Neural PDE Solvers: New Work on Training, Reuse, and Spatial Adaptivity

A new preprint offers a unified framework for solving partial differential equations (PDEs) that merges two established paradigms. Physics-informed neural networks (PINNs) are typically trained by minimizing pointwise strong-form residuals, while finite element methods rely on weak formulations. The proposed approach combines Gauss-Newton optimization with a Petrov-Galerkin projection, potentially improving accuracy and convergence for both neural and hybrid solvers.

Another preprint tackles the cost of training separate PINNs for every equation. The authors introduce graph hypernetworks that encode the structure of a PDE into a reusable solver, so that a single network can be amortized across a family of related equations. This avoids the need to retrain from scratch when only coefficients or operators change.

A third paper focuses on PDEs with spatially heterogeneous dynamics, where properties vary by location due to geometry or material coefficients. It presents a neural operator whose parameters are generated by a hypernetwork conditioned on spatial position, allowing the model to adapt locally rather than applying a global rule.

While the three works all aim to make neural PDE solvers more practical, they differ in approach: one rethinks the training objective, another improves generalization across equations, and the third adds spatial adaptability. Notably, the two hypernetwork-based papers share a common mechanism but apply it to different problems, and none of the sources directly reference the others. Together, they suggest a broad push toward more flexible and efficient PDE learning.