HydroGym: A Reinforcement Learning Benchmark Platform for Flow Control
NewHydroGym is an open-source platform that turns active flow control into a standardized benchmark. Its central claim is that fluid control has not been visibly advanced by reinforcement learning, and that the missing ingredient is not an algorithm but shared benchmark environments and a uniform interface. This framing dictates the platform's shape: it provides 61 validated flow-control environments, wraps finite element, lattice-Boltzmann, finite volume, spectral element and fully differentiable solvers behind one Gymnasium-compatible layer, and lets the same reinforcement learning code move across geometries, Reynolds numbers and dimensionalities so that results can be compared rather than merely asserted.
The paper, "The HydroGym reinforcement learning platform for fluid dynamics", appeared in Nature on 19 August 2026 (volume 657, pages 369-376, DOI 10.1038/s41586-026-10917-6); the preprint is arXiv:2512.17534 (v1 submitted 19 December 2025, v2 revised 30 June 2026). The repository dynamicslab/hydrogym was created on 29 December 2021 under the MIT license, with the copyright line reading Copyright (c) 2021 Jared Callaham. This entry records 2026 as the year of the platform paper; the repository creation year of 2021 corresponds to the copyright start year in the license and serves a different purpose.
1. Overview
The value of this entry lies less in any single drag-reduction number than in a methodologically transferable skeleton: when someone claims a 30% drag reduction from a reinforcement learning controller, can anyone else reproduce that number under the same case and the same baseline? The paper starts from the observation that evaluation practice in fluid control has long been fragmented, with each controller tuned to one geometry and one operating condition, so results cannot be accumulated, transferred or compared. It draws a direct analogy with computer vision and robotics, which advanced through shared benchmarks such as ImageNet and MuJoCo.
A second reason is the platform's stated limits. Platform papers commonly hide their constraints, leaving readers unable to judge where results apply. This one puts the limits in the main text: all current environments use direct numerical simulation with fully resolved turbulent structures and no closure relations; the platform currently reaches Re_tau = 2,200 for internal flows and Re_c = 400,000 for external aerodynamics, below the 10^7 to 10^8 range common in industrial applications; and the authors state explicitly that this ceiling reflects training-phase computational constraints rather than solver capability, since production solvers routinely handle order-10^9 degree-of-freedom simulations on full aircraft geometries, yet embedding such simulations in a training loop would demand more than 500 GPUs and 50 TB of memory per run.
A third reason is the methodological demonstration in the zero-shot transfer experiment. Rather than grinding away on a high-Reynolds-number 3D wing, the authors train in a cheap channel-flow surrogate and deploy the policy directly onto a wing with different geometry, pressure gradient and Reynolds number. That train-on-surrogate, deploy-directly pattern is a useful structural reference for any field where a single high-fidelity simulation is too expensive.
2. How It Works
The software stack has three layers. The bottom layer holds six CFD solver backends: m-AIA LB (lattice-Boltzmann with D3Q27 lattices and BGK or cumulant collision operators for large-scale weakly compressible DNS, using local grid refinement and MPI plus OpenMP plus GPU parallelism); m-AIA FV (finite volume, compressible Navier-Stokes, arbitrary Lagrangian-Eulerian formulation, structured body-fitted curvilinear grids, DNS and LES support, a five-stage low-storage Runge-Kutta scheme, and synthetic jet plus transverse travelling-wave boundary conditions); Nek5000 (spectral element, high-Reynolds-number incompressible flow, PNPN-2 discretization, dynamically coupled to reinforcement learning under MPI); a JAX-based differentiable incompressible solver supporting jax.grad, vmap and jit; JAX-Fluids (fully differentiable high-order CFD for compressible two-phase multiphysics, with WENO5-Z and TENO6-A shock capturing, level-set and diffuse-interface multiphase models, immersed boundary method, and autodiff across multi-node GPU and HPC); and Firedrake (finite element, aimed at code transparency and fast prototyping, using Taylor-Hood elements and PETSc-based distributed parallelism).
