Explore score matching via differentiable physics in this comprehensive conference talk. Delve into the integration of stochastic differential equations and physics operators for modeling non-deterministic physical processes. Learn about replacing SDE drift with differentiable simulators or neural network approximations of physics. Discover how to optimize probability flow ODEs for fitting simulation trajectories and solving reverse-time SDEs for inference. Examine applications to challenging inverse problems, including heat diffusion, buoyancy-driven flow with obstacles, and Navier-Stokes equations. Gain insights into reconstruction MSE versus spectral error and the effects of multiple training steps. Engage with the speaker's expertise during an informative Q&A session.
Overview
Syllabus
- Intro
- Score Matching and Reverse-Diffusion
- Learned Corrections for Physical Simulations
- Combining Physics and Score Matching
- Heat Diffusion
- Reconstruction MSE vs Spectral Error
- Effects of Multiple Steps During Training
- Buoyancy-driven Flow with Obstacles
- Navier Stokes Equations
- Summary
- Q+A
Taught by
Valence Labs
