HydroGym Trains AI Controllers Across 60+ Fluid Environments

HydroGym brings reinforcement-learning control into a shared fluid-dynamics testbed

An international research team launched HydroGym in work highlighted by Michigan Engineering on August 26, 2026. The platform provides more than 60 simulated environments for training and comparing machine-learning controllers that actively modify fluid flows.

The goal is broader than one aerodynamic benchmark. HydroGym targets control problems including drag reduction, lift improvement, noise suppression and thermal management, with applications discussed for aircraft, wind turbines, jet engines and high-performance computing.

The collaboration includes researchers from the University of Washington, University of Michigan, RWTH Aachen University and Technical University of Munich, with additional contributors from institutions across Europe and Asia.

Physics-informed training can reduce trial-and-error

HydroGym combines reinforcement learning with physics-based simulation. The team reports that incorporating physics knowledge into agent training reduced the trial-and-error needed to optimize control strategies by as much as 65% in the studied settings.

The platform supports several fluid solvers, including lattice-Boltzmann, finite-volume, spectral-element and finite-element approaches. Solvers that support automatic differentiation can also be incorporated directly into optimization loops, allowing gradient-based and hybrid methods alongside standard reinforcement learning.

HydroGym can test centralized controllers as well as distributed systems in which multiple local controllers coordinate across a surface. That makes the platform relevant to multi-agent reinforcement learning where one centralized policy may not efficiently process all of the spatial information in a large flow field.

A simple training environment transferred to a harder wing simulation

In one demonstration, researchers trained a controller in a comparatively simple channel-flow environment and then applied it without additional training to a much more complex simulated airplane-wing scenario.

The team reports that the transferred controller reduced surface friction across the wing by 38% and overall drag by 11%. Training in the simpler environment was reported as 100 times faster and 10,000 times cheaper than training directly on the wing simulation.

These are research-team results from selected simulation environments, not guarantees for real aircraft or arbitrary flow-control problems. Hardware, solver fidelity, geometry, turbulence regime, sensing and actuation can all materially change performance.

The open platform is meant to make results easier to compare

A persistent problem in AI-driven flow control is that researchers often build custom environments and report results that are difficult to compare directly. HydroGym provides a shared set of environments and interfaces intended to make controller comparisons and transfer studies more systematic.

The code, documentation and environments are publicly available, and the researchers say the platform is designed to grow with community contributions.

HydroGym is significant because it gives reinforcement-learning researchers a common proving ground for studying whether control policies can transfer across geometries, flow regimes and simulation fidelities. The long-term opportunity is not a single universal fluid controller yet, but a reproducible path for testing how close AI systems can get to that goal.