The middle layer is the unified environment abstraction and actuation interface. Control is applied by imposing time-varying Dirichlet boundary conditions on the surface of the immersed body, specifying a non-zero velocity vector in place of the standard no-slip condition and enforcing it through the interpolated bounce-back scheme. The thermal benchmark cases implement control by varying inflow temperature. Two actuator types are used, synthetic jet and surface rotation, both with a temporal ramping mechanism so that control actions transition smoothly, which the paper identifies as critical for numerical stability.
The top layer handles reinforcement learning and evaluation. The interface is Gymnasium-compatible, with official examples targeting Stable-Baselines3 and RLlib; the paper also names TorchRL, CleanRL, PureJaxRL and Gymnax. Multi-agent reinforcement learning runs through the PettingZoo interface for distributed actuation, and the Firedrake backend adds global stability analysis via SLEPc and modal decompositions via modred. Baseline algorithms are PPO, DDPG and TD3, with gradient-enhanced policy optimization (GPPO) available for differentiable environments.
3. Scope and Assumptions
The numerical premise is strict: every current environment is a DNS with no turbulence model. That buys trusted physical ground truth free of closure assumptions, at the price of cost that grows with grid points and time steps at high Reynolds numbers. The paper quantifies it: individual flow fields up to 75 GB, full test trajectories up to 15 TB.
The Reynolds-number ceiling is Re_tau = 2,200 for internal flows and Re_c = 400,000 for external aerodynamics, adequate for fundamental research and algorithm development but below typical industrial regimes. The stated cause is training-phase compute, not solver capability.
Flow-regime coverage leans toward canonical flows and near-equilibrium turbulent boundary layers. Non-equilibrium flows with strong pressure gradients, separation, reattachment and transient dynamics are not yet included; the authors describe environments under development that introduce adverse and favourable pressure gradients through body forces.
The multiphysics gap is explicit. Future releases are planned to cover aero-acoustics such as jet noise mitigation and trailing-edge noise suppression, reactive flows in combustion systems, and multiphase dynamics. None of this is in the current version.
Transfer generalization has a bounded scope. The zero-shot experiment exploits shared near-wall physics, and the paper states plainly that how far this transfers across flow families lacking those shared features remains open, naming strongly pressure-gradient-driven, separation-dominated and compressible flows.
Experimental validation is absent. Every result is numerical. The paper's framing of future work is that HydroGym agents will eventually be deployed in experimental flow control scenarios, a plan rather than a completed result.
Actuation is idealized. Control acts through numerical boundary conditions; synthetic jets and surface rotation omit the engineering constraints of power draw, actuator efficiency, compressed air supply and installation.
4. Getting Started
The first route is the official Docker examples. Prebuilt containers are recommended to avoid dependency hell. The NVIDIA Hopper and Blackwell image is clagemann/hydrogym-nvhpc-26.1_cuda-12.9_hopper_blackwell:latest, with additional Turing and Ampere variants, and AMD users can pull clagemann/hydrogym-rocm-6.3.3:latest.
The second route is the Firedrake backend. Enter examples/firedrake/getting_started, run ./run_example_docker.sh to test the environment interactively, then ./run_example_docker.sh train to start Stable-Baselines3 training. A typical configuration looks like this:
from hydrogym import FlowEnv
import hydrogym.firedrake as hgym
from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import DummyVecEnv, VecNormalize
def make_env():
env_config = {"flow": hgym.Cylinder, "flow_config": {"mesh": "medium", "Re": 100}, "solver": hgym.SemiImplicitBDF, "solver_config": {"dt": 1e-2}, "actuation_config": {"num_substeps": 2}}
return FlowEnv(env_config)
env = DummyVecEnv([make_env])
env = VecNormalize(env, norm_obs=True, norm_reward=True)
model = PPO("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=100000)
The third route covers the MAIA backend and multiphysics coupling. In examples/maia/getting_started, run ./run_example_docker.sh to prepare the workspace and download cases from the Hugging Face Hub, then execute under MPMD (one Python process plus one MAIA process on the GPU); training uses ./run_example_docker.sh train.
