Explore a cutting-edge approach to solving Partial Differential Equations (PDEs) in this 25-minute DS4DM Coffee Talk presented by Dan Assouline from Mila, Université de Montréal. Delve into the innovative Mesh Agnostic Neural PDE Solver (MAgNet), which addresses the computational limitations of classical numerical methods for high-resolution PDE solutions. Learn how this novel architecture combines Implicit Neural Representations (INR) with Graph Neural Networks (GNN) to predict spatially continuous PDE solutions. Discover MAgNet's capabilities in zero-shot generalization to new non-uniform meshes, long-term physically consistent predictions, and its superior performance across various PDE simulation datasets. Gain insights into how this approach can potentially revolutionize fields like climate prediction by enabling accurate simulations at finer resolutions than currently possible with modern supercomputers.
Overview
Syllabus
MAgNet: Mesh Agnostic Neural PDE Solver, Dan Assouline
Taught by
GERAD Research Center