Reinforcement learning of control strategies for reducing skin friction drag in a fully developed turbulent channel flow

Reinforcement learning is applied to the development of control strategies in order reduce skin friction drag a fully developed turbulent channel flow at low Reynolds number. Motivated by so-called opposition (Choi et al. , J. Fluid Mech. vol. 253, 1993, pp. 509–543), which input so as cancel wall-normal velocity fluctuation on detection plane certain distance from wall, we consider wall blowing and suction input, its spatial distribution determined instantaneous streamwise fluctuations 15 units above wall. A deep neural network used express nonlinear relationship between sensing information it trained maximize expected long-term reward, i.e. reduction. When only measured linear used, present framework reproduces successfully optimal weight for reported previous study (Chung & Talha, Phys. Fluids 23, 2011, 025102). In contrast, when more complex based are obtained. Specifically, obtained switch abruptly strong downwelling high-speed fluid towards upwelling low-speed away respectively. Extracting key features policies allows us develop novel leading reduction rates high 37 %, higher than 23 % achieved conventional same Finding such an effective policy quite difficult relying solely human insights. The results indicate that reinforcement can be through systematic large number trials.