Evo-GTransNet for Parabolic PDEs: A Fixed-Feature Galerkin Method of Lines with Quadrature-Mass Orthonormalization

What the math gives to ML

The paper's transferable asset is not the use of neural features itself, but the separation of a large fixed nonlinear dictionary from a smaller, numerically resolved coefficient space. A quadrature-weighted SVD identifies feature directions that are observable under the discretized domain inner product, while a second rescaling makes the retained functions orthonormal without changing their span. This converts ill-conditioned neural-feature Galerkin systems into identity-mass latent dynamics with explicit energy and contractivity guarantees, suggesting a practical preconditioning and rank-compression layer for neural PDE solvers and continuous-domain latent models.