Machine Learning from LES Data to Improve Coarse Grid RANS Simulations

Abstract

Reynolds-averaged Navier-Stokes (RANS) simulations remain the workhorse for engineering computational fluid dynamics (CFD). However, they are still prohibitively expensive for system thermal hydraulic (TH) analysis. One of the ways to reduce the computational cost is to perform simulations on a coarse grid (CG). Unfortunately, this will introduce larger discretization errors in addition to the uncertainties of the turbulence models. Therefore, further advances are needed in CG RANS modeling techniques. In this work, two high-to-low data-driven (DD) approaches are investigated to reduce grid- and turbulence model-induced errors. The approaches are based on: (1) a turbulence model to predict eddy viscosity; (2) correction of errors in velocity. Neural networks (NNs) are trained using large eddy simulation (LES) data for upper plenum of a gas-cooled reactor facility. For approach (1) an inverse optimization problem is solved to extract the eddy viscosity from the LES data.