Preconditioned Physics-Informed Neural Operator Training

DOI10.48550/arXiv.2609.36216MaRDI QIDQ7361993arXivFDO

Authors Siddhartha Mishra, Marius Zeinhofer, Shizheng Wen

Publication date 28 September 2026

Published in arXiv

Tldr We train neural operators from the governing partial differential equation using a preconditioned residual loss function, achieving mesh-independent conditioning for elliptic problems and greatly improved conditioning for saddle point problems; this allows label-free training to match supervised accuracy on three model problems.

Full work available at URL https://doi.org/10.48550/arXiv.2609.36216

arXiv classification

cs.LGmath.NA

Cites work

• AmgX: a library for GPU accelerated algebraic multigrid and preconditioned iterative methods
• FEM-based Neural Networks for Solving Incompressible Fluid Flows and Related Inverse Problems
• Fourier Neural Operator for Parametric Partial Differential Equations
• Learning the solution operator of parametric partial differential equations with physics-informed DeepOnets
• Optimization in SciML -- A Function Space Perspective
• Parallel Multilevel Preconditioners
• Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
• Physics-Informed Neural Operator for Learning Partial Differential Equations
• Preconditioning and Numerical Stability in Neural Network Training for Parametric PDEs
• The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems
• Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks
• When and why PINNs fail to train: A neural tangent kernel perspective

This page was built for publication: Preconditioned Physics-Informed Neural Operator Training

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7361993)