The fourth route is Nek5000 and multi-agent control. Under examples/nek/getting_started there are 1_nekenv_single for a single-agent environment, 3_pettingzoo for PettingZoo with SB3 training, and 6_zeroshot_wing_demo for the zero-shot wing transfer demo, run via ./run_pettingzoo_docker.sh.
The fifth route covers JAX differentiable environments. Under examples/jax/getting_started: 1_kolmogorov tests the Kolmogorov flow, 2_channel tests channel flow with a strong_actuation option, and 3_ppo trains with a command such as ./run_ppo_docker.sh --env channel --num-envs 1 --num-steps 10 --num-minibatches 5.
For data access, all required environment checkpoints download on first creation from the Hugging Face dynamicslab/HydroGym-environments repository and require connectivity. On HPC compute nodes without external network access, pre-download the environment files as described in examples/maia/README.md.
One practical selection tip: the suite covers both 2D and 3D, and the same flow usually has an inexpensive 2D version alongside its 3D high-fidelity counterpart. Iterating in 2D and switching to 3D once the policy is confirmed is the workflow the suite layout implies.
5. Comparison with Alternatives
Against bespoke control studies, the paper notes that most fluid reinforcement learning work has concentrated on computationally inexpensive 2D configurations at moderate Reynolds numbers, with extension to highly turbulent flows, complex 3D geometries and multiphysics representing the current frontier. A team building a solver, training and evaluation loop from scratch cannot compare its results with anyone else's; the platform moves that cost into shared infrastructure.
Against mainstream reinforcement learning benchmarks such as MuJoCo or Isaac Gym, the form is shared but the physics cost structure is not. Robot interaction costs milliseconds per step; fluid control interaction costs hours of CFD. The paper states plainly that training an effective agent typically requires thousands of environment interactions while each evaluation demands expensive CFD on high-performance computing clusters, making sample efficiency a feasibility precondition rather than a tuning metric.
Against industrial CFD solvers, the two serve different purposes. Industrial solvers target one-shot high-accuracy simulation of a given case; HydroGym targets the controllable, repeatable, parallelizable and differentiable interface a training loop needs. The m-AIA FV backend itself supports DNS and LES, structured body-fitted grids and compressible flow, so raw capability is not the differentiator; design intent is.
Against surrogate models such as neural operators, both routes attack the same problem of making expensive simulation affordable but intervene at different points. Neural operators learn a surrogate outside the solver; the HydroGym transfer route keeps the solver and swaps only the training configuration. The first fails when the surrogate extrapolates; the second fails when control priors do not carry across flow families. Their convergence is visible in the differentiable environments, where gradient-enhanced policy optimization becomes available.
Against classical control baselines, the paper names opposition control and uniform blowing as gold standards, and the zero-shot transfer policy outperforms both. In the platform's framing, comparison with classical linear control is the default baseline.
6. Benchmarks and Evidence
The platform ships systematic benchmark results over four benchmark configurations, using PPO, DDPG and TD3, and the training of all validated baselines consumed more than 150,000 GPU hours. Two-dimensional results appear in Figure 4 and three-dimensional results in Figure 3.
Specific numbers follow. The 2D fluidic pinball at Re = 100 achieves roughly 90% drag reduction by counter-rotating the downstream cylinder, which is the origin of the abstract's "exceeding 90%" figure and corresponds to a particularly favourable configuration. The 3D fluidic pinball at Re = 150 converts irregular asymmetric shedding into a steady symmetric wake and improves velocity deficit and pressure recovery. The 2D cavity at Re = 4,200 and the 3D cavity at Re = 7,500 test shear-layer instability control. Multi-agent control on the 3D cylinder at Re = 3,900 yields about 8% drag reduction, which the paper describes as comparable to the cited work; the multi-agent framework exploits the spatial locality of flow control by decomposing the global problem into local pseudo-environments, each managing a pair of ZNMF jets distributed spanwise, and differs from the cited work through shared gradient updates and a unified replay buffer. On the NACA0012 transverse gust at Re = 1,000, load oscillations drop by roughly 20%.
