ThousandWorlds

ThousandWorlds is a benchmark for emulating exoplanet climates: 1760 simulations across 5 GCMs, 8 planet parameters, and atmospheric variables on a 32 x 64 x 10 latitude-longitude-pressure grid. It includes three nested benchmark subsets, two evaluation protocols, and eight released baseline methods.

Inputs are 8 continuous planet parameters plus the source GCM label. Outputs are time-averaged climate fields on a 32 x 64 latitude-longitude grid: three-dimensional variables are stored as pressure-level channels, and two-dimensional variables are stored as single-level fields.

Want to look around before downloading anything? Try the ThousandWorlds Explorer! Built by Hamza Ali Shahjahan!

Quickstart

The easiest way to use the benchmark is through the Python code:

git clone https://github.com/edstevenson/ThousandWorlds.git
cd ThousandWorlds
pip install -e .

import thousandworlds as tw

tw.download_dataset()
bundle = tw.load("single-complete", data_dir="dataset")

See the GitHub repository for the full quickstart, notebooks, baseline code, evaluation utilities, and reproducing paper results.

Files

The release includes:

• archives/dataset.tar.gz: the ThousandWorlds dataset.
• archives/results-baselines-*.tar.gz: baseline predictions for the 3 subsets.
• croissant.json: Croissant metadata.
• archives/*.sha256: checksum sidecars.

Dataset Contents

The dataset contains gridded fields (NumPy), input metadata (CSV), predefined train/test splits, normalization statistics, and spherical harmonic coefficients plus inverse-SHT weights for spectral methods.

Subsets

The dataset is organized into three subsets of increasing complexity and realism:

Subset	Simulations	Fields	Description
single-complete	256	48	Smaller subset; simulations from a single GCM, complete observations only.
multi-complete	1659	48	All 5 GCMs, still with no missing fields.
multi-partial	1760	53	Full dataset; all 5 GCMs, with missing fields represented as NaNs.

The subset split files contain:

File	single-complete	multi-complete	multi-partial
train.csv	206	1538	1626
test.csv	50	90	100
test_shared_planets_only.csv	-	58	60
held_out_aux.csv	-	31	34

held_out_aux.csv is excluded from train and test to prevent train-test leakage (it contains simulations from auxiliary GCMs that correspond to identical planets present in the test set).

Inputs

Each simulation has one row in dataset/inputs.csv, keyed by simulation_id. The public model inputs are stellar temperature, stellar flux, radius, gravity, rotation period, surface pressure, CO2, CH4, and gcm_label. The metadata also includes is_target_gcm, in_target_physical_domain, planet_id, and source.

Parameter	Range
Radius (Earth radii)	[0.7, 1.4]
Surface gravity (m s^-2)	[6.0, 16.0]
Rotation period (days)	[0.1, 1000.0]
Surface pressure (bar)	[0.5, 5]
CO2 volume fraction (%)	[0, 100]
CH4 volume fraction (%)	[0, 5]
Incident stellar flux (W m^-2)	[500, 1500]
Stellar temperature (K)	[2500, 5800]

Outputs

Target fields include surface temperature, 3D temperature, specific humidity, cloud fraction, east-west wind, north-south wind, absorbed shortwave radiation, and outgoing longwave radiation. Gridded targets are provided on a 32 x 64 latitude-longitude grid, with vertical fields stored on relative pressure levels.

Variable	Dimensionality	Unit
Surface temperature	2D	K
Temperature	3D	K
Specific humidity	3D	dex
Cloud fraction	3D	1
East-west wind	3D	m s^-1
North-south wind	3D	m s^-1
Absorbed shortwave radiation	2D	W m^-2
Outgoing longwave radiation	2D	W m^-2

The gridded field archives are:

File	Shape	Contents
dataset/fields/all-obs.npz	(1760, 53, 32, 64)	Field archive covering all 5 GCMs with structured whole-field missingness.
dataset/fields/complete-obs-only.npz	(1659, 48, 32, 64)	Complete-observation field archive.

Spectral Coefficients: The spectral coefficient archives mirror those field archives with T21 spherical harmonic coefficients: dataset/coefficients/*.npz stores coefficients with 484 coefficients per field and a field_mask for missing fields. Whole-field missingness is represented as all-NaN gridded channels and as false entries in the spectral field_mask.

Evaluation

The package includes loaders and metrics for two benchmark protocols:

• Standard: the main test protocol, ideal for ML model comparison.
• Shared-planets: evaluate on planets shared across target and auxiliary GCMs; used to assess performance relative to inter-GCM error, i.e. how close a model gets to the epistemic uncertainty floor of the problem.

Released baselines include train mean, kNN, PCA ridge, PCA-MLP, Coord-MLP, Coord-DeepONet, PPCA-ICM, and GPLFR. Baseline artifacts include predictions, resolved configs, and metrics JSON files.

Links

• DOI: https://doi.org/10.57967/hf/8695
• Code: https://github.com/edstevenson/ThousandWorlds
• Paper: https://arxiv.org/abs/2606.18338

Citation

If you use ThousandWorlds, please cite the paper:

@article{thousandworlds2026,
  title = {ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets},
  author = {Stevenson, Edward T. and Mak, Mei Ting and Wolf, Eric and Sergeev, Denis E. and Hammond, Tobi and Mayne, N. J. and Cranmer, Miles},
  year = {2026},
  eprint = {2606.18338},
  archivePrefix = {arXiv},
  doi = {10.48550/arXiv.2606.18338}
}