My research vision is to develop machine learning (ML) methods for accelerating scientific simulation and discovery, while opening new frontiers in machine learning research (AI + Science). This lies in the interdisciplinary field of machine learning, scientific computing, and physical sciences.
Towards this goal, my past research has pioneered and made important advances to learning structured and compressed representations for accelerating large-scale and multi-scale simulations in physical sciences, including fluid, plasma, and more generic PDEs and N-body systems. My research has enabled ML-based surrogate models to scale to dynamical systems with two orders of magnitude higher dimensions
and 15x faster than prior ML models. The ML models I developed are being deployed for fluid simulation in industry and will also be used for modeling laser-plasma systems in Stanford National Accelerator Laboratory (SLAC). Besides ML for simulation, I have introduced ML methods for discovering symbolic theories (published in a top physics journal) and relational structures from observations, and have theoretically revealed the origin of phase transition phenomena for the compression vs. prediction tradeoff
in representation learning.