Stochastic Quadrature Rules for Solving PDEs using Neural Networks
Abstract
In this article, we consider issues surrounding integration when using Neural Networks to solve Partial Differential Equations. We focus on the Deep Ritz Method as it is of practical interest and sensitive to integration errors. We show how both deterministic integration rules as well as biased, stochastic quadrature can lead to erroneous results, whilst high order, unbiased stochastic quadrature rules on integration meshes can significantly improve convergence at an equivalent computational cost. Furthermore, we propose novel stochastic quadrature rules for triangular and tetrahedral elements, offering greater flexibility when designing integration meshes in more complex geometries. We highlight how the variance in the stochastic gradient limits convergence, whilst quadrature rules designed to give similar errors when integrating the loss function may lead to disparate results when employed in a gradient-based optimiser.