Applying deep reinforcement learning to active flow control in weakly turbulent conditions

Machine learning has recently become a promising technique in fluid mechanics, especially for active flow control (AFC) applications. A recent work [Rabault et al., J. Fluid Mech. 865, 281–302 (2019)] demonstrated the feasibility and effectiveness of deep reinforcement (DRL) performing AFC over circular cylinder at Re = 100, i.e., laminar regime. As follow-up study, we investigate same problem an intermediate Reynolds number, 1000, where weak turbulence poses great challenges to control. The results show that DRL agent can still find effective strategies, but requires much more episodes learning. remarkable drag reduction around 30% is achieved, which accompanied by elongation recirculation bubble turbulent fluctuations wake. Furthermore, also perform sensitivity analysis on learnt strategies explore optimal layout sensor network. To our best knowledge, this study first successful application weakly conditions. It therefore sets new milestone progressing toward strong flows.