Simultaneous CNN Approximation on Manifolds with Applications to Boundary Value Problems

Researchers introduce a CNN approximation framework for elliptic boundary value problems on compact Riemannian manifolds, proving Sobolev convergence rates governed by intrinsic dimension d rather than ambient dimension D, directly attacking the curse of dimensionality. The physics-informed CNN variant replaces standard L2 boundary penalties with a spectral loss built on the boundary Laplace-Beltrami operator, enforcing proper Sobolev trace control that classical PINNs lack. Practitioners solving PDEs on curved geometries, medical imaging meshes, or climate models on spherical domains should take note, as this offers theoretically grounded convergence guarantees absent in vanilla PINN formulations.