Using deep neural networks for detecting spurious oscillations in discontinuous Galerkin solutions of convection-dominated convection-diffusion equations (Q6080848)

Application of discontinuous Galerkin finite element methods to convection-dominated convection-diffusion equations produce usually sharp layers, but also exhibit large unphysical oscillations. Slope limiters are used to reduce spurious oscillations. In this paper, deep neural networks are constructed for detecting mesh cells on which slope limiters are applied. The paper is focused on constructing a multilayer perceptron (MLP) model for limiting the discrete solution of an elliptic problem, i.e., convection-diffusion equations in the convection-dominated regime. Several neural network architectures are trained with the MLP limiter and are tested numerically. It was shown how data from a lowest order discretization can be used to train a limiter that then can be applied to the discrete solution of higher order methods. The results show that the limiter works almost equally well as classical methods for higher order methods for the same problem but somewhat worse than these methods when applied to the solution of a different problem.