Reproduce “Towards a Foundation Model for Partial Differential Equations Across Physics Domains” 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

Eduardo Soares, Emilio Vital Brazil, Victor Shirasuna, Breno W. S. R. de Carvalho, Cristiano Malossi

IBM Research introduces PDE-FM, a foundation model for physics-informed machine learning designed to generalize across diverse Partial Differential Equation (PDE) systems. PDE-FM achieved state-of-the-art accuracy in six out of twelve heterogeneous datasets from The Well benchmark, demonstrating a 46% mean VRMSE reduction compared to prior operator-learning baselines while maintaining efficiency with a Mamba state-space backbone.

Key insights

  • Pretraining on a diverse corpus of physical simulations induces emergent cross-physics generalization, allowing transferable representations of physical dynamics across varied domains.
  • Combining a Mamba state-space backbone with an FNO-based decoder offers superior expressivity and computational efficiency for modeling continuous physical fields.
  • Dual spatial and spectral tokenization, alongside physics-aware conditioning, is crucial for effectively capturing both local patterns and global coarse structures in heterogeneous PDE systems.

Summary by alphaXiv.

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