Surrogate modeling of fluid dynamics with a multigrid inspired neural network architecture

Algebraic or geometric multigrid methods are commonly used in numerical solvers as they a multi-resolution method able to handle problems with multiple scales. In this work, we propose modification the commonly-used U-Net neural network architecture that is inspired by principles of methods, referred here U-Net-MG. We then demonstrate proposed U-Net-MG can successfully reduce test prediction errors relative conventional when modeling set fluid dynamic problems. total, an improvement velocity and pressure fields for canonical dynamics cases flow past stationary cylinder, 2 cylinders out-of-phase motion, oscillating airfoil both propulsion energy harvesting modes. general, while models model systems well RMSEs less than 1%, use further between 20% 70%. • A multigrid-inspired CNN proposed. Canonical such foil tested. Model extrapolates across different geometries inlet boundary conditions. reduces RMSE 20%–70% baseline model. Scientific computing concepts improve ML design engineering