Mitigating the lift of a circular cylinder in wake flow using deep reinforcement learning guided self-rotation

This study applies deep reinforcement learning (DRL) to reduce lift forces on a circular object in the wake of another. The proximal policy optimization (PPO) method is utilized to control the self-rotation of the cylinder using the feedback of sensors. The flow environment is simulated using a high-fidelity computational fluid dynamics (CFD) solver, accelerated by graphics processing unit (GPU). Remarkably, even in the most challenging tandem distance L∗ = 5.0, the DRL agent devises an effective strategy within 800 episodes, resulting in a 98% reduction in lift fluctuation. Flow structure analysis reveals that the learned policy speeds up the shear layer development of the rear cylinder, subsequently adjusting its interaction with the front cylinder's wake. Furthermore, the policy's generalization is assessed at different distances, observing a notable reduction in the effectiveness of lift fluctuation suppression (ranging from 75% to 80%). To enhance the generalization of trained strategies, we optimize the sensors distribution based on the Proper Orthogonal Decomposition (POD) analysis. This leads to faster convergence and consistently better performance, achieving over 88% reduction in lift fluctuation across various distances. This study sheds light on a promising approach for mitigating bluff-body vibrations in complex flows.