HydroGym AI cuts drag 11% on a simulated wing

Inside a computer model of an airplane wing, a controller trained on a much simpler flow cut surface friction by 38% and overall drag by 11%. The result comes from HydroGym, a new platform built by researchers at the University of Washington, the University of Michigan, RWTH Aachen University and the Technical University of Munich to train and compare artificial-intelligence systems that control fluids.

HydroGym uses reinforcement learning, a method in which an AI agent improves through interaction with its environment. The researchers add knowledge of physics to reduce the blind trial and error that normally makes these problems expensive. In their tests, that approach cut the trial and error needed to optimize control strategies by as much as 65%.

The wing result began with a simpler experiment: a channel made from two flat surfaces punctured with holes, like air-hockey tables. The controller learned to manage air entering and leaving those holes, disrupting turbulent flow while keeping the incoming and outgoing air in balance. When transferred to the more complex simulated wing, the model needed no additional training. Training in the simple setting was 100 times faster and 10,000 times cheaper than training directly on the wing.

The platform is designed as a common proving ground rather than a single controller. It contains more than 60 environments covering different surfaces, flows and control methods, including shape-changing surfaces, tiny flaps, spinning elements and jets. It can also test distributed systems, in which smaller controllers coordinate over different parts of a large surface, and it generates simulated data while offering several physics-modeling approaches.

So what changes in practice? Researchers can train on inexpensive surrogate problems, compare methods on the same ground and then test whether the learned control transfers to more realistic geometries. That could support work on aircraft, wind turbines, jet-engine noise and supercomputer cooling. The limit is clear: these results are simulations, and the figures reported by the research team still need confirmation in real-world conditions.