Deep learning of turbulent heat transfer in channel flow

With the recent development of artificial intelligence(A.I.) and its wide applications, a fundamental question arises such as 'can turbulence be learned by A.I.?' In order to provide an answer to this question, we applied deep learning to the prediction of turbulent heat transfer based only on the wall shear and pressure information which was obtained by DNS. Through this study, we also tried to see whether deep learning can help our understanding of the physics of turbulent heat transfer. Under the assumption that the wall normal heat flux can be explicitly expressed through multilayer nonlinear network in terms of nearby wall shear stresses and wall pressure fluctuations, we applied convolutional neural networks to predict the local heat flux. After an optimization of the network models, we found that the network can predict heat transfer very well with correlation of 0.980 between DNS and prediction by the network for trained $Re$, and show similar performance at a $Re$ three times higher than the trained one, indicating that relations between the wall shear and the heat flux are almost independent of $Re$ within tested range. Additionally, through a sensitivity analysis of the trained model, we found which part of the input data is important in the prediction of heat flux.