AI Agents for Science
Complex Scientific Simulations
We build AI workflows that help researchers translate scientific questions into simulations, diagnose failures, and assess whether the results are physically and numerically sound. In collaboration with Lawrence Livermore National Laboratory and Stanford colleagues, we work with GEOS, a simulation platform for coupled subsurface flow and rock mechanics, with applications including geothermal energy and carbon storage.
Our approach combines language models for planning with reproducible computational tools and independent checks of conservation laws, numerical convergence, and other physical constraints. A central question is when agents can proceed reliably and when expert judgment is needed.
Experimental Data Analysis
We develop scientific assistants that connect natural-language questions to large datasets, analysis tools, and published knowledge. Through ORACLE, our collaboration with SLAC, we are developing workflows for particle physics and the Deep Underground Neutrino Experiment (DUNE).
These workflows will help researchers investigate unusual observations, test competing explanations, and distinguish detector or reconstruction problems from meaningful physics. Related work on SpectraQuery combines battery spectroscopy measurements with scientific literature. Across these projects, we seek answers that researchers can inspect and reproduce, with conclusions linked to the data, assumptions, and computations supporting them.
