Published June 30, 2026 | Version 1.0.1
Software Open

HydroGym: A Multi-Solver Reinforcement Learning Platform for Fluid Dynamics

Description

HydroGym is a multi-solver reinforcement learning platform for data-driven modeling and control of fluid dynamics. This release (v1.0.1) extends HydroGym from a single-solver backend to a native multi-solver reinforcement learning framework for flow control, with environments spanning six CFD solver backends: Firedrake (finite element), MAIA LBM (lattice Boltzmann), MAIA structured FV (finite volume), NEK5000 (spectral element), JAX, and JAX-Fluids (differentiable). It provides 61 pre-configured, Gymnasium-compatible environments ranging from canonical 2D benchmarks to 3D turbulent flows at Reynolds numbers up to 400,000, compatible with standard RL libraries (Stable-Baselines3, RLlib, etc.).

This release introduces a data manager abstraction that handles environment data natively and uniformly across solver backends. Accompanying environment checkpoints are hosted in the companion HuggingFace dataset repository (https://huggingface.co/datasets/dynamicslab/HydroGym-environments) and are auto-accessed at runtime. Documentation is available at https://dynamicslab.github.io/hydrogym.

This archive corresponds to the software accompanying the associated Nature publication.

Files

hydrogym-1.0.1.zip

Files (65.2 MB)

Name Size
md5:71a1388cdab3f8b7e1e2b9a7d40f7d63
65.2 MB Preview Download

Additional details

Additional titles

Alternative title
The HydroGym Reinforcement Learning Platform for Fluid Dynamics

Identifiers

URL
https://github.com/dynamicslab/hydrogym/tree/v1.0.1
URL
https://github.com/dynamicslab/hydrogym
arXiv
arXiv:2512.17534

Dates

Updated
2026-06-30

Software

Repository URL
https://github.com/dynamicslab/hydrogym
Programming language
Python
Development Status
Active