Benchmark RadarThousandWorlds
ThousandWorlds is a benchmark for climate emulation of potentially habitable exoplanets. It provides a dataset of approximately 1800 simulations from five global climate models, mapping eight planet parameters to 3D atmospheric fields. It includes three nested benchmark subsets: single-simulator regression, multi-simulator regression with complete observations, and multi-simulator regression with structured missingness. Two evaluation protocols are provided: one for ranking methods and one measuring performance relative to inter-model disagreement.
- Released
- 2026-06-16
- Readiness
- Runnable
- Primary field
- Science & Research
Why it matters
Machine-learning emulators could accelerate exoplanet climate modeling, but progress was limited by the lack of a curated multi-model dataset. This benchmark fills that gap, enabling reproducible comparison of emulation methods and addressing a regime of low-data, multi-simulator regression where deep learning underperforms.
Motivation
The search for life beyond Earth will depend on detecting faint signatures in the atmospheres of potentially habitable exoplanets.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.