Comparative study of reinforcement learning and reduced-order model-based control for mitigating vortex-induced vibration

Various active flow control (AFC) algorithms have been developed for vortex-induced vibration (VIV) suppression, but comparative studies on different control strategies remain limited. This study compares reinforcement learning (RL)-based and reduced-order model (ROM)-based closed-loop control algorithms for mitigating VIV. A transversely oscillating cylinder confined between two walls is employed to assess both control strategies, with AFC achieved through the blowing and suction of two synthetic jets mounted on the cylinder. We first introduce and validate the two control frameworks, demonstrating their effectiveness in suppressing VIV at a Reynolds number of 100. Next, dynamic mode decomposition is applied to extract eigenvalues and energy distributions of flow modes during suppression to analyze the differences between the two control strategies. Our results show that the RL-based strategy reduces VIV amplitude to less than 10% of its initial value within 5–6 oscillation periods, whereas the ROM-based strategy requires about 14 periods. Most modal energy concentrates in the first few modes, indicating that these modes primarily govern the flow field characteristics during control for both methods. We find that the RL-based strategy exhibits larger decay rates in the dominant modes, which corresponds to the faster decrease in VIV amplitude in the early control stage. However, the RL-based strategy exhibits low-energy modes with growth rates nearing or exceeding zero, whereas the ROM-based strategy ensures all modal growth rates remain negative. This results in better control performance for the ROM-based strategy during the later stages.