Audrey Woods, MIT CSAIL Alliances | September 21, 2026
Imagine you’re using generative AI to simulate water flowing over a dam. It might look realistic enough, maybe even indistinguishable from reality. But for scientific machine learning and safety-critical applications, there’s a big difference between looking right and being certifiably, technically correct.
“Centuries of work have gone into understanding science through equations,” says MIT CSAIL Graduate Student Utkarsh. “But as AI and machine learning have gotten really big, a lot of that information is assumed to be captured implicitly from data,” which is not good enough for rigorous scientific AI. “If I'm building a rocket, or a space mission, I want exact correctness.”
As a fifth-year PhD student in the Julia Lab studying under CSAIL Professor Alan Edelman and Research Affiliate Christopher Rackauckas, Utkarsh is working to solve one of the biggest hurdles in scientific ML: teaching generative models to strictly respect hard mathematical and physical constraints. The problem is not solved yet, but Utkarsh says, “the scientific implications are enormous. AI for science is becoming as big as it possibly can, and we're doing our best to help get it to the point where it starts to transform everyone's lives.”
BUILDING PHYSICAL REALISM INTO GENERATIVE
AI The core of Utkarsh’s research centers around the question of creating scientifically robust generative AI. “We’ve all seen these nice images, which may be great for memes or personal use. But can you apply these models to build physical simulations? Can you deploy them in industry, in real systems?” To address this, Utkarsh’s approach is to preserve the broad capabilities of foundation models trained on massive amounts of data, while developing post-training and inference-time methods that make them more reliable for physical simulation. “The goal is not simply to make these models obey known mathematical and physical constraints, but to use that structure to guide them toward better generations that also obey physical laws as much as possible” For example, Utkarsh and his colleagues published the idea of physics-constrained flow matching at NeurIPS 2025, which introduced a method for a zero-shot inference framework that enforces constraints—like the laws of physics—in pretrained generative models. Utkarsh says this project has had “real uptake in academia. People are using it, and I've heard good feedback.” Recently, the team expanded the idea into molecule generation, which could help scientists working on drug discovery, protein generation, and using AI to develop novel medical treatments. “AI for science is becoming very big, but there's always the risk that these models miss physics completely. The outputs look real, but when you check the metrics, they violate them.” Utkarsh hopes his method might mitigate that risk.
Beyond MIT, Utkarsh has also been working with NVIDIA to apply his research to weather forecasting, a notoriously complex challenge. “The problem NVIDIA is trying to solve is this: we've collected an enormous amount of weather data since the 1980s, and we get more every day.” These data are a potential goldmine for AI training and could be critical for predicting extreme future events. But how to be sure that the generated weather simulations are physically and scientifically accurate? While the project is still ongoing and not yet public, Utkarsh says he is using his background in physical simulations and computational physics to allow models to do super-resolution, predicting weather at a 5-kilometer resolution as fast and accurately as possible.
Recently, Utkarsh has been asking, “What happens when you have a richer structure? What if the code is for a simulation? Then you can't just test yes or no. You need to test whether the answer is correct in a continuous sense.” For that, he’s created a way to use reinforcement learning with verifiable physics, which offers both hard program-validity checks as well as continuous physics rewards for accuracy and consistency. Unlike his inference-time constraint work, this approach uses physics directly as feedback during post-training of LLMs. The goal of this project is to train a broadly capable model that can generate reliable simulation code across different physical systems, rather than having to train a separate model for every new problem or condition. By applying physics-based feedback on the code, Utkarsh hopes scientists will be able to get to reliable simulation code faster.
THE JULIA LAB: PERFORMANCE & REPRODUCIBILITY
Utkarsh was drawn to the Julia lab because, “We don't just do research, write a paper, and call it a day. We actually want to implement things and be as transparent as possible.” This aligns with his previous experience. Earlier on in his research, Utkarsh worked on solvers that scaled well on GPUs. “The solvers I developed were 20 to 100 times faster than what was available, and people could use them off the shelf.” His solvers are now used at national labs for astrophysical simulations and at companies like Pfizer to accelerate drug discovery.
This experience made him passionate about developing methods that work across domains and “transfer as much as possible,” an ethos he shares with his Julia Lab colleagues. “They genuinely care about what scientists want, and they come from both worlds. They can speak the mathematical language and the computational language.” He aims to release as much of his work as possible as open-source tools and build them using the Julia programming language, which was designed to combine the accessibility of a high-level programming language with the performance needed for scientific computing. “That's one of Julia's success stories in scientific computing: it's popular because of how easily people can write code in it and still get performance quickly.”
LOOKING FORWARD: ALPHAFOLD FOR PHYSICS & SCIENTIFIC IMPACT
The release of AlphaFold by Google DeepMind was an enormous moment in biotech because “it solved the very hard problem of protein-structure prediction. It took decades of engineering, scientific research, biology, and machine learning to reach that level.” Utkarsh says that same moment hasn’t yet happened in physics-constrained AI and ML. “High-fidelity simulations of complex physical systems can be extraordinarily computationally expensive, which is why all these computational techniques were developed. The question I've been trying to answer is how to make them as fast as possible.” Incorporating physics equations and prior knowledge into large AI models is a fundamentally different problem from natural language processing, because natural language does not come with known governing equations analogous to those in physics. ChatGPT and other LLMs were trained with books and written data. But physics will require integrating “the structure we already know” into AI generation. “That isn't solved yet, although we've been making progress.”
At this critical moment in AI transformation, Utkarsh is “proud that what we do is somewhat different from other people. We try to understand things from first principles. We don't want to take it on faith that the data will learn everything. We want to reason about it and check ourselves, to ask whether we're doing the right thing.” While recent advances in AI-assisted mathematics and mathematical proofs have made him increasingly optimistic about the scientific capabilities of foundation models, his goal remains harnessing that same growing capability to build faithful physical simulations from large AI models, while incorporating the mathematical and physical structure scientists already know. The importance of that pursuit is not lost on him, with questions about how exactly AI is going to impact lives and shape the future of science still up for debate. He says, “I feel very lucky to be here at this moment, which is a career-defining moment for a lot of people.”
Learn more about Utkarsh on his website, LinkedIn, or CSAIL page.
