Paperlane

no. 8 · Wednesday 4 November 2026

A posteriori analysis of neural network approximations of partial differential equations

Thomas Führer and Sergio Rojas

math.NA · v2

Abstract

With the aim of defining robust losses for the neural network approximation of partial differential equations, we study, in a general setting, a posteriori estimates used in finite element analysis. We show that the error is equivalent to the sum of two residuals provided that the underlying variational formulation is well-posed. The first contribution is the projection of the residual to a finite-dimensional space and is therefore computable, while the second one can be reliably estimated by a computable upper bound in many practical scenarios. Two important applications can be considered during the neural network optimization (training) process: first, the computable lower bound is used as loss functional and both contributions are used to monitor the error at each solver step; or second, the two estimators are included in the loss functional, thereby providing total control over the error. As a model problem, we consider a second-order elliptic partial differential equation and discuss several variational formulations, including different approaches for incorporating boundary conditions into the estimators. Various numerical experiments are presented to validate our findings.

fields
Numerical Analysis
pdf ↗ page arXiv:2507.06017 ↗