Active Flow Control Strategies Based on Deep Reinforcement Learning

Abstract We demonstrate the effectiveness of deep reinforcement learning (DRL) in active flow control around a cylinder at Re = 500. A DRL-based algorithm was employed to manipulate synthetic jets for controlling the flow field. The results show that the DRL agent achieves a 49% reduction in drag coefficient and significantly stabilizes lift coefficient fluctuations, showcasing its capability to adaptively optimize control strategies. The vorticity evolution during training reveals a progressive suppression of the Kármán vortex street, leading to a fully stabilized wake and smoother flow dynamics. Substantial improvements are observed in the streamwise and cross-stream velocity fields, with the wake transitioning from a chaotic, unsteady state to a stabilized and streamlined structure. Additionally, the pressure field evolves from an asymmetrical, fluctuating distribution to a stabilized, symmetrical profile, reducing drag and ensuring uniform pressure recovery. These findings highlight the potential of DRL-based strategies for robust and energy-efficient flow control.