Reproduce “Towards General Neural Surrogate Solvers with Specialized Neural Accelerators” 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

Chenkai Mao, Robert Lupoiu, Tianxiang Dai, Mingkun Chen, Jonathan A. Fan

Researchers at Stanford University developed Specialized Neural Accelerator-Powered Domain Decomposition Methods (SNAP-DDM), integrating specialized Fourier Neural Operators with classical Domain Decomposition Methods to solve large-scale parametric PDEs. This framework enables accurate and generalizable solutions for problems with arbitrary domain sizes and boundary conditions, achieving near-unity accuracy for subdomain problems and global errors around 5% for electromagnetics.

Key insights

  • Decomposing large, arbitrary PDE problems into smaller, fixed-size subdomains allows neural operators to overcome limitations in scalability and handling diverse boundary conditions.
  • Training specialized neural operators for different types of subdomain physics (e.g., PML, sources, materials) enhances their individual accuracy and robustly enables the overall domain decomposition process.
  • The Self-Modulating Fourier Neural Operator (SM-FNO) architecture improves model expressivity and training stability through dynamic modulation and explicit residual connections, achieving near-unity accuracy for subdomain solutions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.