Deep Reinforcement Learning Discovers a Novel Control Algorithm for Mitigating Flow-Induced Vibrations in Underactuated Tandem Cylinders

Abstract

This study presents the first experimental implementation of deep reinforcement learning (DRL) for the active real-time suppression of flow-induced vibrations in simultaneously vibrating tandem cylinders using rotary actuation, considering fully actuated and underactuated configurations. In the fully actuated case, where both cylinders are independently controlled, the DRL agent discovers a high-frequency, phase-locked bang-bang control strategy that suppresses the vibrations of both cylinders by more than 95\%. Analysis of the training dynamics reveals a physically interpretable learning process in which the agent first identifies the optimal phase relationship between the actuators before refining the actuation frequency. In the underactuated configuration, where only the upstream cylinder is actuated, equally weighted rewards produce ineffective control, suppressing vibrations only in the actuated cylinder. Introducing asymmetric reward weighting enables the DRL agent to discover a low-frequency lock-on strategy that achieves 70\% and 90\% vibration suppression in the upstream and downstream cylinders, respectively. For staggered arrangements with lateral offset, conventional training fails to converge, requiring a curriculum learning approach. The resulting two-stage curriculum identifies a statically biased bi-harmonic rotational control signal capable of suppressing vibrations in both cylinders. The success of the underactuated control strategy highlights its potential to reduce energy consumption and hardware complexity in multi-body flow control systems.