Bridging Numerical Methods and Deep Learning with Physics-Constrained Differentiable Solvers

Explore cutting-edge machine learning strategies for neural PDE solvers in this hour-long talk by UC Berkeley Assistant Professor Aditi Krishnapriyan. Delve into self-supervised learning techniques for solving fluid dynamics and transport PDE problems using spectral methods, and discover "simulation-in-the-loop" approaches that incorporate PDE-constrained optimization as a neural network layer. Learn how machine learning methods can be seamlessly integrated with numerical methods in fully differentiable settings, addressing challenges in spatiotemporal modeling and limited data scenarios. Gain insights into physics-inspired machine learning, geometric deep learning, and inverse problems, with applications in molecular dynamics and fluid mechanics.

DDPS | Bridging numerical methods and deep learning with physics-constrained differentiable solvers