Physics-Informed Coarsening for Multigrid Graph Neural Surrogates

Anonymous Authors

Paper : Paper will be released upon acceptance.

Code : Code will be released upon acceptance.

Datasets: Datasets will be released upon acceptance.

Overview

We introduce a multigrid graph neural network for solid mechanics simulations on unstructured meshes. Our approach leverages physics-informed coarsening based on residual signals to prioritize physically important regions.

This enables improved long-horizon rollout stability and better generalization compared to single-scale graph neural networks.

plate_project.mp4

Plate Dataset

rode_project.mp4

Rode Dataset

spindle_project.mp4

Spindle Dataset

Page updated

Google Sites

Report abuse