Study of Feedback Control Strategy Obtained by Deep Reinforcement Learning for Flow Separation Control around Airfoil in 3D Simulation

Abstract

In this study, we performed deep reinforcement learning-based flow control using plasma actuator in a 3D CFD to improve the control system of a wind tunnel experiment. To achieve this, we analyzed the control strategy based on the trained controller using detailed visulizations of the flow field and control history. The target flow field is around the NACA0015 airfoil, and the flow condition is Re = 6.3 × 10⁴, AoA = 14°, which is over 3° higher than the stall angle. As a result, we confirmed that the same control strategy was trained in the 3D CFD as in the experiment. Moreover, the trained controller achieved a higher L/D than conventional predetermined controls by combining F⁺ = 1 and F⁺ = 6.