HydroGym Platform Trains AI to Control Fluid Dynamics, Reducing Drag by 38%

Fluid dynamics has always been a problem of overwhelming complexity. The equations that govern how air moves over a wing, how water flows through a pipe, how heat dissipates from a computer chip—these involve so many variables, changing so rapidly across space and time, that engineers have traditionally relied on expensive wind tunnels, trial-and-error testing, and incremental improvements. An international team of researchers has now built a different kind of laboratory: HydroGym, a machine learning platform designed to train artificial intelligence systems to actively control fluid flows, and in doing so, to solve problems that have resisted conventional approaches.

The platform emerged from a collaboration between researchers at the University of Washington, University of Michigan Engineering, RWTH Aachen University, and the Technical University of Munich, with funding from the National Science Foundation and Boeing. The core insight is straightforward but powerful: instead of trying to predict fluid behavior directly through calculation, use reinforcement learning—the same machine learning technique that has recently transformed protein folding and nuclear fusion research—to train AI agents that learn to manipulate flows by interacting with simulated environments. By embedding physics knowledge into the training process, the team reduced the computational trial-and-error typically required by as much as 65 percent.

The practical demonstration was elegant. Researchers created a simple scenario: a channel formed by two flat surfaces, perforated with holes like an air hockey table. They trained a machine learning model to control the air entering and exiting those holes, learning to disrupt the turbulent flows that create friction while maintaining balance between incoming and outgoing air. The model learned this task in a relatively inexpensive simulation. Then came the test: they applied that same controller to a far more complex scenario—a section of a simulated airplane wing. The results were striking. Surface friction across the wing dropped by 38 percent. Overall drag fell by 11 percent. And the training process had been 100 times faster and 10,000 times cheaper than training directly on the wing itself.

This transfer of learning from simple to complex geometries points toward something larger. Steven Brunton, the Boeing Professor in AI and Data-Driven Engineering at the University of Washington and a senior author of the study published in Nature, framed the stakes: fluid flows are central to industries worth several trillion dollars annually—energy, transportation, health, defense. An improved ability to control these flows could reshape efficiency across aviation, wind power generation, jet engine acoustics, and industrial cooling. The platform's potential lies not just in solving individual problems but in discovering general principles that apply across different scenarios.

HydroGym itself is built for flexibility. It can train centralized control systems—a single AI brain managing an entire surface—or distributed systems in which multiple smaller controllers manage different regions while coordinating with neighbors. This matters because on large surfaces, a single controller would be overwhelmed by information. The platform generates simulated datasets on the fly rather than relying on historical data, allowing users to choose from multiple physics modeling strategies: lattice Boltzmann, finite-volume, spectral-element, and finite-element methods. Some of these solvers support automatic differentiation, meaning they can be embedded directly into the training loop, opening pathways to gradient-based optimization alongside standard reinforcement learning.

The team prepared more than 60 testing environments—varying control strategies, surface shapes, and flow types—where users can train models and benchmark them against one another. The code, documentation, and full set of environments are freely available on GitHub. Christian Lagemann, the first author and former postdoctoral researcher at UW, described the shift this represents: instead of isolated demonstrations with no common framework for comparison, researchers can now study how control strategies transfer across different geometries and flow types, or train in inexpensive surrogate environments and test in realistic scenarios. The vision is toward a single model of fluid dynamics—one that captures enough physics to apply to different levels of turbulence, any surface shape, and both liquid and gas flows or mixtures of the two.

What makes HydroGym significant is not that it solves any single problem definitively, but that it creates a shared proving ground. The international research community can now collaborate systematically rather than working in isolation. Each improvement, each new control strategy, each successful transfer from simple to complex geometry adds to a collective understanding. The platform is actively growing with contributions from researchers at institutions across South Korea, Sweden, France, the United Kingdom, and Germany. For industries built on the physics of flowing fluids—and that is most of them—this represents a shift from individual breakthroughs to systematic discovery.