Fluid dynamics has a new training ground for artificial intelligence, and its name is HydroGym. An international research team has launched a platform designed to help machine-learning systems control turbulent flows around wings, turbines, engines and other engineered surfaces. The system allows researchers to train and compare artificial-intelligence controllers under standardized conditions, with the goal of reducing drag, increasing lift, suppressing noise and improving heat transfer. The project brings together scientists from the University of Washington, the University of Michigan, RWTH Aachen University and the Technical University of Munich, along with collaborators from institutions across Europe, Asia and the United Kingdom. Their study, published in Nature, presents HydroGym as a potential foundation for developing broadly transferable models of fluid behavior rather than isolated solutions for individual engineering problems.
Fluid flows are central to transportation, energy production, health technologies and defense, but predicting them remains one of the most difficult challenges in applied physics. Air moving over an aircraft wing, water passing through a turbine or coolant circulating around a computer processor is governed by the Navier–Stokes equations, a set of nonlinear equations that describe the conservation of mass, momentum and energy. In realistic environments, these equations interact with turbulence, surface geometry, pressure gradients, temperature differences and changing boundary conditions. The number of variables becomes so large that even powerful computers may require enormous amounts of time and energy to calculate the flow directly. HydroGym addresses this difficulty by combining high-fidelity numerical simulations with reinforcement learning, allowing artificial agents to improve their decisions by interacting with simulated fluid environments.
In reinforcement learning, an agent repeatedly observes a system, takes an action and receives a numerical reward based on the consequences. For fluid control, the actions might include changing the angle of a small flap, morphing part of a surface, rotating a tiny actuator or adjusting jets that inject and remove fluid. The reward can be designed to favor lower drag, higher lift, reduced turbulence or more efficient heat removal, while penalties prevent the controller from using excessive energy or violating physical constraints. HydroGym adds knowledge of fluid mechanics to this process, reducing the wasteful trial and error that can make conventional reinforcement learning impractical. According to the researchers, incorporating physics into training reduced the amount of experimentation required to optimize control strategies by as much as 65 percent.
The platform is particularly important because flow-control research has often been divided into highly specialized demonstrations. A controller developed for one wing, pipe or cavity may perform well in that setting but fail when the geometry, Reynolds number or type of fluid changes. HydroGym creates a common framework in which algorithms can be tested on the same problems and judged using comparable performance measures. Its more than 60 environments represent different surfaces, flow conditions and control mechanisms. Researchers can therefore examine whether an algorithm learns a general physical principle or merely memorizes the details of one simulation. They can also train a controller in a relatively inexpensive environment and then test it on a more complex and computationally demanding model.
The team demonstrated this transfer process using turbulent flow between two flat surfaces perforated with holes, similar in appearance to the surfaces of an air-hockey table. By controlling air entering and leaving through the holes, an artificial agent learned to disrupt the organized structures that contribute to turbulent skin friction. The controller had to reduce friction while maintaining a balance between incoming and outgoing flow, a requirement that prevented it from achieving an apparent improvement simply by adding or removing mass from the system. Although the training problem was comparatively simple, the resulting strategy was then transferred to a simulated section of an airplane wing, where the geometry and flow behavior were considerably more complex.
On the wing, the transferred controller reduced surface friction by 38 percent and total aerodynamic drag by 11 percent. The result is notable because the model was applied without additional training in the new geometry, a capability the researchers describe as zero-shot transfer. The computational advantage was equally striking: training in the simpler environment was approximately 100 times faster and 10,000 times less expensive than training directly on the simulated wing. The finding suggests that a controller can learn features of turbulence that are not tied to a particular shape. Instead of observing every possible aircraft configuration during training, an artificial agent may be able to extract the underlying relationships between actuation, vortices, pressure and friction, then use those relationships in a new setting.
That possibility could transform the economics of aerodynamic design. Drag reduction is valuable because aircraft and wind turbines must overcome aerodynamic resistance throughout operation, and even modest improvements can translate into lower fuel consumption, longer range or greater energy output. Active flow control could also help aircraft maintain lift during difficult maneuvers, reduce noise generated by turbulent wakes or improve the performance of compact propulsion systems. In other fields, strategically placed jets or deformable surfaces could manage heat around data centers, improve combustion efficiency in engines or control the movement of fluids in industrial equipment. The same principles may eventually apply to liquids, gases and mixtures in which multiple phases interact, although the complexity of those systems remains a major challenge.
HydroGym is not limited to a single centralized artificial “brain” that receives information from an entire surface and issues commands to every actuator. The platform also supports distributed and multi-agent control, in which smaller controllers manage separate regions while communicating with neighboring agents. This approach may be essential for large wings, industrial pipelines or turbine blades, where the amount of information required by a centralized system would be too great to process efficiently. Local agents can respond to conditions in their own region while coordinating with the broader flow. Because fluid behavior follows the same physical laws across a surface, a collection of controllers may learn reusable local strategies even when the exact timing and position of vortices vary.
Another feature of HydroGym is that its training data can be generated during the simulation rather than taken from a fixed historical dataset. Users can select among several numerical approaches, including lattice Boltzmann, finite-volume, spectral-element and finite-element methods. These techniques represent fluid motion in different ways and offer different compromises between speed, accuracy and computational cost. Several of the supported solvers, including JAX-Fluids, provide automatic differentiation, which allows the simulator to calculate how a small change in a control input affects a selected objective. This makes it possible to combine reinforcement learning with gradient-based optimization and hybrid methods. Researchers can thus compare not only different agents but also fundamentally different ways of searching for effective control policies.
The creators of HydroGym hope the open platform will move the field away from isolated demonstrations and toward a shared science of controllable flows. Its code, documentation and environments are freely available, enabling researchers to reproduce published results, add new geometries and test competing algorithms. The project was primarily supported by the U.S. National Science Foundation and Boeing, with additional support from the University of Michigan, the U.S. Army Research Office, the German Research Foundation and the European Research Council. The researchers emphasize that a universal model capable of controlling every turbulent flow does not yet exist, and transferring strategies between radically different physical regimes will require further validation. Even so, the wing demonstration offers a compelling glimpse of an AI system learning in a simple world and then acting successfully in a far more realistic one—a step that could make intelligent control of the invisible forces around us faster, cheaper and widely accessible.
Subject of Research: Artificial intelligence, reinforcement learning and active control of fluid flows
Article Title: The HydroGym reinforcement learning platform for fluid dynamics
Web References: https://dynamicslab.github.io/hydrogym/docs/introduction; https://www.youtube.com/watch?v=SQrPBk6f0GY
References: Nature study, DOI: 10.1038/s41586-026-10917-6
Keywords
Fluid dynamics, reinforcement learning, artificial intelligence, turbulence, aerodynamic drag, flow control, aircraft wings, computational fluid dynamics, multi-agent systems, HydroGym
