Turbulent cylinder wake stabilization with intelligent distributed plasma actuators powered by deep reinforcement learning

This study demonstrates experimental deep reinforcement learning (DRL) control of a turbulent cylinder wake at a Reynolds number of 28 000. Two dielectric barrier discharge plasma actuators operated independently to seed disturbances to the cylinder boundary layer, and two hot-wire sensors submerged within the wake provide real-time feedback. Following the Deep Q-Network framework, the closed-loop control law is parameterized by a fully connected neural network and executed at 1 kHz on a field-programmable gate array. Results show that the DRL agent is able to identify an effective control policy in a single training run without prior knowledge, achieving the same-order drag reduction magnitude as the steady open-loop control by utilizing only 40% actuator power consumption of the latter. The learning process, however, is sensitive to the setting of hyperparameters, and poor reward configurations can lead to an unstable learning process. Diagnostics via a high-speed particle image velocimetry and a dimensionality reduction approach reveal that the agent learns a conditional actuation strategy. Namely, when the wake has a tendency to deflect significantly toward one side, intense plasma actuation should be applied on the opposite side to suppress this lateral oscillation, stabilizing the flow to high-reward states and thereby reducing drag reduction. These results underscore the potential of DRL for adaptive, energy-efficient control in turbulent flows.