An Artificial Neural Network trained through Deep Reinforcement Learning achieves Active Flow Control of the 2D Karman Vortex Street behind a Cylinder at moderate Reynolds number

The Karman vortex street has attracted much attention for over a century. However, it still offers a topic of investigation for flow stability and control. These are critical for industrial applications and outstanding scientific questions. Active flow control remains mostly inaccessible due to the combination of non-linearity, high dimensionality, and time dependence implied by the Navier Stokes equations. Here we report that Artificial Neural Networks trained through Deep Reinforcement Learning can achieve active flow control of the Karman vortex street behind a cylinder in 2D simulations at Re = 100. In particular, the Neural Network manages to reduce drag by around 8% through increasing the size of the recirculation area by 125% while decreasing the pressure drop. This active flow control is achieved through synthetic jets normal to the surface of the cylinder blowing perpendicular to the flow. The jets have very low mass flow rates, in average no more that 0.5% of the incoming mass flow rate intersecting the cylinder diameter. These results show that Deep Reinforcement Learning is a promising tool for attacking the still largely unsolved problem of optimal flow control.