Numerical investigation of wall modeling for LES using convolutional neural network

Wall modeling in large eddy simulation (LES) is crucial as the computational cost of LES rises significantly for high Reynolds number wall-bounded flows which can make the LES impossible for such flows. Therefore, it is important to develop a wall model with reasonable accuracy and computational cost. In the recent years, with the increase in the computational resources and the proven ability of data-driven approaches in making predictions, they have become a popular tool for different applications in the fluid mechanics including turbulence modeling in LES. In the present study, the convolutional neural network (CNN) is used as a tool to develop a data-driven nonlocal wall stress model for the LES of turbulent channel flow. First, a hyperparametric study is performed and the model performance is checked in the a priori test. Finally, the model is embedded in an actual LES to investigate how well the model performs in the simulation. In the a posteriori test, initially, the model is tested for the same condition as used for the training. Then, the generalizability of the model is checked by using the model under various conditions different from those used for the training data. The results show that CNN is successful in establishing a wall model with simple input choices and has reasonable accuracy in predicting the wall shear stress and flow field.