Research on vorticity driven reward for active flow control over airfoil based on deep reinforcement learning

Abstract

An intelligent flow control on the flow separation over an airfoil under weak turbulent conditions is investigated and solved by deep reinforcement learning (DRL) method. Both single and synthetic jet control at the airfoil angles of attack of 10°, 13°, 15° are compared by training a neural network for closed-loop active flow control strategy based on the soft actor-critic (SAC) algorithm. The training results demonstrate the effectiveness of the deep reinforcement learning-based active flow control method in suppressing the flow separation at high angles of attack, validating its potential in complex flow environments. To improve the stability of the shedding vortex alley over airfoil, a novel reward function considering the vorticity statistics in terms of both vortex and asymmetric shear intensity is first proposed in this work. This vorticity driven reward is demonstrated to perform better in suppressing the rotation and shear intensity and the aerodynamic optimization than the traditional one. Moreover, it can accelerate the convergence speed during the exploration phase. Moreover, it can accelerate the convergence speed during the exploration phase. This study provides valuable insights for future applications of DRL in active flow control under more complex flow conditions.