic flow, safety, better imagery of street-level activity and sidewalks, and much more. The book also examines, for instance, urban greenery. While satellite imagery can show us how much tree cover and green spaces cities have, near-ubiquitous images from phones and other sources can also reveal to what extent people glimpse greenery in everyday life, a factor in reported wellness.

“The real promise of visual AI is not simply that computers can look at millions of images,” Zhang says. “It is that we can connect what is visible in those images — streets, buildings, greenery, traffic, public space — with larger questions about how cities function and how people experience them.”

Better image recognition by AI even extends to urban interiors. By using images from 400,000 AirBnB listings across the world, one recent Senseable City study shows that, contrary to some claims, interior design styles are not becoming globally more homogeneous, but reflect significant geographic differences. 

“No matter what it is, we can learn from what we can see and then use it as urban designers, planners, policymakers, and citizens,” Mazzarello says. “It can be our eyes, or cameras with computers, but in the end it’s the same methodology, and now we are trying to optimize the ways we can use these tools.”

Promise and pitfalls

If the promise of visual AI for urban studies is vast, the pitfalls are concerning. In “How AI Sees the City,” the authors outline multiple potential problems with the technology, including the intrusiveness of widespread visual surveillance and the potential for bias being reinforced through AI systems. 

The installation of ubiquitous cameras can quickly raise concerns about surveillance. London, an early adopter of CCTV, has about 210 cameras per square mile. But eight of the world’s 10 most camera-heavy cities are in China; Shanghai has over 5,000 cameras per square mile. Such surveillance practices have raised controversy in other parts of the world, with debate over the uses of traffic cameras bubbling up in the U.S. this year as well. 

In evaluating the potential safety gains from intensive video recording, the authors write, “the benefits must be weighed against the significant erosion of personal freedom and the potential for abuse inherent in a system of constant monitoring.”

Meanwhile, AI systems can reinforce social biases as well, leading to the production of data that reinforce prior perceptions as much as underlying realities — about people, neighborhoods, and whole cities. If AI models are trained on majority population groups, they may not evaluate minority groups the same way. 

“We need to teach AI to see, and depending on how you teach it, it will see what what is embedded in the culture,” Duarte says. “AI is not neutral.”

Still, as Mazzarello adds, “our eyes are not neutral, either. Every tool has to be guided in the right way, and trained in the best way.” 

Other scholars have praised “How AI Sees the City.” Michael Batty of University College London has called it a “fascinating book” that “shows how we are beginning to interpret the world of urban design, suggesting ways in which we might improve design using urban analytics, AI and large language models.”

Ultimately, though the authors think there is great value in deploying visual AI to learn more about our cities, how they function, and how they might be improved. With caution and independent thinking, progress is possible. Or, as they conclude in the book, “We should explore this wisely, critically, and creatively.” 

David Siegel SM ’86, PhD ’91, a computer scientist, entrepreneur, and philanthropist, will serve as the next MIT Innovation Fellow during the 2026-27 academic year. Working with the MIT Schwarzman College of Computing, Siegel will explore how artificial intelligence can accelerate scientific discovery at the Institute and beyond.

“From the MIT Schwarzman College of Computing to the MIT Siegel Family Quest for Intelligence, David has been a superb thought partner for me and other Institute leaders on a range of very significant initiatives, so we're delighted to have him join us now as an MIT Innovation Fellow,” says MIT President Sally Kornbluth. “Our community has long benefited from David's exceptional technical insight, entrepreneurial experience, instinct for connecting people, and infectious love for MIT. We look forward to working with him now as he helps us identify new opportunities at the intersection of AI and scientific discovery.” 

“MIT has played a foundational role in shaping how I view technology’s potential to address complex challenges,” says Siegel. “I’m thrilled to return to campus as an Innovation Fellow to collaborate with brilliant students, researchers, and faculty at a pivotal juncture in how technology shapes our world.”

