A control policy learned by deep reinforcement learning can set wall blowing and suction from local wall measurements and cut turbulent skin friction by roughly a third, but the benefit shrinks as the Reynolds number rises. The learned strategy is not simple opposition control: it builds a virtual wall further from the surface and suppresses the pressure-strain redistribution that feeds the wall-normal velocity fluctuations, so the Reynolds shear stress in the buffer layer falls and, through the FIK identity, so does the drag. The degradation at higher Reynolds number is traced to large-scale outer motions, which the wall sensors see only as a superposed footprint while the small-scale near-wall turbulence they actually modulate remains beyond the controller's reach. This is the case study for why a control law learned at one Reynolds number does not simply transfer to another, and for how budget equations and scale decomposition turn a black-box policy into a physical explanation.
Deep reinforcement learning (DRL) is employed to develop control strategies for drag reduction in direct numerical simulations (DNS) of turbulent channel flows at high Reynolds numbers. The DRL agent…