Active flow control of a bubble of recirculation in a turbulent boundary layer through deep reinforcement learning

Abstract: The phenomenon of a wing near stall can be resembled in a simplified manner through a turbulent boundary layer subjected to an adverse pressure gradient that originates a bubble of recirculation. From classical theory, in order to control this kind of flows, a periodic forcing with a zero-net-mass flux (ZNMF) is performed through the application of blowing and suction of rectangular jets placed right upstream the bubble of recirculation. In this talk I will compare the results obtained using deep reinforcement learning (DRL) as the new approach with respect to the classical control, using the same setup in both cases. A DRL agent controls the magnitude of the blowing or suction of the control jets while ensuring a ZNMF too. We found that DRL manages to obtain a higher bubble reduction than classical control. Furthermore, the reduced bubble in the periodic forcing is highly correlated to the phase of the forcing, while DRL strategy provides a smoother control. Furthermore, a entire new framework that communicates DRL libraries (TF-agents) with a GPU-solver (SOD2D), through memory (SmartSim) has been developed, suited for the upcoming generation of exascale computing systems.

Bio: Francisco Alcántara Ávila is a postdoctoral researcher at KTH Royale Institute of Technology. Francisco obtained his PhD degree in 2021 at the Universidad Politécnica de Valencia, Spain, where he conducted direct numerical simulations of heat transfer at high Reynolds numbers. His main field of research is the application of deep reinforcement learning into active flow control problems, which he has been working on during his postdoctoral research.