Data availability is explicit. Checkpoints and trajectories are hosted in the Hugging Face dynamicslab/HydroGym-environments dataset repository. Direct inspection shows 1513 files across 90 environment-named top-level directories, roughly 5400 downloads, and a most recent update of 23 October 2026. Directory names encode configuration and operating condition directly, such as NACA0012_3D_Re200000_AOA0, Cylinder_3D_Re3900, DRA2303Jet_3D_Re400000_Ma07 and TCFlarge_3D_Re206, with grid resolution variants marked by _fine_FD or _medium_FD suffixes.
Data volume carries an explicit caveat: the largest direct numerical simulation datasets cannot be deposited in full, since individual flow fields reach 75 GB and full test trajectories reach 15 TB. Representative compressed subsets sufficient to reproduce the reported results are provided, and complete datasets are available from the corresponding authors on reasonable request.
One counting nuance deserves attention. The paper body uses 61 and 61+ while the abstract says more than 60 and the GitHub README says 61+, whereas the Hugging Face repository shows 90 directories. The gap comes from counting convention: the paper counts validated flow scenarios, while the repository counts directories that include grid-resolution variants of the same scenario, such as Pinball_2D_Re100 and Pinball_2D_Re100_fine_FD. Any quoted number should state which level it refers to.
7. Role in Physics-AI Workflows
As shared ground for reinforcement learning control research, the platform unifies six backends, more than two hundred solver configurations, several actuator and sensor families under one interface, sparing each group the work of building a trainable CFD environment while placing results back in a comparable coordinate system.
As a vehicle for differentiable physics, the fully differentiable solvers combine with GPPO to form a path that backpropagates analytic gradients into policy parameters. The paper notes this combination supports both accelerated policy optimization and inverse design, working backwards from desired flow features to the required forcing or geometry, and describes such environments as computational experiments probing the limits of what is physically achievable.
As a candidate pretraining corpus, the paper's end-state architecture is a foundation model for flow control, pretrained across the diversity of HydroGym environments and adaptable to new configurations with few or zero shots. The analogy drawn is to large language models, whose generalization stems from the availability of standardized, large-scale, diverse pretraining corpora rather than any single algorithmic breakthrough. The zero-shot transfer result is presented as an indicator that this path is meaningful.
As a surrogate training tier for industrial-scale problems, the paper supplies a concrete engineering pattern for fields where high-fidelity simulation is too costly per episode: train control priors in cheap configurations, deploy directly in expensive ones.
Relative to other work on this site, HydroGym occupies the environment and interface layer. It is complementary to surrogate-model reduction and intersects concretely with multiphysics coupling through the MAIA MPMD execution mode. Keeping that division of labour clear avoids reading an environment benchmark as a reduction scheme. Concrete comparisons are available in the site's fluid simulation toolkit for solver selection, numerics and PDE solvers for discretization and differentiability, surrogate modeling and reduced-order modeling and neural operators for the alternative reduction routes, and performance engineering and GPU acceleration for compute provisioning.
On application cases, the platform's wing and aerodynamic environments map directly onto aerodynamics entries such as the AirfRANS airfoil high-fidelity surrogate benchmark on the data-supply side and SU2 aerofoil shape optimization on the adjacent optimization task, while wind turbine aeroelastic simulation and wake control covers a concrete control object and multiphysics coupling simulation corresponds to the multiphysics and reactive-flow directions planned for future releases.
8. Limitations and Failure Modes
It does not support engineering delivery commitments. Every result is numerical and the actuators are idealized boundary conditions. Control gains carry no model of drive power, electrical supply, compressed air or installation feasibility, so quoting a 90% drag reduction in a product selection or tender response is not defensible.
It does not support direct training at industrial Reynolds numbers. The stated ceiling is Re_c = 400,000, and although the authors attribute this to training-phase compute rather than solver capability, the practical consequence is that a high-fidelity simulation cannot enter a training loop at all, which is the same issue as the coverage limit viewed from another angle.
Transfer failure modes are specific. The zero-shot result depends on shared near-wall turbulence mechanisms. The paper names the flow families where it does not hold: strongly pressure-gradient-driven, separation-dominated and compressible flows. A second mode is unstable performance when geometry and operating point mismatch, with zero-shot performance varying across scenarios while subsequent fine-tuning consistently benefits from the pretrained initialization even where immediate transfer looks limited.
