Tensor-Based Neural Network Non-Linear Subgrid-scale Model for Large-eddy Simulation

A tensor-based neural network (TBNN) is employed to obtain a physically-informed data-driven model of the subgrid-scale (SGS) stress tensor for large-eddysimulation (LES) of turbulent flow. Direct numerical simulation (DNS) data of a turbulent channel flow, carried out using a pseudo-spectral method at theReynolds number 𝑅𝑒 = ۲۸۰۰, based on the mean velocity and channel half height, is used for training of the neural network. The model predictions of SGS stress tensor were in good agreement with the filtered DNS data. Anisotropy invariant map of the SGS stresses also showed that model predictions meet realizability conditions and a proper level of SGS anisotropy in case of turbulent channel flow at Re = ۲۸۰۰

Keywords:

Large-eddy simulation , tensor-based neural network , subgrid-scale model