Abstract
We propose a combination of machine learning and flux limiting forproperty-preserving subgrid scale modeling in the context of flux-limitedfinite volume methods for the one-dimensional shallow-water equations. Thenumerical fluxes of a conservative target scheme are fitted to the coarse-meshaverages of a monotone fine-grid discretization using a neural network toparametrize the subgrid scale components. To ensure positivity preservation andthe validity of local maximum principles, we use a flux limiter that constrainsthe intermediate states of an equivalent fluctuation form to stay in a convexadmissible set. The results of our numerical studies confirm that the proposedcombination of machine learning with monolithic convex limiting producesmeaningful closures even in scenarios for which the network was not trained.