Reproduce “PGD-NO: A Neural Operator with Precomputed Geometry Decomposition for 3D Million-scale Physics Simulations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weiheng Zhong, Jing Bi, Victor Oancea, Hadi Meidani
Researchers developed PGD-NO, a neural operator that uses precomputed geometry decomposition to enable 3D physics simulations on meshes with over 10 million nodes. This approach addresses the memory and scalability limitations of existing neural solvers by decoupling geometry encoding, achieving competitive predictive accuracy and processing meshes up to 60 million nodes.
Key insights
- PDE solutions can be effectively learned using non-learnable, precomputed geometric primitives, which significantly reduces the memory footprint of neural operators.
- Decoupling geometry encoding from solution querying enables linear memory scalability, allowing processing of meshes with tens of millions of nodes on standard hardware by shifting the memory bottleneck to surface node operations.
- Iterative refinement of geometry tokens through multiple Geometric Token Layers is crucial for enhancing predictive accuracy, especially for complex 3D physics simulations, as evidenced by consistent performance improvements with increasing layer depth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.