Hannes Stärk, a fourth-year PhD student at CSAIL who built VoltzGen, says the model works because it actually learns, draws inferences from the data it was trained on, and generates novel ideas inspired by that data. Machine learning requires models to generalize from the data they are trained on, Stark says. He spent seven months creating VoltzGen, often working 12 hours a day. “Otherwise, the solution is already in the training data,” he says. Mr. Stark has also built a network of more than 30 scientists at and outside of MIT who are researching drug development, metabolomics, structural biology, and the design and application of molecular binders for the treatment of cancer, autoimmune diseases, and genetic diseases. “It’s great to have one model that can do all this,” he says. Training across all these areas also improves the generalization of the model.
Beyond drug discovery
As labs working on drug development continue to reap the benefits of AI, other researchers across the institute are busy applying existing AI tools or, in many cases, developing their own models for use in a myriad of fields and applications. An interdisciplinary group involving the Department of Electrical Engineering and Computer Science (EECS), CSAIL, and Massachusetts General Hospital has launched MultiverSeg, a tool that rapidly annotates regions of interest in medical images, potentially helping scientists develop new treatments and map disease progression. MIT researchers are also designing and operating an AI-driven automated laboratory to accelerate and refine the process of discovering sustainable materials and new components for solar panels. Ahmed’s MechE group also develops AI models that help automakers design high-performance vehicles and determine whether large transport ships are seaworthy. Ahmed also teaches a course titled AI and Machine Learning for Engineering Design. Offered for the first time in 2021, it attracts mechanical, civil, and environmental engineers, as well as students from aerospace engineering, Sloan, and more.
MIT Technology Review
“The goal is to harness diverse types of raw data and turn it into something that helps us understand what’s putting species at risk.”
sarah beery
Meanwhile, Priya Donti, assistant professor at EECS and principal investigator at the Institute for Information and Decision Systems (LIDS), has developed an AI-powered optimization approach to help schedule generation resources on the power grid. The machine learning tools her group builds will help utility operators address many unavoidable power grid problems. “The big challenge is that the power grid has to maintain a precise balance between the amount of power that is produced and put into the grid and the amount of power that is taken out on the other side,” she explains. “When you have large fluctuations from solar, wind, and other power sources whose output changes with the weather, the power grid has to be more tightly tuned to maintain that balance.” Information about the physics of how the power grid works is embedded in Donti’s AI model, so it functions and reacts much like a real power grid.
MIT researchers are applying AI tools to explore and analyze the natural world. Sara Beery, an assistant professor at EECS specializing in AI and decision-making, discovers and mines ecological data collected by a wide range of remote sensing technologies to develop AI methods to analyze and predict how species and ecosystems around the world are changing. These technologies have enabled Biery and his colleagues to collect data on far more endangered species than ever before at an unprecedented scale. Historically, most ecological research has focused on collecting incredibly rich data on a single species in a very small area, but “we realized that that’s not enough,” she says. For example, information gleaned from just one small part of a river’s ecosystem won’t help us understand or stop what she calls “the exponential increase in species extinction rates that we’re currently facing.” Already, Bialy says, “We are using multimodal AI to enable experts to quickly search vast repositories of image data and discover data points that were previously very difficult to find.” But the goal, she says, is to make different types of raw data readily available, from satellite and bioacoustic sensor data to camera images and DNA, “so that we can actually turn it into the kind of scientific insights that will help us understand what’s putting species at risk.”
Men’s et Manus With AI
While some MIT researchers are successfully leveraging AI to invent technologies ranging from new cancer treatments to safer, more powerful cars, others are also using machine learning and other AI tools to determine whether these technologies work as promised or can be successfully and economically produced at scale. Connor Coley, SM ’16, PhD ’19 is an associate professor of chemical engineering and EECS who designs new molecules and recipes for creating new molecules, primarily small organic molecules, for potential use in pharmaceutical, agricultural, and other chemical companies. Former Corey MIT Technology Review Innovators Under 35 winners have developed “genetic” algorithms that use biologically inspired processes such as selection and mutation. The tool effectively encodes potential polymer blends extracted from a large database of polymers into a digital chromosome, and algorithms are refined to generate the most promising material combinations.
Corey, who works at the intersection of chemistry and computer science, believes AI could one day help his lab discover polymer blends that lead to improved battery electrolytes or customized nanoparticles for safer drug delivery. He and his lab are also working on developing machine learning tools to streamline the discovery and production process. “If you want AI to be the brains of the science you do, you also need hands,” says Corey, who was one of the first MIT faculty members hired into the MIT Schwarzman College of Computing. He and his group combined a robotic liquid handling platform with an optimization algorithm. In a project designed to find the best polymer blend, the autonomous system not only selects polymer solutions to test, but also performs physical tests. The system was able to generate and test 700 new polymer blends per day and identified one that performed 18% better than any of its components.
Systems with similar levels of autonomy could also have a major impact on early-stage drug discovery. He believes one of the benefits will be to speed up the time it takes to get drugs from the lab to clinical trials. But the real question, he says, is: “What can we do in the future that we couldn’t do before with a reasonable amount of resources?”
Dr. Alexander Shimen (PhD ’25) also uses AI both to search for new materials and to control robots that test the physical properties of those materials. For his doctoral thesis, Shimen built a fully autonomous, AI-powered robotic laboratory from scratch to discover and test sustainable, high-performance materials for solar panels. The system incorporates computer vision, machine learning, and optimization algorithms and operates 24 hours a day.
