PhD Defense: Machine Learning with Differentiable Physics Priors

https://umd.zoom.us/j/6621676794

Differentiable physics is a powerful and novel approach to learning and control problems that involve physical objects and environments. We have developed differentiable scalable, powerful, and efficient differentiable simulators. These include the state-of-the-art differentiable physics for rigid body, cloth, fluids, articulated body, deformable solid, hybrid traffic system, NeRF-based representation, and even quantum dynamics, which as a whole built a closed-loop differentiable pipeline to learn the physics world. A diagram in my research statement visualizes the structure of such a differentiable system. Our physics priors can serve as a strong prior of our world and greatly improve the data efficiency when training AI algorithms. It can be integrated with applications like embodied AI (articulated body), AI for fashion and design (cloth), animation (soft body), ML for science (fluids, soft materials, quantum), autonomous driving (traffic), and quantum computing.