Reproduce “Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes” on OpenResearch

About this paper

Tobias Würth, Niklas Freymuth, Clemens Zimmerling, Gerhard Neumann, Luise Kärger

PI-MGNs, developed by researchers at Karlsruhe Institute of Technology, introduce a framework that combines MeshGraphNets with physics-informed training to solve non-stationary and nonlinear partial differential equations on arbitrary meshes without requiring simulation data. This approach achieves computational speedups of several orders of magnitude over traditional FEM while demonstrating strong generalization to unseen and larger geometries.

Key insights

• Physics-informed training within a GNN framework enables robust generalization to arbitrary mesh geometries, diverse material distributions, and unseen process conditions without relying on pre-computed simulation data.
• An FEM-based physics-informed loss function can effectively train neural solvers for complex nonlinear and non-stationary PDEs, bypassing the need for iterative solvers during training and mitigating error accumulation in time-dependent simulations.
• Architectural components like temporal bundling for multi-step prediction, relative positional encoding, and noise injection are critical for enhancing the accuracy, robustness, and generalization performance of PI-MGNs, especially for long simulation rollouts.

Summary by alphaXiv.