Deep Reinforcement Learning for Active Drag Reduction in Wall Turbulence

Reinforcement Learning is a framework in which an agent learns to take decisions through a trial-and-error process. When the learning process is supported by deep neural networks, it is referred to as deep reinforcement learning (DRL).

In this work, a numerical simulation is designed as a first step to apply DRL control in wall-bounded turbulent flows. A two-dimensional channel with a recirculation bubble on the lower wall is considered as first test case. The goal of the control is to minimize the size of the recirculation region, also called reattachment length. Since the control introduces perturbations in the velocity field, the reattachment length is measured using moving average over time.

Existing state-of-the-art DRL algorithms are tested in this setting in order discover novel control strategies, beyond traditional opposition control. In particular, we use proximal policy optimization (PPO) to determine the time variation of the amplitude of a volume forcing positioned upstream of the bubble. The magnitude of the forcing is limited in order to force the agent to learn an energy-efficient policy for the control.

Once an effective control policy is obtained in 2D, we will consider the more challenging case of drag reduction in 3D fully-turbulent channel flow.