File size: 6,565 Bytes
8d5a693 d761306 cb312a7 18e3500 8d5a693 93830e3 f581fd1 8d4bfc3 f581fd1 ca600b1 8d5a693 d761306 8d5a693 f20c22c 8d5a693 f581fd1 8d5a693 f581fd1 8d5a693 f581fd1 8d5a693 f581fd1 8d5a693 f20c22c 8d5a693 6120d9e 8d5a693 6120d9e 8d5a693 f20c22c 8d5a693 ca600b1 8d5a693 f581fd1 8d5a693 f581fd1 8d5a693 f581fd1 8d5a693 85719d5 8d5a693 85719d5 8d5a693 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 | ---
pretty_name: ThousandWorlds
license: cc-by-4.0
size_categories:
- 1K<n<10K
task_categories:
- tabular-regression
- other
tags:
- benchmark
- datasets
- physical-sciences
- scientific-machine-learning
- exoplanets
- climate
- astronomy
- emulation
- simulation
- physics
- pde
- parameter-to-field-regression
- structured-outputs
- multi-simulator-transfer
- spatiotemporal
viewer: false
---
# ThousandWorlds
<img src="imgs/MASCOT.png" align="right" width="220" style="margin-top: -1.25rem;" alt="ThousandWorlds mascot">
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.
[](https://github.com/edstevenson/ThousandWorlds)
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.

## Quickstart
The easiest way to use the benchmark is through the
[Python code](https://github.com/edstevenson/ThousandWorlds):
```bash
git clone https://github.com/edstevenson/ThousandWorlds.git
cd ThousandWorlds
pip install -e .
```
```python
import numpy as np
import thousandworlds as tw
tw.download_dataset(".")
bundle = tw.load("single-complete", data_dir="dataset")
pred = np.broadcast_to(bundle.Y_train.mean(axis=0), bundle.Y_test.shape)
scores = tw.evaluate.rmse(pred, bundle.Y_test, bundle.field_mask_test, bundle.field_names)
scores["per_variable"]
```
See the GitHub repository for 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
- Archival mirror: https://doi.org/10.7910/DVN/8IEH6Q (Harvard Dataverse)
- Paper: coming soon!
## Citation
If you use ThousandWorlds, please cite the paper:
```bibtex
@misc{thousandworlds2026,
title = {ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets},
author = {{ThousandWorlds authors}},
year = {2026},
note = {Manuscript in preparation}
}
```
|