Yorktown Heights, N.Y. – September 15, 2026 -- IBM and NASA released the open-source NASA-IBM Lunar Foundation Model, one of the first publicly available foundation models built for scientific exploration of the Moon, outperforming widely used detection methods by up to 23% in identifying craters, ice deposits and volcanic formations.
Model cuts ice-detection error by up to 22% versus baseline
The model reduced root-mean-square error in identifying areas with high potential for lunar ice by up to 22% compared with the SwinV2-B (ImageNet) benchmark, according to a technical paper authored by IBM and NASA researchers. Permanently shadowed lunar regions -- among the hardest environments to observe -- may hold subsurface ice, a resource NASA considers essential for a future Moon base and for producing rocket fuel for Mars missions.
Crater mapping accuracy rises nearly 19% at context-scale resolution
At roughly 100-meter context-scale resolution, the model outperformed SwinV2-B by nearly 19% while using only half the training data typically required, the paper shows. At meter-scale resolution, it matched state-of-the-art accuracy with lower fine-tuning costs, data NASA uses to select safe landing sites and plan long-term lunar infrastructure.
Volcanic feature detection improves 3% with lower fine-tuning costs
Using imperfect labels, the model captured the extent of Irregular Mare Patches -- volcanic features tied to the Moon's thermal history -- 3% more accurately than SwinV2-B, delivering comparable accuracy at greater efficiency, per the study.
Dataset unifies 30 layers from nine instruments across four missions
Alongside the model, IBM and NASA built the first open-source, machine-learning-ready lunar dataset of its kind, aggregating more than 30 spatially aligned layers from nine instruments across four missions. The dataset combines tens of thousands of images and maps from NASA's Lunar Reconnaissance Orbiter and GRAIL missions with complementary data from the Japan Aerospace Exploration Agency's SELENE/Kaguya mission.
"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," said Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters. IBM Research Europe, UK and Ireland Director Juan Bernabe-Moreno said the model gives scientists a shared foundation to explore the Moon at scale by connecting observations across instruments and surfacing patterns difficult to detect in isolation.
Model joins IBM's Prithvi family spanning geospatial and heliophysics domains
The lunar model extends the Prithvi family of open foundation models, which already covers geospatial, weather and heliophysics data, under a shared IBM-NASA approach that lets researchers adapt one base model to multiple scientific tasks rather than building separate systems for each question.