Airfoil pitch control in the wake of a backward-facing step via spatiotemporal deep reinforcement learning

Closed-loop lift control for an airfoil in a strongly nonuniform inflow is challenging, especially when only partial surface-pressure observations are available. This study investigates pitch control of a NACA (National Advisory Committee for Aeronautics) 0012 airfoil descending through the wake of a backward-facing step. A spatiotemporal deep-reinforcement-learning-based control framework is developed, in which the controller adopts sparse surface pressure and airfoil kinematics as feedback. The proposed MSALSTM policy, trained with proximal policy optimization, combines multi-head self-attention for spatial pressure encoding with long short-term memory for temporal inference of hidden flow states. Among the learned architectures, MSALSTM provides the most stable lift tracking and the smallest pitch oscillation. It also outperforms the fixed-angle and gain-scheduled proportional–integral–derivative baselines under the same descent condition. When a drag penalty is added to the reward function, the learned policy further increases the mean lift-to-drag ratio by 46%. Further analysis reveals that the attention module adaptively redistributes sensor emphasis according to the evolving flow state, while the recurrent memory organizes pressure-history information into structured hidden-state trajectories. These results demonstrate that spatiotemporal learning can reconstruct effective flow-state information from sparse wall-pressure data for high-Reynolds-number aerodynamic control.