Zero-shot generalizable deep reinforcement learning for drag-reduction control on turbulent wing sections

This study presents a novel approach to applying deep reinforcement learning (DRL) for drag-reduction control of the turbulent boundary layer (TBL) over a wing surface. While most prior DRL-based wing control efforts have focused on two-dimensional simulations, this work targets the suction side of a 3D wing section at a chord-based Reynolds number of 200,000 and an angle of attack of 5^○.

Although DRL has shown great potential, the training process remains computationally expensive due to the need for many iterations and trajectories to learn an optimal policy. The cost is further amplified in wing applications due to the wide range of turbulent scales in TBLs, making each iteration time-consuming. Moreover, spatial inhomogeneity limits the effective control area.

To address these challenges, we propose a two-step strategy: (1) Divide the control region into blocks and train DRL policies in channel flows that replicate each block’s local flow characteristics; (2) Deploy the learned policies to their corresponding regions on the wing.

This method significantly enhances computational efficiency and enables wide-area control, outperforming the state-of-the-art opposition control approach.