DeepMind Hamiltonian Dynamics Suite: 17 Simulated Physical Systems for Learning Latent Dynamics
- Provider
- Google DeepMind (Botev, Jaegle, Wirnsberger, Hennes, Higgins)
- License
- Apache-2.0 (source)
- Stored at
- Google Cloud Storage, gs://dm-hamiltonian-dynamics-suite (us)
- Access
- Free and public. No account with the data provider is needed.
- Category
- Physics
- Source
- https://arxiv.org/abs/2111.05458
About this dataset
Simulated dynamical systems built by DeepMind to test models that learn physics from image observations, including Hamiltonian and Lagrangian neural networks. Seventeen datasets in four groups: Toy Physics (mass-spring, pendulum, double pendulum, two-body, each with colour-randomised, friction and long-trajectory variants, integrated with scipy.integrate.solve_ivp at dt = 0.05 and friction coefficient 0.05 for the non-conservative variants), Molecular Dynamics (Lennard-Jones systems of 4 and 16 particles), Multi Agent (matching pennies and rock-paper-scissors replicator dynamics in non-transitive zero-sum games) and Mujoco Room (3D camera motion on circle and spiral paths). Each dataset has 50,000 training and 20,000 test trajectories; every trajectory carries rendered frames, the ground-truth phase-space state, its first time derivative and the per-trajectory parameters that were randomised (mass, spring constant, gravitational constant, pivot length, colours). Measured 2026-11-15: 65 objects, about 259 GB, one folder per system holding train.tfrecord, test.tfrecord and a features.txt schema; the two Lennard-Jones shards are the largest at 30 GB and 12 GB. Read with tf.data or the loaders in dm_hamiltonian_dynamics_suite. Apache-2.0 per the bucket LICENSE; cite Botev et al. 2021, Which priors matter? Benchmarking models for learning latent dynamics.
How to access this dataset in Python
import gcsfs
fs = gcsfs.GCSFileSystem(token='anon')
fs.ls('dm-hamiltonian-dynamics-suite/', detail=False)[:10]Lists the files with anonymous access. Runs on any machine, with no cloud account.
Data structure
Files
Example file: molecular_dynamics/lj_16/features.txt
Analyze this dataset without downloading it
On Scigantic, DeepMind Hamiltonian Dynamics Suite: 17 Simulated Physical Systems for Learning Latent Dynamics opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.
This page is also available as Markdown and as JSON.
Related Physics datasets
- The Well: Acoustic Scattering with Random Inclusions
- TIDE Helical Forced Turbulence: 3D Incompressible Navier-Stokes DNS at 256^3 (Dai et al.)
- hls4ml LHC Jet Dataset (150 Particles): Five-Class Jet Tagging for FPGA Trigger Studies
- Quantum Physics 0.6
- ColliderML Release 1: Simulated HL-LHC Collision Events in the Open Data Detector
- Cold Plate CFD: 2,000 Steady Conjugate Heat Transfer Simulations of a Dual-Chip Pin-Fin Liquid Cold Plate
- Fusion Equilibrium Challenge
- VKI-LS59: 2D Turbine Cascade RANS Dataset under Geometric Variations (PLAID)