Physical scaling matters in deployment. Observations are standardized with spatially varying viscous scales on the wing, and because the viscous time unit evolves along the chord, deployed action update frequency is scaled by the local streamwise-averaged friction velocity of each block. Such scaling depends on wall-turbulence similarity, which fails where the physics differs.
Benchmark comparability has its own edge. The paper explicitly leaves robustness to other configurations and uses to the reader's judgement, so cross-configuration generalization currently rests on a single empirical path, the zero-shot transfer experiment.
The DNS compute trap is real. At high Reynolds numbers, DNS cost for grid and time step grows together. Platform backends are powerful, yet wiring one directly into a training loop drains the compute budget quickly; cost should be computed against the training configuration, not the deployment configuration.
Community maturity is modest. The repository shows 572 stars, 85 forks and 49 open issues, with only three formal releases: v0.1.2, v0.1.2.1 and v1.0.1, the last published on 24 April 2026. Version cadence is sparse, so anyone pinning specific backend versions should verify environment compatibility directly.
9. Version and Source Notes
On publication and dating: the platform paper appeared in Nature on 19 August 2026, volume 657, pages 369-376, DOI 10.1038/s41586-026-10917-6. The preprint is arXiv:2512.17534, with v1 submitted 19 December 2025 and v2 revised 30 June 2026. The repository was created on 29 December 2021, and the MIT license carries a 2021 copyright line. The original entry recorded year 2001, a dating error; this entry records 2026, the paper year, while the body also gives the repository creation year and the license copyright year.
On titles: the arXiv preprint is titled "The HydroGym Reinforcement Learning Platform for Fluid Dynamics", the Nature version "The HydroGym reinforcement learning platform for fluid dynamics", the README banner "HydroGym: Reinforcement Learning for Fluid Dynamics", and the repository description "An RL-Gym for Challenge Problems in Data-Driven Modeling and Control of Fluid Dynamics". Citation should track the specific source.
On authorship and institutions: the paper lists 21 authors, with Christian Lagemann first and Ricardo Vinuesa and Steven L. Brunton as senior authors. Institutions span the University of Washington, RWTH Aachen, TU Munich, the University of Michigan, Princeton, Cornell and the University of Florence. The README asks users to cite two items: the Lagemann et al. conference paper at the 7th Learning for Dynamics & Control Conference (PMLR, 2025, pages 497-512), and the arXiv preprint of the Nature paper.
On environment counts: the paper body says 61, the abstract says more than 60, the README says 61+, and the Hugging Face repository shows 90 directories. These reflect different conventions and must be labelled when quoted.
On solver backends: the README table lists six backends with environment counts of Firedrake 20, MAIA LBM 55, MAIA Structured FV 8, NEK5000 2, JAX 2 and JAX-Fluids 2, totalling 89 configurations; the paper body instead groups environments into 14 geometric configurations and 61 flow scenarios. Table counts and scenario counts are not interchangeable.
On drag numbers: the abstract's "exceeding 90%" corresponds to the 2D fluidic pinball at Re = 100 under counter-rotation, a particularly favourable configuration rather than a platform-wide level. The 3D cylinder multi-agent figure is about 8%. The zero-shot transfer yields 38% local skin-friction reduction and 11% total drag reduction, and the two must not be merged.
On cost figures: the four orders of magnitude comes from the core-hour comparison between channel surrogate training and direct on-wing training, matching the stated exploration efficiency gain exceeding 10^4. A separate number applies even with periodic validation on the full wing geometry, where total wall-clock time falls by roughly a factor of 60. These rest on different comparison baselines and cannot be collapsed into a single universal speedup claim.
On the experimental validation claim: the original entry mentioned validation of repeatable deployment in a wind tunnel or plasma experimental facility. That does not match the paper, which states all results come from direct numerical simulation and treats experimental deployment as future work. The discussion cites laboratory rotating-cylinder studies as supporting evidence for the viability of reinforcement learning control, with drag reductions of 25-30% at Reynolds numbers up to 140,000, but that is referenced prior work rather than a platform result.
