AI-Assisted Wall-Modelled LES and Flow Control¶
Matthias Meinke (RWTH Aachen)
Abstract¶
Two applications of artificial intelligence in aerodynamics and fluid mechanics will be presented. The first application concerns wall functions for wall-modelled large-eddy simulation (WM-LES). Results obtained using the classical Spalding wall model are presented for the canonical case of a zero-pressure-gradient turbulent boundary layer and for transonic airfoil flow involving shock buffet. Ongoing work on AI-derived wall models is then summarized, with a focus on improving the prediction of fluctuating wall-shear-stress components. An autoencoder is trained on wall-resolved LES data and subsequently applied to spatially filtered velocity fields, or used within WM-LES, to reconstruct the velocity components in the near-wall region. A-priori tests and preliminary a-posteriori results show improved predictions of fluctuating flow-field components at the wall. The second application addresses the effective control of fluid flows, which is important in many fields, including aerodynamics and aeroacoustics. Fluid motion, however, is difficult to control because it is governed by high-dimensional, nonlinear, and multiscale dynamics that challenge conventional approaches. To facilitate the development and training of AI-based tools, HydroGym has been implemented as a solver-independent reinforcement learning (RL) platform that provides more than 60 validated and openly available flow-control environments, ranging from canonical laminar flows to complex turbulent configurations.