The realm of artificial intelligence research is evolving at an unprecedented pace, with groundbreaking innovations emerging from laboratories at the Massachusetts Institute of Technology. Through the prestigious Undergraduate Research Opportunities Program (UROP), talented undergraduates actively participate in cutting-edge artificial intelligence research initiatives. Over the past two years, the MIT Quest for Intelligence has successfully placed 329 students in revolutionary research projects designed to expand the boundaries of computing and AI technology, while simultaneously leveraging these advanced tools to transform brain research methodologies, enhance medical diagnostic and treatment approaches, and discover extraordinary materials with remarkable properties.
Rafael Gomez-Bombarelli, an assistant professor in MIT's Department of Materials Science and Engineering, has engaged several Quest-supported undergraduates in his pioneering mission to identify novel molecules and materials through artificial intelligence applications. "These students bring fresh perspectives and tremendous enthusiasm to our research endeavors," he notes. "The Quest program has enabled us to connect with students from diverse academic backgrounds who might not have otherwise considered joining our research initiatives."
While some students participate in laboratory research for a single semester, others develop lasting connections to their research teams. Nick Bonaker exemplifies this long-term engagement, now in his third year collaborating with Tamara Broderick, an associate professor in the Department of Electrical Engineering and Computer Science, to develop innovative assistive technology solutions specifically designed for individuals with severe motor impairments.
"Nick consistently demonstrates remarkable aptitude for rapidly mastering new tools and concepts, impressing both myself and our research collaborators," Professor Broderick observes. "His thoughtful engagement with the needs of the motor-impaired community particularly stands out. He has meticulously incorporated feedback from motor-impaired users, charitable organization partners, and academic researchers throughout the development process."
This autumn, the MIT Quest commemorated two years of supporting UROP students in their artificial intelligence research endeavors. Below, we showcase four exemplary projects from the recent semester that highlight the remarkable achievements possible when undergraduate talent meets cutting-edge AI technology.
Revolutionizing Solar Energy Harvesting with Machine Learning
As solar technology continues advancing, the cost of solar energy production steadily declines. Experimental solar cells have achieved nearly 50 percent efficiency in laboratory settings, but researchers see no reason to stop there, according to Sean Mann, a sophomore specializing in computer science.
Through his UROP collaboration with Giuseppe Romano, a researcher at MIT's Institute for Soldier Nanotechnologies, Mann is developing an advanced solar cell simulator that enables deep learning algorithms to systematically identify superior solar cell designs. Historically, efficiency improvements required evaluating new materials and geometries involving hundreds of variables. "Conventional design exploration methods prove costly because simulations only measure the efficiency of one specific configuration," Mann explains. "This approach doesn't provide guidance for improvement, necessitating either expert knowledge or extensive additional experimentation to enhance performance."
Mann's project aims to create a differentiable solar cell simulator that not only calculates cell efficiency but also indicates how adjusting specific parameters will enhance performance. With this information, artificial intelligence systems can predict which modifications among countless combinations will most significantly boost cell performance. "Integrating this simulator with a neural network designed to maximize cell efficiency will ultimately yield exceptional designs," he predicts.
Currently, Mann is building an interface between AI models and traditional simulation tools. The most significant challenge thus far has been debugging the simulator, which solves complex differential equations. After several sleepless nights meticulously verifying his equations and code, he identified the issue: an array offset by one element, causing skewed results. With this obstacle overcome, Mann now focuses on finding algorithms to help the solver converge more rapidly—a crucial step toward achieving efficient optimization.
Applying Physics-Informed Neural Networks for Stress Fracture Detection
Modern jet engines contain internal sensors that alert operators when malfunctions occur. However, precisely diagnosing the specific failure often requires engine disassembly. To obtain clearer diagnostic information more efficiently, engineers are experimenting with physics-informed deep learning algorithms to interpret these sensor distress signals.
"This technology would significantly simplify identifying problematic components, eliminating the need for complete engine disassembly," notes Julia Gaubatz, a senior studying aerospace engineering. "Such advancements could save substantial time and resources in industrial applications."
