Deep reinforcement learning-guided active control of turbulent flows
Phys. Rev. Fluids 11, 043903 – Published 8 April, 2026
DOI: https://doi.org/10.1103/y7cc-2xh7
Abstract
In this study, we present a deep reinforcement learning (DRL) framework for closed-loop control of canonical flow past a circular cylinder under fully developed turbulent conditions. To enable computationally efficient training for each episode, we employ a graphics processing unit–accelerated flow solver based on the generalized interpolation-supplemented cascaded lattice Boltzmann method. Moreover, we further employ a two-stage exploration process, beginning with a coarse mesh and minimal spanwise flow resolution before progressing to a finer mesh with well-resolved spanwise flow. This approach significantly reduces computational time, making the implementation of intricate and time-consuming DRL-guided active flow control (AFC) feasible in turbulent flows. The resulting DRL-guided AFC achieves an optimal drag reduction of and a lift fluctuation reduction of due to recovered pressure at the rear side, attenuated wake vortical structures, reduced recirculation bubble length, and mitigated downstream turbulent fluctuations. Additional numerical experiments applying second-order smoothing to the DRL-determined control actions confirm that effective turbulent flow control does not necessarily require high-frequency actuation. This work verifies DRL as a powerful machine-learning tool for tackling challenges of AFC in turbulent flows. The revealed phenomena also highlight rich opportunities for further explorations of the underlying physics of flow under control.
