Particle.news
Get it on Google Play
Download on the App Store
13 ARTICLES
·
23h ago

NASA and IBM Release Open Lunar Foundation Model

Researchers gain a unified, machine-learning-ready lunar dataset plus a pretrained model that speeds crater mapping, finds likely ice deposits, and informs Artemis landing planning

Reference and model-generated lunar polar ice-prospectivity maps, with yellow indicating higher prospectivity and blue indicating lower prospectivity
Image
Image
Image courtesy of NASA

Overview

  • The NASA‑IBM Lunar Foundation Model and its SomBench dataset were released Thursday, Sept. 10, 2026, and are available open‑source on Hugging Face with code and tools hosted on GitHub.
  • SomBench is a machine‑learning‑ready benchmark of roughly 2 million aligned image tiles that layers more than 30 spatially aligned datasets from nine instruments across four missions to unite multiresolution observations.
  • NASA and IBM report the model cuts error or raises accuracy by about 22–23% on selected tasks versus common baselines and can match or improve crater mapping, volcanic‑feature detection, and ice prospectivity estimates while requiring fewer labeled examples.
  • The release includes integration with the TerraTorch toolkit and explicit illumination metadata so the model can handle extreme lunar lighting, but NASA and IBM caution its outputs are research aids that need independent validation before use in operations.
  • The effort builds on the Prithvi family of science foundation models and was trained mainly on 17 years of LRO data supplemented by GRAIL, Lunar Prospector and JAXA’s SELENE, with the next phase focused on community adoption, fine‑tuning and mission‑specific testing.