Numerically consistent data-driven subgrid-scale model for large-eddy simulation

In this work, we train a data-driven subgrid-scale (SGS) model for large-eddy simulation (LES) which incorporates numerical error arising from the LES equations. LES solves the low-pass filtered Navier-Stokes equations while using a model for unclosed SGS terms. The typical approach for generating data-driven closure models in LES computes required SGS terms using a filtering operator on direct numerical simulation (DNS) data. Recent research highlights that the numerical error between the derivative and filter operators is comparable to the modeling error of traditional SGS models. We develop an artificial neural network (ANN) trained on both filtered DNS data and computed commutation error to act as a numerically consistent closure model in LES. Our goal is to create a high fidelity SGS model for LES which can adapt to different numerical methods in computational fluid dynamics solvers. The results of this ANN-augmented LES model are evaluated in the case of forced isotropic turbulence.