Answers are grounded in this paper and cite their sources.
In the present work, deep reinforcement learning (DRL) is used to control the flow past an infinite-span SD7003 wing section at a Reynolds number of Re=60,000, and an angle of attack of AoA=14º, where massive turbulent separation is present. Large-eddy simulations are used to train a DRL controller capable of selecting the appropriate jet velocity at each instant. In addition to assessing the controller under fully turbulent conditions, two pressure-sensing strategies are investigated: a reference configuration employing distributed wake sensors and a practical configuration relying exclusively on wall-mounted sensors. Both strategies successfully suppress flow separation, significantly reducing wake turbulence and improving the aerodynamic performance of the wing. Remarkably, the wall-sensor configuration achieves slightly higher aerodynamic performance, providing approximately 58% drag reduction and 75% lift enhancement while relying on substantially fewer and more localized measurements. Flow-field and Reynolds-stress analyses reveal that the wall-sensor controller generates larger jet velocity fluctuations, enhancing near-wall momentum transfer, reducing the extent of the separation bubbles, and producing a narrower wake. These results demonstrate that DRL-based active flow control remains effective in fully turbulent three-dimensional flows while showing that sparse wall-mounted sensing is sufficient to achieve high-performance closed-loop control, representing an important step toward practical aerodynamic applications.
No takes yet. Share an insight, caveat, or question.
Montalà et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: