Active flow control on airfoils by reinforcement learning

Abstract

Active flow control is a widespread practice for airfoil aerodynamic performance enhancement. Within active flow control, reactive strategies are very effective, but the adequate design of these strategies is often complex. This study proposes a reactive control strategy based on a Reinforcement Learning (RL) agent to effectively govern the motion of a rotating flap implemented on a NACA0012 airfoil. With this objective, first different Computational Fluid Dynamics (CFD) simulations are conducted to gather data about the tested case. Then, a numerical model based on Artificial Neural Networks (ANN) is developed to model the discussed case. Finally, the RL agent is trained and tested under different conditions. The results show that the trained RL agent is able to provide a fast and reliable response for every tested condition, setting the adequate position of the flap and obtaining an appropriate aerodynamic performance of the airfoil for all the tested conditions. In comparison with the optimum conditions, the absolute error in the position of the flap set by the agent is below 2.2 for all the angles of attack, resulting in an aerodynamic performance very close to the optimum, being only 0.39%–3.05% lower, depending on the case.