Overview
- IBM and NASA published the NASA‑IBM Lunar Foundation Model and the accompanying unified lunar dataset on Hugging Face, a move announced on Sept. 10, 2026 that makes the tools publicly available to researchers.
- The released dataset combines more than 30 spatially aligned data layers drawn from nine instruments on four missions, including NASA’s Lunar Reconnaissance Orbiter and GRAIL, and incorporates tens of thousands of images and maps at multiple resolutions.
- In benchmark tests the foundation model identified key lunar features up to 23% more accurately than common methods, with roughly 19% better crater detection and reductions in ice‑prospectivity errors of as much as 22% versus a SwinV2‑B baseline.
- The system uses multimodal, multiresolution training with masked‑token learning and explicit lighting and spacecraft metadata to handle extreme lunar shadows and scale differences, and it supports lightweight fine‑tuning methods so teams can adapt it without retraining from scratch; the developers stress it is a pattern‑recognition aid that does not replace direct surface measurements.
- The open release builds on IBM–NASA foundation models for Earth and the Sun and is aimed at accelerating site selection, resource prospecting, and scientific study ahead of Artemis-era crewed missions, while inviting the global research community to refine and validate the model.