A neural-network-based mixed model of the subgrid-scale stress for large-eddy simulation of forced isotropic turbulence

An artificial neural network (ANN) is employed to develop an accurate subgrid-scale (SGS) stress model for large-eddy simulation (LES) of forced homogeneous isotropic turbulence. The input variables considered include the filtered strain-rate tensor, the modified Leonard stress tensor, and a combination of them. The data for training the ANN are obtained from a direct numerical simulation of three-dimensional incompressible isotropic turbulence with linear forcing. Both a priori analysis and a posteriori calculation are conducted to evaluate the performance of ANN-based SGS models. It is demonstrated that incorporating the modified Leonard stress tensor into the network architecture significantly improves the predictive performance of ANN-based models. Moreover, the proposed ANN-based mixed SGS model is shown to outperform the traditional dynamic models, such as the dynamic Smagorinsky model, the dynamic Clark model, and the dynamic two-parameter mixed model. In addition, the developed ANN-based mixed model trained using only the database of forced isotropic turbulence performs well in LESs of the transient decaying turbulent flow and the Taylor-Green vortex flow with various Reynolds numbers.