An open reinforcement learning platform for fluid dynamics wants to do for flow control what shared benchmarks did for robotics.
HydroGym, described in a new arXiv paper, provides more than 61 validated simulation environments covering flows from laminar to turbulent, with Reynolds numbers up to 400,000 and both 2D and 3D configurations. It supports five solver backends — finite-volume, spectral-element, finite-element, lattice-Boltzmann, and fully differentiable — so researchers are not locked into a single simulation stack. Across those environments, RL agents found control strategies yielding drag reductions exceeding 90% in standard test cases.
The more striking result is zero-shot transfer: an agent trained only on simplified channel flow cut friction drag by 38% on an unseen 3D wing section at a chord Reynolds number of 200,000, without any additional tuning. The authors say that reduction in exploration cost is roughly four orders of magnitude compared with optimizing directly on the wing — which, if it holds up under scrutiny, would meaningfully shrink the compute bill for aerodynamic control research.
Fluid dynamics has historically resisted the benchmark-driven compounding that accelerated progress in robotics and protein structure prediction, because researchers typically tuned each controller to a specific geometry and operating condition. HydroGym is an explicit attempt to break that pattern — though whether the community actually converges on it, or fragments across competing platforms as has happened in other ML subfields, remains to be seen.