Optimizing aerodynamic stability in compressible flow around a vibrating cylinder with deep reinforcement learning

This paper explores the use of Deep Reinforcement Learning (DRL) to improve the aerodynamic stability of compressible flow around a vibrating cylinder. In uncontrolled conditions, the cylinder experiences a drag coefficient of 1.35 and an oscillatory lift coefficient with an amplitude of 0.35. By applying a classic Deep Q-Network (DQN), the lift oscillation amplitude is significantly reduced to ±0.025, marking an improvement of over 100%. The study further investigates the effects of episode count, neural network architecture, and DQN variants on performance, demonstrating the robustness of the approach. While changes to the neural network structure within the classic DQN yield limited improvements in reducing lift oscillations, both the classic and dueling DQN variants effectively control lift oscillations. Notably, the dueling DQN provides greater stability, reducing lift oscillation amplitude to as low as ±0.001. The paper also examines the effect of varying jet positions, offering valuable insights into the adaptability and generalization of the proposed DRL-based control strategy.