Loading page…
The team wants models to uncover rules governing groups of interacting components, with mathematical guarantees that predictions hold up under noisy experimental data.
Listen to this story
The $1 million NSF-backed effort will test whether AI can infer governing rules from interactions among groups of components, rather than only pairwise relationships. UCSB researchers plan to build mathematical foundations and reliability guarantees, then evaluate the framework on synthetic data and datasets from several scientific fields. The goal is to move beyond forecasting toward identifying equations and physical principles. Whether the methods transfer across disciplines and remain trustworthy on noisy experimental data is still to be demonstrated.
The team is targeting overlapping higher-order interactions, including relationships that span multiple scales or change over time.
Researchers plan to explore Kolmogorov-Arnold networks and topological neural networks as possible routes to identifying governing equations.
Yao Qin’s work focuses on mathematical guarantees for model reliability, with potential applications in molecular biology and drug discovery.
Discovering a scientific law takes more than predicting what happens next. Four UC Santa Barbara researchers have received a $1 million National Science Foundation award to develop AI that can uncover rules governing groups of interacting components. The university’s announcement describes a research program that starts with mathematical foundations, then tests whether the methods work across different scientific settings.
The project, “Foundations for Discovering Higher-Order Scientific Laws with Deep Learning,” brings together Nina Miolane, Haewon Jeong, Sanjukta Krishnagopal and Yao Qin. It is the first federal grant for UCSB’s Real AI Initiative, which launched earlier in 2026. Miolane, Jeong and Qin co-direct the initiative.
Many current AI approaches in science represent relationships two components at a time, Miolane said: two atoms, two brain regions or two stars. But a group can produce behavior that its individual pairs cannot. The researchers want models to capture those higher-order interactions—relationships involving several components together—including overlapping groups, different scales and changes over time.
Miolane’s contribution draws on topological deep learning, which uses mathematics to represent group interactions in data. Her work has applied that approach to the brain. For this project, the goal is to give models a way to calculate how multiple parts act together, then use that representation to investigate the rules governing the whole system.
Jeong’s group studies machine learning for complex scientific systems. She wants the models to do more than deliver accurate forecasts: they should also reveal what they have learned. The team hopes to explore architectures including Kolmogorov-Arnold networks and topological neural networks as routes toward identifying governing equations and physical principles.
How can we build AI systems that understand not only human language, but also the physical world?
Haewon Jeong, UCSB assistant professor of electrical and computer engineering
Qin is leading the project’s certified robustness work: developing mathematical guarantees about model reliability. She said some promising architectures for scientific research lack the theoretical foundations needed to tell scientists when predictions can be trusted amid the noise in real experimental data.
That concern is especially important in molecular biology and drug discovery, Qin said, where small prediction errors can have serious downstream consequences. Her part of the project aims to build a foundation for provable robustness, rather than relying on enthusiasm for models that scientists are already adopting.
Krishnagopal brings a focus on interconnected social and biological systems, informed by her physics background. She wants the framework to explain how an entire system behaves over time, not just describe a single snapshot. The team’s broader aim is a mathematically consistent approach that does not depend on one particular scale or dataset.
The planned work follows a sequence, with tests spanning very different scientific settings. Neuroscience data will include datasets drawn from Miolane’s research; molecular biology and cosmology provide other settings for checking whether the approach extends beyond a single discipline or kind of system.
For Qin, broader usefulness would be one measure of success. She described the desired shift in AI for science as moving toward models that are not only accurate, but demonstrably trustworthy—a goal the funded work will now pursue.
Loading discussion...
Join the conversation
Explain when testing alone would—or would not—be enough.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.