A long-standing connection to MIT

Siegel’s relationship with MIT began during his graduate studies, where he earned a master’s degree in 1986 and PhD in 1991, after earning his bachelor’s degree in electrical engineering and computer science from Princeton University in 1983. Immersed in the field during a foundational era for computer science, he worked at the MIT Artificial Intelligence Lab (now the Computer Science and Artificial Intelligence Laboratory) in Professor Tomás Lozano-Pérez’s research group on human-machine interaction, contributing to the development of a pioneering humanlike robotic hand. 

Siegel has remained deeply involved with the Institute in the years since. He is a life member of the MIT Corporation, previously served on its Executive Committee, and co-chairs the External Advisory Committee for the MIT Schwarzman College of Computing. Additionally, Siegel was an early champion of the MIT Quest for Intelligence, an Institute-wide initiative studying intelligence in brains and machines, recently renamed the MIT Siegel Family Quest for Intelligence.

Entrepreneurship, philanthropy, and AI leadership

After completing his PhD, Siegel founded several early internet ventures before co-founding Two Sigma in 2001. A leading global investment firm, Two Sigma approaches investing through a data science and engineering lens, echoing Siegel’s experience in the MIT AI Lab. With the scientific method embedded in its culture, the firm uses artificial intelligence, machine learning, and advanced quantitative modeling. Siegel retired from day-to-day management of Two Sigma in 2024; however, he remains co-chair. 

Currently, Siegel’s work spans several fields, with a strong emphasis on science, technology, and philanthropy. Through the Siegel Family Endowment, a philanthropic foundation that he established in 2011, Siegel supports leaders, researchers, and organizations that are examining how technological change affects society and how to guide that shift for the public good. The endowment backs organizations such as the Scratch Foundation, Center on Rural Innovation, Khan Academy, Pursuit, and The Aspen Institute.

Recognizing the critical resource gap between academic research labs and frontier AI, Siegel founded the nonprofit Open Athena in 2024. Open Athena equips academic research labs with elite AI talent, data engineering expertise, and computational resources to enable groundbreaking discoveries at scale. The organization is also developing Marin, a 535-billion-parameter foundation model built entirely in public. By sharing every dataset and experiment in real-time, Marin ensures that the science of frontier AI remains a shared public asset for researchers and innovators worldwide. Open Athena works with leading global institutions including MIT and is funded by philanthropic partners including Bloomberg Philanthropies, Google, The Huang Foundation, and Schmidt Sciences. 

Siegel actively serves on several governance and advisory boards. He is vice-chair of the Scratch Foundation, which he co-founded in 2013 with MIT Professor Mitch Resnick, a member of the Cornell Tech Council, and a board member of organizations such as Re:Build Manufacturing, Khan Academy, and NYC FIRST Robotics. In 2025, Siegel was appointed to the U.S. Department of Energy’s Office of Science Advisory Committee, providing counsel on complex scientific and technical issues impacting federal scientific research programs.

Outside of philanthropy, Siegel remains actively engaged in the global AI ecosystem as an investor, hands-on advisor, and thought leader. Through his family office Shinrai Management, he focuses on supporting entrepreneurs and investing in high-growth startups, including several founded by MIT students and alumni.

Siegel’s debut book, “When Machines Act: The Promise and Peril of Navigating Our Agentic AI Future,” co-authored with Yale University’s Jeffrey Sonnenfeld and Stephen Henriques, will be published by MIT Press next March, coinciding with his residency as an MIT Innovation Fellow. Drawing on interviews with top tech leaders and off-the-record discussions with over 300 CEOs, the book provides a practical roadmap for autonomous AI, outlining where to deploy it, how to govern it, and which rules truly matter.

“David’s ties to MIT date back to his doctoral research in the AI Lab and have deepened through his many contributions to the Institute, including his significant involvement with the Quest for Intelligence and the MIT Schwarzman College of Computing,” says MIT Provost Anantha Chandrakasan. “His long-standing commitment to MIT, together with his vision for the future of AI and science, makes him an especially fitting Innovation Fellow. We look forward to the contributions he will make and the connections his work will foster across campus."