Other corrections to the original entry: the claim of six CFD solver backends is accurate; "dozens of pre-configured environments" understates the 61 (or 61+ by README) actually provided; the attribution to the University of Washington, RWTH Aachen and TU Munich is directionally right but omits participants such as the University of Michigan; the description of a fully differentiable JAX solver is correct but needs JAX-Fluids added as a distinct second differentiable backend; and the NACA 0012 mention does not distinguish steady, gust, spanwise, 2D and 3D variants.
On versions and dependencies: the formal release is v1.0.1, published 24 April 2026 as the Full Multi-Solver RL Environments Release, preceded by v0.1.2 and v0.1.2.1, both on 26 December 2022. The versioned code snapshot used in the paper is archived on Zenodo under DOI 10.5281/zenodo.21062222, record version 1.0.1, published 30 June 2026, MIT licensed, with prebuilt containers hosted in the clagemann Docker Hub repository. The default branch is main, with the most recent push on 11 November 2026.
On data availability: the Hugging Face repository provides checkpoints and trajectories, but the largest DNS datasets cannot be deposited in full, since individual flow fields reach 75 GB and full test trajectories reach 15 TB; complete data must be requested from the corresponding authors. Any claim that all data is already public needs correcting.
References
- Lagemann C. et al., "The HydroGym reinforcement learning platform for fluid dynamics", Nature, vol. 657, pp. 369-376, 19 August 2026, DOI 10.1038/s41586-026-10917-6, https://www.nature.com/articles/s41586-026-10917-6
- Lagemann C. et al., "The HydroGym Reinforcement Learning Platform for Fluid Dynamics", arXiv preprint 2512.17534 (v1 19 December 2025, v2 30 June 2026), https://arxiv.org/abs/2512.17534
- dynamicslab/hydrogym repository and README, covering the environment list, solver backend table, Docker image names, example directory layout and the two-item citation requirement, https://github.com/dynamicslab/hydrogym
- MIT license file of dynamicslab/hydrogym, Copyright (c) 2021 Jared Callaham, https://github.com/dynamicslab/hydrogym/blob/main/LICENSE
- Lagemann C. et al., "HydroGym: A Reinforcement Learning Platform for Fluid Dynamics", 7th Annual Learning for Dynamics & Control Conference, PMLR, 2025, pp. 497-512, the conference paper listed in the README citation block
- dynamicslab/HydroGym-environments, Hugging Face dataset repository holding environment checkpoints and trajectories, directly inspected at 1513 files across 90 environment directories, https://huggingface.co/datasets/dynamicslab/HydroGym-environments
- Versioned HydroGym code snapshot, Zenodo record 21062222, version 1.0.1, 30 June 2026, DOI 10.5281/zenodo.21062222, https://doi.org/10.5281/zenodo.21062222
- HydroGym official API documentation, https://dynamicslab.github.io/hydrogym
- HydroGym flow configuration documentation at docs/FlowConfigurations.md, https://github.com/dynamicslab/hydrogym/blob/main/docs/FlowConfigurations.md
- GitHub API repository metadata covering creation date 2021-12-29, license, star, fork and issue counts, topics and latest push time, https://api.github.com/repos/dynamicslab/hydrogym
- GitHub API release listing showing v1.0.1 published 2026-04-24 and v0.1.2 with v0.1.2.1 published 2022-12-26, https://api.github.com/repos/dynamicslab/hydrogym/releases
- Prebuilt container repository clagemann with CUDA 12.9 Hopper and Blackwell, Turing and Ampere, and ROCm 6.3.3 images, https://hub.docker.com/repositories/clagemann
- Figures 3 and 4 of the paper, showing PPO, DDPG and TD3 training curves and test rollouts on the 2D and 3D benchmark environments, referenced in item 1
- Figure 7 of the paper, showing zero-shot transfer from an Re_tau = 206 channel-flow surrogate to a Re_c = 200,000 3D NACA0012 wing with roughly 38% local skin-friction reduction and 11% total drag reduction, referenced in item 1