Uncovering dynamically critical regions in near-wall turbulence using 3D Convolutional Neural Networks

Near-wall regions in wall-bounded turbulent flows experience strong intermittent events involving ejections of slow-moving fluid parcels away from the wall, and `sweeps' of faster moving fluid towards the wall. Here, we train a three-dimensional Convolutional Neural Network (CNN) to predict the intensity of ejection events that occur in Direct Numerical Simulation (DNS) of a periodic channel flow. The trained network is able to predict burst intensities accurately for flow snaphshots that are sufficiently removed from the training data so as to be temporally decorrelated. More importantly, we probe the trained network to reveal regions of the flow where the network focuses its attention in order to make a prediction. We find that these salient regions correlate very well with fluid parcels being ejected away from the wall. Moreover, the CNN is able to keep track of the salient fluid parcels as the flow evolves in time. This demonstrates that CNNs are capable of discovering dynamically critical phenomena in turbulent flows without requiring any a-priori knowledge of the underlying dynamics. Remarkably, the trained CNN is able to predict ejection intensities accurately for data at different Reynolds numbers, which highlights its ability to identify physical processes that persist across varying flow conditions. The results presented here highlight the immense potential of CNNs for discovering and analyzing nonlinear spatial correlations in turbulent flows.Comment: 10 pages, 7 figure