SGS backscatter effects in coarse-grid LES predicted by a machine-learning-based SGS model

In this talk, a machine-learning-based sub-grid scale (SGS) modeling for coarse-grid large-eddy simulation (LES) is proposed. The proposed SGS model consists of an unsupervised and supervised machine learning model to enable accurate prediction of the SGS stresses for unconventionally coarse grid LES.

In the a posteriori test using the fully-developed turbulent channel, the coarse-grid LES with the proposed model shows a good prediction of the Reynolds shear stress and the resultant mean streamwise velocity, while a conventional SGS model fails the prediction. The difference in prediction accuracies between the two SGS models originates from the near-wall Reynolds shear stress.

Budget analyses of the Reynolds normal stresses reveal that the SGS backscatter predicted by the proposed SGS model significantly increases the spanwise Reynolds stress in the near-wall region. The near-wall spanwise stress is then redistributed to the wall-normal component through the pressure-strain term, giving rise to the increased near-wall Reynolds shear stress. In contrast, the conventional SGS model without the backscatter does not show such a process, leading to the under-prediction of the near-wall Reynold shear stress and the consequential over-prediction of the mean velocity.