Reproduce “Deep Equilibrium Based Neural Operators for Steady-State PDEs” 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

Tanya Marwah, Ashwini Pokle, J. Zico Kolter, Zachary C. Lipton, Jianfeng Lu, Andrej Risteski

Researchers from Carnegie Mellon University and Duke University introduce FNO-DEQ, a Fourier Neural Operator variant that leverages Deep Equilibrium Models to solve steady-state Partial Differential Equations by modeling their solutions as fixed points of an implicit, weight-tied neural network. This architecture achieves superior prediction accuracy and robustness to noise with fewer parameters compared to existing FNO-based methods.

Key insights

  • Weight-tying acts as a powerful inductive bias for neural operators solving steady-state PDEs, leading to more efficient and accurate models by aligning the architecture with the mathematical problem structure.
  • Deep Equilibrium Models provide a principled way to build infinite-depth networks with O(1) training memory, addressing scalability challenges in deep learning for scientific computing.
  • The integration of DEQs with FNOs can significantly enhance model robustness to noise in both input parameters and observed solutions, which is critical for practical applications with imperfect data.

Summary by alphaXiv.

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