Extended Galerkin neural network approximation of singular variational problems with error control (Q6968020)

The authors are concerned with the extension of the Galerkin neural network approach considered in their previous works for second- and fourth-order self-adjoint elliptic PDEs and for singularly perturbed elliptic systems with boundary layer type solutions to quite broad classes of linear non-self-adjoint and/or indefinite problems. The PDEs are supposed to be well-posed in the sense that they admit a unique solution that depends continuously on the data. First, the authors review the basic Galerkin neural network framework, then they provide the extension to non-self-adjoint and non-positive-definite problems, i.e., they get least-squares formulations for bilinear and linear forms of the variational formulations for these problems. The basic idea is to split the solution of the PDE into a part of the solution with high Sobolev regularity and a singular part with low regularity. The latter, also known as a knowledge-based function, highly depends on the geometrical shape of the integration domain. Successfully solving a problem depends on a thorough understanding of these knowledge-based functions. Two interesting examples, one concerning the Poisson problem and the other the Stokes problem, both formulated on non-standard domains, are solved numerically and discussed in detail.