An AI Learned to Cut Air Drag in a Wind Tunnel Without Any Model of the Airflow

A learning algorithm has cut the aerodynamic drag on a car-shaped model inside a real wind tunnel, working only from sensor readings and with no model of the turbulent air around it. Junjie Zhang and Georgios Rigas of the Department of Aeronautics at Imperial College London, with colleagues there and at Peking University, published the result Sept. 7 in Communications Engineering.

That it ran in hardware is the point. Most machine-learning results in fluid control come out of computer simulations; this one learned in moving air, from measurements taken as it acted. Drag is fuel, and on a blunt shape most of it comes from the churning wake behind the body, the part of the physics engineers model worst.

The system, which the team calls REACT, ran on an Ahmed body, a simplified car shape used as a standard test object in aerodynamics labs. It carried only onboard sensors and small servo-driven surfaces. From sparse tunnel measurements, it converged on a policy that reduced drag and still returned a net energy saving once the power its surfaces drew was counted, the paper reports. Given no prior knowledge of the flow physics, the agent found for itself that damping the organized, repeating swirls in the wake is what saves the most energy.

Set against model-based controllers the team ran as baselines on the same rig, REACT reached what the paper describes as "two to four times greater performance." A single policy trained offline stayed effective without retraining across Reynolds numbers of 86,400 to 518,400. Reynolds number measures how turbulent a flow is, and that range spans a sixfold spread of conditions.

The authors describe the work as a demonstration of autonomous closed-loop reinforcement-learning control in a high-Reynolds-number wind-tunnel environment, and say it suggests "a path toward data-driven, state-dependent control of turbulent flows."