Reinforcement learning (RL)-based closed-loop flow control shows great potential for managing nonlinear and complex aerodynamic flows. In this study, we investigate RL-based flow control to enhance the lift-to-drag ratio of an NLF(1)-0115 airfoil at a chord-based Reynolds number of \( Re_c = 20{,}000 \) and an angle of attack \( \alpha = 5^\circ \) . Key control parameters, including the reward function, agent action time, observed state, and actuator placement, are systematically examined. Our results reveal that physics-informed tuning significantly improves control performance. An optimal agent action time of \( \tau = 0.07 \) , corresponding to approximately 11% of the primary oscillation period, was identified. It broadens the induced forcing spectrum, enhancing interaction with a wider range of flow structures. This finding establishes a clear physical connection between the RL agent action time and the unsteady flow dynamics. However, excessively short agent action time reduce the forcing amplitude, limiting control authority. Adjusting the observed state from wake-region sensors to surface-mounted pressure sensors yields comparable improvements in lift-to-drag ratio, ranging from \( 34.1\% \) to \( 35.5\% \) , demonstrating the practical feasibility of using surface measurements. Forcing placement based on stability analysis significantly enhances control effectiveness. To improve data efficiency, the optimized 2D RL controller is transferred to a 3D CFD environment through prescribed spanwise wavenumber superposition. This lower-cost 3D controller effectively suppresses flow separation and significantly enhances aerodynamic performance. The results present a promising and practical alternative to direct 3D RL training for airfoil flow control.
