High-Fidelity Simulations of Active Flow Control Over an Airfoil with Deep Reinforcement Learning

Abstract

A dielectric barrier discharge plasma actuator (PA) is a flow control device that can suppress flow separations. Previous studies have shown that burst actuation is especially effective for this purpose. An experimental study applied deep reinforcement learning (DRL) to the flow separation control and conducted feedback control to dynamically determine the non-dimensional burst frequency F⁺, according to the flow field. In this study, we make a framework coupling the DRL and large-eddy simulations whose fidelity reproduces the unsteady flow features to investigate the detailed flow fields controlled by PA with DRL. The framework found the optimal burst frequency clarified by the previous study at the angle of attack, 12 degrees. The suppression of flow separation was not completely achieved, but the results showed that the DRL models temporarily improved the lift coefficient, and a characteristic burst drive appeared at the angle of attack of 15 degrees.