NASA and IBM Research have released an open-source foundation model built for lunar science. The model turns 17 years of observations into reusable features that researchers can adapt to tasks such as finding craters and estimating ice near the Moon’s poles.
Training used SomBench, described by the team as the largest co-registered multimodal lunar dataset to date. It contains nearly two million aligned tile bundles across 11 kinds of data and two spatial scales. Roughly one million high-resolution images came from the Lunar Reconnaissance Orbiter’s Narrow Angle Camera, with almost 964,000 multispectral images from its Wide Angle Camera.
The collection also incorporates measurements from missions including GRAIL, Lunar Prospector and Japan’s Kaguya/SELENE. A foundation model is pretrained on large amounts of mostly unlabeled material, then adapted to a specific problem with a smaller labeled set. That approach fits planetary science, where observations are plentiful but expert annotations are expensive.
The reported results include lower error when predicting polar ice deposits and stronger crater detection. These are task-specific evaluations, not proof that one model can answer every lunar-science question. Its practical value will depend on how independent teams reproduce those gains and fine-tune it for different instruments, regions and scientific objectives.