Deep Reinforcement Learning for Active Flow Control of Vortex-Induced Vibrations: An Experimental Study

Deep reinforcement learning (DRL) has recently emerged as a promising data-driven approach for active flow control (AFC). A particularly important and practically relevant application of AFC is the suppression of vortex-induced vibrations (VIV), which arise when unsteady wake forces couple with the natural modes of freely oscillating structures. Prior studies have shown that DRL can discover effective control strategies for VIV suppression in simulated environments. However, the application of these methods to real-world systems remains limited, primarily due to challenges such as hardware integration, sensor noise, latency in control loops, and the stochastic nature of physical flow environments.

To address these limitations, we implement a DRL-based closed-loop control strategy in a physical experiment aimed at suppressing VIV by controlling the rotational speed of a transversely oscillating circular cylinder. The experimental setup features a circular cylinder mounted on a low-friction air-bearing system, allowing transverse oscillations under vortex-induced vibrations in a water tunnel at Reynolds number 4000. Actuation is applied via a DC motor to control the cylinder’s angular velocity. Real-time sensor data—including transverse displacement (via laser range sensor), acceleration (via accelerometer), and hydrodynamic forces (via six-axis force transducer)—are acquired through a synchronized measurement system and used to construct the control state vector. A deep reinforcement learning agent based on the proximal policy optimization (PPO) algorithm is trained directly on the physical system using a reward function that penalizes transverse displacement. Training accounts for sensor noise, actuation delays, and unsteady flow conditions. The resulting control policy achieves substantial vibration reduction. Moreover, flow visualization using particle image velocimetry (PIV) reveals suppression of unsteady wake structures and a reorganization of vortex shedding. These results demonstrate the viability of DRL for experimental active flow control and offer guidance for its deployment in complex fluid–structure systems.