Active Airfoil Flow Control Based on Reinforcement Learning

We develop a reinforcement learning-based closed-loop active flow control strategy for alleviating laminar separation from the suction surface of a wing section with NLF(1)-0115 airfoil at five degrees angle of attack. The impact of key parameters such as actuation location, observed states, and reward function on the effectiveness of the control are investigated for a chord Reynolds number of Re = 5, 000. We then test the control agent at higher Reynolds numbers. Control authority is maintained at Re = 10, 000 and Re = 20, 000. While the controller increases lift at Re = 10, 000. At Re = 20, 000 it reduces lift and increases the aerodynamic efficiency. A proper orthogonal decomposition indicates that this change can be attributed to a transition of the primary unsteady mode of the uncontrolled flow from a wake mode to a shear layer mode at Re = 20, 000. The present study offers valuable insights into estimating the effectiveness of reinforcement learning-based control agents at higher Reynolds numbers based solely on their low-Reynolds number performance. It highlights the crucial role played by flow instabilities on reinforcement learning-based active flow control strategies and establishes a predictive framework for the scalability of such strategies in aerodynamic applications.