NASA-IBM Lunar Foundation Model cuts ice error up to 22%

The release sits inside the NASA-IBM "AI for Science" collaboration, which has run under a Space Act Agreement since early 2022. The two organizations put out their first Prithvi model on Hugging Face in August 2023, trained on Landsat and Sentinel-2 imagery of the contiguous US and adapted for flood and wildfire mapping. The lunar model follows that pattern, and its base, TerraMind, was developed by IBM in 2025 with ESA and Forschungszentrum Jülich for Earth observation, so the underlying architecture was not built for the Moon specifically.

The training corpus is unusually broad for lunar work. It brings together more than 30 spatially aligned data layers from nine instruments and four missions, including NASA's GRAIL gravity data and Lunar Prospector hydrogen readings plus JAXA's Kaguya/SELENE mineralogy data. The team split the corpus geographically by map zones rather than randomly distributing tiles, a choice aimed at preventing leakage between training, validation, and test data. One design decision stands out: rather than asking the model to infer lighting from raw pixels, the team feeds illumination angles, sun position, and tile extent as explicit context, since lighting geometry shapes how the lunar surface looks far more than its actual properties do.

The results are mixed by task. The largest gains appeared in polar ice prediction, while on meter-scale crater detection and Irregular Mare Patches the model roughly ties the strongest baselines, with differences falling within the variance between training runs; IBM claims a 3 percent lead on IMPs, but the reported results look more comparable. The authors present the model as a reusable foundation for downstream work, not a substitute for physical measurement instruments, and controlled experiments isolating each innovation are still pending. How much of the edge comes from the lunar pretraining versus the architecture is accordingly an open question, and it likely matters most to researchers deciding whether to build on this model or on their own task-specific pipelines.