Exoplanet Climate Benchmark · AstroAutomata
ThousandWorlds
A multi-simulator benchmark dataset of exoplanet general circulation models for spatial climate emulation.
What ThousandWorlds is
ThousandWorlds is an exoplanet climate emulation benchmark developed by AstroAutomata to evaluate machine learning models on predicting planetary atmospheric states from physical orbital and stellar parameters. General circulation models (GCMs) provide detailed physics-based climate simulations, but their computational cost prevents comprehensive parameter space explorations. ThousandWorlds formulates this challenge as a structured parameter-to-field regression task across multiple planetary simulators.
The dataset aggregates 1,689 simulations generated across five distinct GCMs, varying eight planetary and stellar characteristics. Targets consist of two-dimensional and three-dimensional time-averaged climate fields mapped to a standardised spatial grid, accompanied by spherical harmonic spectral coefficients.
Designed to address multi-simulator transfer and spatial physical surrogate modelling, the benchmark includes predefined splits across three nesting tiers and two formal evaluation protocols, accompanied by ten baseline algorithms.
How it was made
The dataset compiles simulations run across five GCMs. Inputs comprise eight continuous physical features (stellar temperature, incident stellar flux, planetary radius, surface gravity, rotation period, surface pressure, and atmospheric volume fractions of carbon dioxide and methane) alongside a categorical GCM identifier. Target outputs are time-averaged atmospheric fields standardised onto a 32 × 64 latitude-longitude grid across 10 relative vertical pressure levels for 3D variables.
The atmospheric outputs are distributed in two forms: gridded NumPy arrays (including 2D surface fields and 3D variables stored as pressure-level channels) and T21 spherical harmonic coefficients (484 coefficients per field) with inverse spherical harmonic transform weights. Simulations are structured into three nested evaluation tiers: single-complete (277 simulations from one GCM), multi-complete (1,595 simulations across five GCMs with complete variables), and multi-partial (1,689 simulations across five GCMs containing structured whole-field missingness recorded as NaNs and tracked via spectral masks).
Key facts
- Total simulations
- 1,689 (v2.0.0)
- Source climate models
- 5 GCMs
- Input parameters
- 8 continuous physical parameters plus source GCM label
- Spatial resolution
- 32 × 64 latitude-longitude grid across 10 vertical pressure levels
- Output variables
- 8 climate variables (up to 53 total channels in full grid)
- Spectral representation
- T21 spherical harmonics (484 coefficients per field)
- Data formats
- NumPy (.npz) and CSV
- Release version
- v2.0.0
- Benchmark subsets
- single-complete, multi-complete, multi-partial
What you can use it for
- Climate Surrogate Modelling
- Train spatial machine learning models, such as Fourier neural operators, DeepONets, or convolutional decoders, to emulate 2D and 3D planetary atmospheric fields directly from scalar planetary and stellar inputs.
- Multi-Simulator Transfer Learning
- Develop models capable of transferring learned climate dynamics across different numerical GCM architectures using the multi-GCM splits.
- Missing Field Imputation
- Evaluate regression methods on predicting incomplete or missing planetary climate fields under structured whole-field observation masks.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from AstroAutomata/ThousandWorlds at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 1,689 rows of input_planets/all
ch4
float- min
- 0
- median
- 0
- mean
- 4.5e-4
- max
- 0.04
co2
float- min
- 0
- median
- 3.5e-4
- mean
- 0.03
- max
- 1
gcm_label
string_label- exoplasim66%
- um17%
- exocam-pre20229%
gravity
float- min
- 4.70
- median
- 9.81
- mean
- 9.75
- max
- 20
in_target_physical_domain
bool- True54%
- False46%
is_target_gcm
bool- False77%
- True23%
planet_id
int- min
- 0
- median
- 757
- mean
- 770.32
- max
- 1,596
radius
float- min
- 1.6M
- median
- 6.4M
- mean
- 6.4M
- max
- 18M
rotation_period
float- min
- 0.25
- median
- 13.03
- mean
- 22.83
- max
- 220
simulation_id
int- min
- 0
- median
- 844
- mean
- 844
- max
- 1,688
source
string_label- Paradise et al. [2022b]28%
- this work25%
- Paradise et al. [2021]23%
stellar_flux
float- min
- 400
- median
- 1,100
- mean
- 1228.91
- max
- 3,100
stellar_temperature
float- min
- 2,500
- median
- 3,000
- mean
- 3725.76
- max
- 5,777
surface_pressure
float- min
- 10,000
- median
- 100k
- mean
- 239k
- max
- 1.2M
Strengths and limits
Good at
- Standardises outputs from five separate exoplanet GCMs onto a unified 32 × 64 grid and spherical harmonic representation.
- Separates shared simulated planets into dedicated splits to prevent data leakage and benchmark model performance against the inter-GCM epistemic uncertainty floor.
- Supplies ten benchmark baselines spanning linear methods, gradient boosting, coordinate networks, and operator learning frameworks.
Watch out for
- Variables are restricted to time-averaged climate states rather than time-resolved temporal dynamics.
- Spatial outputs are constrained to a low-resolution 32 × 64 grid with 10 pressure levels.
- Inputs are bounded to terrestrial and sub-Neptune parameter ranges, covering radii between 0.7 and 1.75 Earth radii and surface pressures up to 5 bar.
Questions
Can I look inside ThousandWorlds without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is ThousandWorlds under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What planetary parameters are varied across the dataset?
The inputs include planetary radius (0.7 to 1.75 Earth radii), surface gravity (6.0 to 17.0 m s^-2), rotation period (1.0 to 200.0 days), surface pressure (0.5 to 5 bar), atmospheric CO2 volume fraction (0 to 100%), atmospheric CH4 volume fraction (0 to 5%), incident stellar flux (500 to 1500 W m^-2), and stellar temperature (2500 to 5800 K).
Which atmospheric variables are predicted?
Target variables comprise two-dimensional surface temperature, absorbed shortwave radiation, and outgoing longwave radiation, alongside three-dimensional specific humidity, cloud fraction, east-west wind, north-south wind, and atmospheric temperature across vertical pressure levels.
What is the purpose of the shared-planets evaluation protocol?
The shared-planets protocol evaluates models exclusively on planet configurations that were simulated by multiple GCMs, measuring whether predictions reach the limit of inter-model disagreement.
How to cite it
Stevenson, Edward T. and Mak, Mei Ting and Wolf, Eric and Sergeev, Denis E. and Hammond, Tobi and Mayne, N. J. and Cranmer, Miles, 'ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets', arXiv:2606.18338, 2026. doi:10.48550/arXiv.2606.18338
Written from ThousandWorlds's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.