Physics-Tailored Machine Learning
Building machine learning models that respect physical symmetries and constraints to discover new physics from experimental data.
Standard machine learning treats data as abstract vectors. Physics-tailored ML goes further: it embeds known structure — symmetries, conservation laws, known functional forms — directly into the model architecture, making learning more data-efficient and the results more interpretable.
The payoff is not just better predictions but genuine scientific discovery: when the physics-tailored model disagrees with theory, it is telling you something real.
Key result: Applied to dusty plasmas — complex systems of ions, electrons, and charged particles — a physics-tailored neural network trained on 3D particle trajectories recovered the inter-particle forces with high precision and revealed that the measured particle charge deviates from theoretical predictions. The result suggests that ML can serve as a tool for deriving physical laws in many-body systems beyond the reach of analytic methods (PNAS 2025, 2025 Cozzarelli Prize).