Boundary-preserving randomized neural network methods for Poisson equations on general two- and three-dimensional domains

Abstract

In this paper, we propose a boundary-preserving randomized neural network method for Poisson equations with Dirichlet boundary conditions in two and three dimensions. A linear lifting imposes the boundary conditions exactly, without balancing penalty weights between interior and boundary residuals. We construct the liftings by generalized barycentric coordinates and transfinite interpolation on convex and nonconvex polygons and polyhedra, and by radial parameterization on star-shaped curved domains. Numerical experiments on different domains show that the proposed method substantially reduces errors compared with the standard randomized neural network method.