Deep reinforcement learning for active flow control in a turbulent separation bubble

Abstract

The control efficacy of deep reinforcement learning (DRL) compared with classical periodic forcing is assessed for a turbulent separation bubble (TSB) reaching a friction Reynolds number of Re𝝉=750. The TSB is a simplified representation of the separation phenomenon naturally arising in wings, and a successful reduction of the TSB has practical implications in the reduction of the aviation carbon footprint. We use two different grid resolutions so that the DRL training is run on the coarse grid for computational simplicity. Since this coarse grid captures the most important features of the flow, the obtained strategy can be directly applied on the fine grid, reaching a very good performance. While the classical zero-net-mass-flux (ZNMF) periodic control is able to reduce the TSB length by a 6.8%, the DRL-based control achieves 8.9% reduction. Furthermore, DRL control provides a smoother control strategy while also being ZNMF. To the best of our knowledge, the current test case is the highest Reynolds-number flow that has been successfully controlled using DRL to this date. Last, we provide details of our open-source computational framework suited for the next generation of exascale computing machines.