modeling the subgrid-scale kinetic energy in a turbulent channel flow using artificial neural network

Modeling the subgrid-scale energy, , has important applications in large-eddy simulation, including the lattice boltzmann method and formulation of advanced subgrid-scale models. in this study, a deep neural network is specifically developed to predict  for large-eddy simulation of turbulent channel flow. to produce the training data for the neural network, a direct numerical simulation of turbulent channel flow at the friction reynolds number  is performed using an existing highly accurate pseudo-spectral method. the impact of the neural network configuration on its predictions is studied by examining the mean, probability density function and skewness of . moreover, the correlation with the filtered data, its relative and root mean square error are also examined using a priori analysis. appreciable improvements in the predictions were observed with increasing the number of neurons in the hidden layer, up to 64. increasing the number of hidden layers to two and three showed small improvements in the predictions.the performance of the neural network is also compared with a dynamic subgrid-scale model. the comparison reveals that the neural network predictions reach correlation coefficients higher than 90% with the filtered direct numerical simulation data, whereas the dynamic subgrid-scale model predictions only reach up to about 50%. also, a closer agreement was observed with the filtered data for the neural network predictions of , compared with the dynamic subgrid-scale model.