WorksMathematics
A Staggered Discontinuous Galerkin Neural Network Method with a Computable Error Decomposition
Abstract
We develop a staggered discontinuous Galerkin neural network (SDGNN) method for scalar elliptic problems with a computable error decomposition. The staggered discontinuous Galerkin (SDG) method requires only local continuity of its trial functions and gives a square, inf-sup stable system without stabilization parameters; we carry this structure over to neural networks. Networks defined on the patches of a staggered mesh satisfy the SDG continuity conditions for all values of the network parameters, and the loss is the squared dual norm of the SDG weak residual. We show that the loss is equivalent, with computable constants, to the square of the training error, the distance between the SDG projection of the network and the SDG finite element approximation, so that the full error is bounded by the discretization error, a multiple of the square root of the loss, and the network projection error. The weak residual requires no derivatives of the networks, the boundary and convection terms carry no penalty or stabilization weights, and the scalar trial functions may jump across the midlines of the primal cells. Numerical experiments show that the square root of the loss tracks the error during the training, that prescribed jumps across a square and a circular interface are captured consistently over the initializations, and that the method is more accurate than physics-informed neural networks at the same parameter counts.