Floquet-guided reinforcement learning for active flow control of a low-Reynolds-number three-dimensional wing

Abstract

Reinforcement-learning (RL) control of three-dimensional separated flows is computationally demanding because the number of sensor inputs and independent actuator outputs increase rapidly with spanwise resolution. This study develops a Floquet-guided RL strategy for active flow control of a National Advisory Committee for Aeronautics 0012 straight wing at a chord-based Reynolds number of Re=500. Floquet analysis of the two-dimensional time-periodic baseline flow identifies the most amplified spanwise wavelength, which is embedded in a harmonic actuation basis so that the agent optimizes only the time-dependent amplitude and phase of each blowing slot. The resulting four-dimensional action space replaces a much larger set of independently controlled spanwise actuators. The learned policy increases the period-averaged lift by approximately 7.29%. Flow-field analysis shows that the spanwise-modulated blowing generates streamwise vortices that transport higher-momentum fluid toward the separated upper-surface region, extend the low-pressure region downstream, and improve resistance to the adverse pressure gradient. These results demonstrate that linear stability information can provide a physically interpretable low-dimensional control basis for RL in a laminar/transitional wing flow.