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

Abstract

Near-wall regions in wall-bounded turbulent flows experience intermittent ejection of slow-moving fluid packets away from the wall and sweeps of faster moving fluid towards the wall. These extreme events play a central role in regulating the energy budget of the boundary layer, and are analyzed here with the help of a three-dimensional (3D) Convolutional Neural Network (CNN). A CNN is trained on Direct Numerical Simulation data from a periodic channel flow to deduce the intensity of such extreme events, and more importantly, to reveal contiguous three-dimensional salient structures in the flow that are determined autonomously by the network to be critical to the formation and evolution of ejection events. These salient regions, reconstructed using a multilayer Gradient-weighted Class Activation Mapping (GradCAM) technique proposed here, correlate well with bursting streaks and coherent fluid packets being ejected away from the wall. The focus on explainable interpretation of the network's learned associations also reveals that ejections are not associated with regions where turbulent kinetic energy (TKE) production reaches a maximum, but instead with regions that entail extremely low dissipation and a significantly higher tendency for positive TKE production than negative production. This is a key finding of the study, and indicates that CNNs can help reveal dynamically important three-dimensional salient regions using a single scalar-valued metric provided as the quantity of interest, which in the present case is the ejection intensity. While the current work presents an alternate means of analyzing nonlinear spatial correlations associated with near-wall bursts, the framework presented is sufficiently general so as to be extendable to other scenarios where the underlying spatial dynamics are not known a-priori.