Abstract


In this talk, some recent applications of deep reinforcement learning (DRL) in active flow control (AFC) on a toy model will be introduced. Here the term AFC means that the control is realized by injecting a small amount of energy into existing flow systems. Compared to its counterpart, i.e., passive flow control, AFC is adaptive and on-demand, and hence has a much wider operating range. First, the use of DRL for AFC is demonstrated through an interesting hydrodynamic problem - hydrodynamic stealth, where the actuation is realized using a group of windward-suction-leeward-blowing (WSLB) actuators. Second, what the DRL agent “thinks” during the learning is interpreted through a vortex-induced vibration control problem. Last, the control is pushed towards the turbulent flow regime. Through these DRL-guided AFC studies, some new and unexpected control strategies have been revealed.