Mixed subfilter-scale models for large-eddy simulation of decaying isotropic turbulence using an artificial neural network

This study is concerned with the development of a new subfilter-scale (SFS) stress model for large-eddy simulation (LES) of decaying isotropic turbulence using an artificial neural network (ANN). Both a priori and a posteriori tests are performed to investigate the effect of input variables on the performance of ANN-based SFS models. Within the range of parameters and flow types considered, the ANN-based model with filtered strain-rate tensor as input is found to show excellent predictions of the resolved statistics in a posteriori test, although it provides low correlation coefficients between the true and predicted SFS stresses in a priori test. However, this model performs poorly in the predictions of the SFS statistics and backscatter. On the other hand, the predictive accuracy of ANN-based models is significantly improved by using a combination of the strain-rate tensor and the modified Leonard stress tensor as input variables. The proposed ANN-based mixed SFS model not only can predict the backscatter, but also exhibits better performance in predicting the resolved and SFS statistics than the traditional dynamic models. In particular, the ANN-based mixed model shows an advantage over the dynamic two-parameter mixed model in terms of the accuracy and computational efficiency.