Closed-loop machine learning control of a cavity flow using adaptive plasma actuators at moderate Reynolds number

A real-time machine learning control of the cavity flow was experimentally implemented at a moderate Reynolds number of 48 000, operating at a 4 kHz interaction frequency. A fast-response dielectric barrier discharge plasma actuator and a hot-wire anemometer probe were served as the flow perturbation device and the state sensor. Aiming to achieve maximum attenuation of the single-tone cavity noise, the particle swarm optimization (PSO) and genetic programming (GP) algorithms were employed to identify optimal actuation parameters within an open-loop framework and derive an optimal closed-loop control law, respectively. This approach yielded three distinct effective control strategies characterized by unique shear-layer vortex dynamics. The first strategy, identified via PSO as the optimal unsteady periodic actuation, modifies the noise spectrum through a “partial-escape partial-clipping” mechanism applied to shear-layer vortices, shifting the dominant characteristic frequency from St = 1.606 to St = 1.2. The second strategy, involving steady high-intensity plasma actuation, significantly suppresses shear-layer instability. This results in the near-exclusive “complete escape” of nascent vortices above the cavity trailing edge, with minimal clipping effect observed. In contrast, the third conditional actuation strategy, derived via GP, employs a novel manipulation approach: vortices are actively directed either downward into the cavity or upward, bypassing the downstream corner entirely. This alternation between “complete-escape” and “complete-clipping” mechanisms induces a 94% attenuation in the modal noise peak. Notably, the associated power consumption for this strategy is merely 31% of that required for the steady actuation strategy.