Active wake control of a rectangular cylinder near a free surface via mode-aware deep reinforcement learning

This study investigates active wake control of a rectangular cylinder near a free surface using a mode-aware deep reinforcement learning (DRL) framework aimed at reducing drag and suppressing vortex shedding. The flow is modeled as a two-dimensional incompressible two-phase system, with the air–water interface captured by the volume-of-fluid method. Proper Orthogonal Decomposition (POD) is first used to identify the region of strongest flow fluctuations, which is then encoded by a convolutional neural network into 12 latent signals as compact observations for the DRL agent. Closed-loop control policies are learned through jet actuation based on near-wake information. The proposed framework achieves robust convergence and significantly reduces both mean drag and fluctuating lift. Free-surface confinement accelerates the upper-side flow and generates a dominant asymmetric vortex, thereby intensifying wake asymmetry. The control jet forms a virtual barrier that suppresses the large-scale upper vortex, restores a bi-vortex wake structure, delays vortex formation, and mitigates flow separation. POD analysis further shows that control redistributes energy among more modes and weakens dominant coherent structures. In particular, the second POD mode exhibits more balanced upper and lower vortex strengths because the jet prevents the upper vortex from compressing the lower one. This decoupling weakens shear-layer interactions, stabilizes the wake, and explains the observed force reduction.