Structure-Informed Neural Networks for Boundary Observation Problems

Abstract: We introduce Structure Informed Neural Networks (SINNs), a novel method for solving boundary observation problems involving PDEs. The SINN methodology is a data-driven framework for creating approximate solutions to internal variables on the interior of a domain, given only boundary data. The key idea is to use neural networks to identify a co-ordinate transformation to a latent space, upon which a well-posed elliptic system of partial differential equations is constructed. The use of elliptic systems enables the low-cost transfer of information from the domain's boundary to its interior. This enables approximate solutions to PDE boundary observation problems to be constructed for generic, and even ill-posed, problems. A further advantage of the proposed method is its ability to be trained on experimental or numerical data without any knowledge of the underlying PDE. We demonstrate the ability of SINNs to accurately solve boundary observation problems by considering two challenging examples of a non-linear heat equation and boundary observation for the Navier-Stokes equations.