Investigation of nonlocal data-driven methods for subgrid-scale stress modelling in large eddy simulation

TL;DR

This paper develops a nonlocal CNN-based subgrid-scale stress model for large eddy simulation, demonstrating improved accuracy, stability, and robustness over local models and traditional SGS models in turbulent flow simulations.

Contribution

It introduces a novel nonlocal CNN-based SGS model trained on DNS data, outperforming local models and traditional approaches in LES of turbulent flows.

Findings

Nonlocal CNN models outperform local models like ANN and Smagorinsky in LES.

The models accurately predict backscatter and maintain stability across different grid resolutions.

The approach effectively extrapolates to higher Reynolds numbers and complex turbulent flows.

Abstract

A nonlocal subgrid-scale stress (SGS) model is developed based on the convolution neural network (CNN), a powerful supervised data-driven approach. The CNN is an ideal approach to naturally consider nonlocal spatial information in prediction due to its wide receptive field. The CNN-based models used here only take primitive flow variables as input, then the flow features are automatically extracted without any $priori$ guidance. The nonlocal models trained by direct numerical simulation (DNS) data of a turbulent channel flow at $Re_{\tau}=178$ are accessed in both the $priori$ and $posteriori$ test, providing physically reasonable flow statistics (like mean velocity and velocity fluctuations) closing to the DNS results even when extrapolating to a higher Reynolds number $Re_{\tau}=600$. In our model, the backscatter is also predicted well and the numerical simulation is stable. The nonlocal models outperform local data-driven models like artificial neural network and some SGS models, e.g. the Smagorinsky model in actual large eddy simulation (LES). The model is also robust since stable solutions can be obtained when examining the grid resolution from one-half to double of the spatial resolution used in training. We also investigate the influence of receptive fields and suggest using the two-point correlation analysis as a quantitative method to guide the design of nonlocal physical models. To facilitate the combination of machine learning (ML) algorithms to computational fluid dynamics (CFD), a novel heterogeneous ML-CFD framework is proposed. The present study provides the effective data-driven nonlocal methods for SGS modelling in the LES of complex anisotropic turbulent flows.