New lunar AI cuts error by 22% on a map of ice-favorable conditions

How can scientists combine lunar photographs, terrain, temperature, radar, and chemical-composition data without building a new algorithm for every question? NASA and IBM have released a foundation model—a pretrained system that learns reusable representations—so teams can adapt it to tasks such as mapping craters, small volcanic features, or conditions favorable to ice preservation. The model, prepared datasets, and code are open for testing.

Training used 963,609 bundles anchored to multispectral images from the Lunar Reconnaissance Orbiter Wide Angle Camera, at roughly 100 meters per pixel, and 1,000,113 bundles anchored to Narrow Angle Camera images, at roughly 1 meter per pixel. Each bundle combines products registered over the same area. In all, the system received 11 modalities, including imagery, terrain, and context from other instruments, along with illumination angles that strongly change how lunar ground appears.

After pretraining, the researchers adapted the same visual encoder to four benchmarks: crater detection at both scales, segmentation of small volcanic features called irregular mare patches, and reproduction of a polar ice-prospectivity map. Data, splits, transformations, and metrics were held fixed for the lunar model and comparison architectures. Reported averages came from five training seeds, revealing how much scores vary from initialization alone.

The widest gap appeared on the polar map. Root mean squared error, a measure that penalizes larger deviations and is better when lower, fell from 0.0377 ± 0.0004 for the best baseline to 0.0293 ± 0.0013 for the lunar model—a relative reduction of about 22%. The target, however, was a map calculated from environmental indicators, not observed ice. In wide-scale crater detection using half the labeled data, mean average precision rose from 0.2313 ± 0.0027 to 0.2541 ± 0.0018. On meter-scale craters and volcanic-patch segmentation, the leaders were treated as comparable because the gaps were smaller than the variation across seeds.

Two design choices help explain transfer across tasks. The system receives illumination geometry explicitly rather than confusing shadows with terrain properties. It also keeps different-resolution inputs in separate blocks before combining their signals. In the polar test, part of the advantage already came from this organization of modalities; lunar pretraining accounted for the remaining gain. The authors have not yet experimentally isolated every architectural contribution.

The release provides a tool for comparing hypotheses and accelerating triage of large archives, not an autonomous cartographer ready for mission decisions. The model lacks an absolute geodetic reference, its high-resolution training coverage is limited to selected sites, and it has not been validated for landing certification or hazard clearance. Its scientific value will depend on outside teams reproducing the tests and checking with instruments and new labels whether the gains persist beyond the four evaluated datasets.