“For the MIT Schwarzman College of Computing, David’s fellowship is a chance to build on an already strong connection and explore how AI can expand the frontiers of science. Having known David since we were both graduate students at MIT, I’m deeply familiar with his ability to advance AI and its application in various fields,” says Dan Huttenlocher, dean of the MIT Schwarzman College of Computing and the Panasonic Professor of Electrical Engineering and Computer Science. “He understands both the college’s aspirations and the challenges ahead, and his perspective will help us identify concrete paths for research, education, and broader engagement. I look forward to working with him over the coming year.”

A year in residence

MIT Innovation Fellows typically spend a year or more in residence at the Institute. They draw on their experience, expertise, and professional networks to engage with faculty and students, participate in public events, and provide strategic counsel to MIT leaders.

The program has brought luminaries from industry and government to MIT. Most recently, Brian Deese, former White House National Economic Council director, served as an Innovation Fellow. Other fellows have included Virginia M. “Ginny” Rometty, former chair, president, and CEO of IBM; Eric Schmidt, former executive chair of Google’s parent company, Alphabet; the late Ash Carter, former U.S. secretary of defense; and former Massachusetts Governor Deval Patrick.

As an Innovation Fellow, Siegel will help guide the Institute’s focus on leveraging artificial intelligence to support scientific discovery, working closely with the MIT Schwarzman College of Computing and departments across the college to help extend their impact beyond MIT. 

“Using AI to accelerate scientific discovery is the ultimate engineering challenge. There is simply no better launchpad in the world for that work than MIT,” says Siegel.

Repeatability in engineering product design ensures that a manufacturing process performs the same way every time and allows for a working prototype to be transformed into a reliable, safe, consistent, and cost-effective mass-market product. For students in class 2.70 (Fundamentals of Precision Product Design), precision and repeatability are the name of the game. 

“[As an engineer], you have an extra responsibility to overlook nothing,” says course instructor Alex Slocum, the Walter M. and A. Hazel May Professor of Mechanical Engineering. “If you miss something, someone could be hurt or die.”

Slocum’s message is serious, but his approach to teaching the material is famously fun — in fact, he prefers the spelling “FUNdaMENTALs” for the first word of the class name. His mother, Mariana Polonsky Slocum, was a mathematics student at MIT. “She taught me, physics doesn't care about your feelings,” he says. “I want [students] to understand that we are governed by the laws of physics, and that is a catalyst for creativity, not a hindrance. It is a hindrance if you forget that.” 

Through the course, students learn deterministic design, selection, and assembly of machine elements to create and manufacture robust precision machines, instruments, and systems. They also apply Slocum’s “Functional Requirements, Ergonomics and Environment, Design Parameters, Analysis, References, Risks, Countermeasures” (FRED PARRC, pronounced like “Fred Park”) model, and engage in peer review and evaluation.

“You get a lot of time working on problems that just pop up in engineering. To me, it felt very [representative] of the grad work that I was doing,” says Mariia Smyk, a graduate student in mechanical engineering. 

Some students may describe the course as “creative chaos,” but tend to agree that their learning experience is one that drives home the fundamentals. 

“It definitely made me more confident knowing that I can look at what I'm designing and be very deliberate in taking steps toward mitigating the risks that anyone would face when they use a product,” says graduate student Adian Salazar. “I feel like I've been able to apply all those really fundamental concepts that I learned in the more theory-heavy classes to real-world machines.”

“The minute you think, ‘I’ve done everything I can possibly do for Lincoln Lab,’ next thing you know, you’re belly crawling through a mock rubble pile,” says Niki Fandel. 

Through hundreds of photo shoots each year, the MT Lincoln Laboratory photographer captures technical innovations — from functional fibers to advanced radar systems ­— and the people who develop them for national security. Her work brings her to every corner of the laboratory’s main campus in Lexington, Massachusetts, and to field sites across the country.  

Video by Tim Briggs/MIT Lincoln Laboratory | 5 minutes, 56 seconds