Convolutional Neural Network based Prediction of Turbulent Channel Flow from Cross-Section Information

Abstract

In the present study, a convolutional neural network model based on U-Net architecture is constructed to estimate a turbulent channel flow from measurements on the wall. By assuming that we can measure pressure and shear stresses in streamwise and spanwise directions on a wall, we estimate velocity and pressure at several planes parallel to the walls inside the flow. We have conducted direct numerical simulation of a turbulent channel flow at friction Reynolds number, Re_τ = 150, to build a dataset for the machine learning. We investigate the estimation performances obtained by the present U-Net model. It is found that the estimation accuracy of pressure and three velocity components is high near the wall but deteriorates rapidly away from the wall. In the log layer, the model shows poor estimation performance especially for velocity components. The model can estimate pressure better than velocity components in the whole domain.