Applying Deep Reinforcement Learning with High-Fidelity Fluid Simulations for Separation Control over an Airfoil¶
Tomoaki Tatsukawa (Tokyo University of Science)
Abstract¶
DBD plasma actuators (PAs) have gained attention as microflow control devices for suppressing separated flows around airfoils. In particular, burst actuation using a non-dimensional burst frequency (F+) demonstrates superior control performance compared to continuous actuation. However, determining the appropriate F+ for varying flow conditions remains challenging. This study proposes a simulation framework that combines high-fidelity simulation and deep reinforcement learning (DRL) to identify effective burst actuation strategies. The framework exchanges information such as state, reward, and action between a supercomputer running the CFD solver and a machine executing the DRL program. The DRL agent determines the optimal F+ based on pressure data obtained from sensors on the airfoil surface. The reward is based on trailing-edge pressure, and repeated learning episodes enable the agent to develop control strategies. The results of the episodes demonstrated that control strategies capable of suppressing trailing-edge separation could be obtained without prior knowledge.