Robust Airfoil Flow Control Under Free-Stream Turbulence via Reinforcement Learning

We investigate reinforcement learning (RL)-based active flow control over a wing section with the NLF(1)-0115 airfoil at a Reynolds number of Re=20,000 and an angle of attack of alpha=5°. A computational fluid dynamics solver serves as the environment, providing flow states and evaluating rewards for the RL-based controller. The controller is trained without free-stream turbulence using three reward functions. Each incorporating a different penalty value of 0, 10 and 20. Among them, the reward r=Cl/Cd-10 yields a better control performance. To assess the robustness of the control, the trained controller is then tested under increasing levels of free-stream turbulence, 0.1%, 0.5% and 1%, representing wind tunnel conditions. Results show significant performance degradation, highlighting the importance of including turbulence effects during training to develop more robust controllers. We also explore the reconstruction of wake velocity fields using surface pressure time histories as input to neural network (NN) and convolutional neural network (CNN) models. While the NN exhibits higher accuracy within the training domain, the CNN shows better generalization to unseen cases. This analysis provides insights into the observability of wake dynamics from surface data and supports the development of data-driven estimators for closed-loop flow control.