Reinforcement Learning-Based Closed-Loop Airfoil Flow Control

Abstract

We systematically investigated a reinforcement learning (RL)-based closed-loop active flow control strategy to enhance the lift-to-drag ratio of a wing section with an NLF(1)-0115 airfoil at an angle of attack 5 degree. The effects of key control parameters, including actuation location, observed state, reward function, and control update interval, are evaluated at a chord-based Reynolds number of Re=20,000. Results show that all parameters significantly influence control performance, with the update interval playing a particularly critical role. Properly chosen update intervals introduce a broader spectrum of actuation frequencies, enabling more effective interactions with a wider range of flow structures and contributing to improved control effectiveness. The optimally trained RL controller is further evaluated in a three-dimensional numerical setup at the same Reynolds number. Actuation is applied using both spanwise-uniform and spanwise-varying control profiles. The results demonstrate that the pretrained controller, combined with a physics-informed spanwise distribution, achieves substantial performance gains. These findings extend the feasibility and scalability of a pretrained RL-based control strategy to more complex airfoil flows.