Gaubatz spent the fall semester programming physical constraints into a deep learning model for her UROP project, working alongside Raul Radovitzky, a professor in MIT's Department of Aeronautics and Astronautics, graduate student Grégoire Chomette, and third-year student Parker Mayhew. Their objective involves analyzing high-frequency signals from components like jet engine shafts to identify specific stress points where cracks might develop. They aim to pinpoint potential failure points by training neural networks on numerical simulations of material breakdown processes to understand the underlying physics.
Working from her Cambridge apartment, Gaubatz constructed a simplified version of their physics-informed model to verify their assumptions. "This approach makes it easier to examine the neural network's weights to understand its predictions," she explains. "It serves as a verification method to ensure the model behaves according to theoretical expectations."
Gaubatz selected this project to apply concepts from her machine learning coursework to solid mechanics, which focuses on how materials deform and fracture under stress. The integration of deep learning into this field represents an emerging frontier, she observes. "It's fascinating to witness how new mathematical approaches might transform traditional engineering practices."
Developing AI Systems with Advanced Visual Reasoning Capabilities
An artificial intelligence model capable of superhuman chess performance might struggle with Sudoku puzzles. Humans, conversely, easily adapt existing knowledge to master new games. To impart similar adaptability to AI systems, researchers developed the ARC visual-reasoning dataset to encourage the creation of novel techniques for solving problems involving abstraction and reasoning.
"Strong performance on this test indicates more human-like intelligence in AI systems," explains first-year student Subhash Kantamneni, who joined a UROP project this fall in the laboratory of Department of Brain and Cognitive Sciences Professor Tomaso Poggio, part of the Center for Minds, Brains and Machines.
Professor Poggio's laboratory aims to conquer the ARC challenge by combining deep learning with automated program-writing techniques to train an agent capable of solving ARC's 400 tasks by generating its own programs. Much of their work utilizes DreamCoder, an MIT-developed tool that learns new concepts while solving specialized tasks. Using DreamCoder, the laboratory has successfully solved 70 ARC tasks, with Kantamneni collaborating this fall with master of engineering student Simon Alford to tackle the remaining challenges.
To address ARC's approximately 20 pattern-completion tasks, Kantamneni created a script to generate similar examples for training the deep learning model. He also developed several mini-programs, or primitives, to solve a separate category of tasks involving logical operations on pixels. With these new primitives, DreamCoder learned to combine existing and new programs to solve ARC's approximately 10 pixel-based tasks.
Despite the challenging coding and debugging process, Kantamneni felt welcomed and valued by the laboratory team. "I don't think they even realized I was a freshman," he remarks. "They genuinely listened to my contributions and respected my input."
Examining Language Computation Through Advanced AI Analysis
Language functions as more than a symbolic system—it enables concept expression, cognitive reasoning, and interpersonal communication. To understand the brain's language processing mechanisms, psychologists have developed methods for tracking how quickly people comprehend written and spoken information. Extended reading times often indicate improper word usage, providing insights into how the brain incrementally derives meaning from word sequences.
For her fall UROP project in Roger Levy's BCS laboratory, sophomore Pranali Vani conducted online sentence-processing experiments developed by a previous UROP student. Each sentence contained a word positioned to create ambiguity or implausibility. The more unusual the sentence construction, the longer human participants required to decipher its meaning. For instance, placing a verb like "tripped" at the end of a sentence—"The woman brought the sandwich from the kitchen tripped"—typically confuses native English speakers. Although grammatically correct, this structure suggests that bringing rather than tripping represents the sentence's primary action, creating reader confusion.
Across three experimental sets, Vani discovered that the most significant processing delays occurred when verbs were positioned in ways that sounded ungrammatical. Vani and her advisor, Ethan Wilcox, a Harvard University PhD student, observed similar results when conducting the experiments on a deep learning model.
"The model exhibited 'surprise' when grammatical interpretations appeared unlikely," Wilcox notes. Although the model received no explicit grammar training, these results suggest that neural networks trained on extensive text corpora effectively learn grammatical rules implicitly.
Vani reports that she enjoyed learning to program in R and shell scripts while gaining appreciation for the persistence required in original research. "Research demands considerable time investment," she reflects. "Every detail and decision throughout the experimental process requires careful consideration and thought."
Partial funding for this semester's MIT Quest UROP projects was provided by the MIT-IBM Watson AI Lab.
