HydroGym: A Reinforcement Learning Platform for Flow Control
HydroGym provides a Gymnasium-compatible interface for training reinforcement learning agents on CFD problems, developed by a collaboration spanning the University of Washington, RWTH Aachen, TU Munich and others, and published in Nature.
What it offers: Dozens of pre-configured flow environments (cylinder, fluidic pinball, open cavity, NACA 0012 airfoil, turbulent channel, Kolmogorov flow, turbulent boundary layer) backed by finite-element (Firedrake), lattice-Boltzmann and finite-volume (MAIA), spectral-element (NEK5000) and fully differentiable (JAX) solvers. It runs from a laptop to exascale HPC and works with Stable-Baselines3 or RLlib.
Reported results: A policy trained on a turbulent channel flow at a friction Reynolds number of about 206 was deployed zero-shot onto 3D NACA 0012 wing sections, reducing surface skin friction by roughly 38% and overall drag by about 11%, at an estimated four orders of magnitude lower core-hour cost than training directly on the wing.
Industrial relevance: Turbulent drag accounts for a large share of energy loss in transport and pipeline flow. HydroGym turns flow control from bespoke one-off studies into a benchmarked, reproducible workflow.
- Wind-Turbine Blade Load Prediction & Surrogate · Data-driven prediction of blade fatigue load and power curve, solving hard-to-ev
- Tokamak Plasma Magnetic Confinement Control · Deep reinforcement learning directly commands tokamak magnetic coils to hold div
- Physics-Informed CNN for Battery Pack Temperature Fields · A physics-informed convolutional neural network whose loss encodes the finite-di
- MACE-MP-0: A Universal Machine-Learned Interatomic Potential · A general-purpose machine-learned interatomic potential built on an equivariant
- TopoDiff: Diffusion-Model-Driven Topology Optimization · A conditional diffusion approach to structural topology optimization from MIT th