Preconditioned Physics-Informed Neural Operator Training

License: See LICENSE file · Updated 2026-09-28
My rating

The paper targets general physics simulation tasks using neural operators trained from governing equations without generating a full offline dataset with a classical solver. This physics-informed setup allows fresh samples to be drawn at each optimization step and avoids the associated large offline data-generation cost. The work focuses on preconditioning the ill-conditioned physics-informed loss to improve the training approach.

Tags: