Reinforcement learning for bluff body active flow control in experiments and simulations

TLDR

Reinforcement learning has proven effective in games and robotics, underscoring its potential for complex control tasks. The study applies reinforcement learning to reduce drag on circular cylinders in turbulent flow. High‑fidelity simulations were employed to investigate the physical mechanisms behind the RL‑derived drag‑reduction strategies. RL achieved a drag reduction of over 30 % quickly and facilitates efficient exploration of new flow‑control strategies, potentially accelerating discovery and design in fluid‑mechanics engineering.

Abstract

Significance Reinforcement learning (RL) has been applied effectively in games and robotic manipulation. We demonstrate the effectiveness of RL in experimental fluid mechanics by applying it to reduce the drag of circular cylinders in turbulent flow, a canonical fluid–structure interaction problem. Although physics agnostic, RL managed to reduce the drag by <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mn>30</mml:mn> <mml:mi>%</mml:mi> </mml:math> or reach another specified optimum point very quickly. Following this discovery, we used high-fidelity simulations to probe the underlying physical mechanisms so that the discovered control techniques can be generalized to other similar flow problems. More broadly, RL-guided active control can lead to efficient exploration of additional flow-control strategies in experimental fluid mechanics, potentially paving the way for accelerating scientific discovery and different designs in flow-related engineering problems.