NASA and IBM released a free AI that spots Moon water ice 22% more accurately

Finding water ice on the Moon is one of the hardest problems in planetary science. Ice is not evenly distributed, it hides in permanently shadowed craters at the poles, and identifying it from orbit requires distinguishing faint spectral signals from instrument noise. IBM and NASA’s Lunar Foundation Model is designed specifically for that problem.

The Lunar Foundation Model is an open-source AI trained on more than two million images from five Moon-orbiting spacecraft: the Lunar Reconnaissance Orbiter, GRAIL, Lunar Prospector, SELENE/Kaguya, and others. It is available on Hugging Face and GitHub through IBM’s TerraTorch geospatial framework, at no cost.

In benchmark testing, the model identifies water ice deposits 22% more accurately than previous methods. It also finds lunar craters 19% more precisely. Notably, it reaches that performance using half the labeled training data that comparable models require — an important characteristic for lunar science, where labeled datasets are expensive to produce.

The practical value is in resource mapping for the Artemis programme. NASA plans to establish a sustained human presence near the lunar south pole, where water ice could be converted to drinking water, oxygen, and rocket propellant. Knowing where the ice actually is, and how much of it exists, directly shapes landing site selection and infrastructure planning.

The model is released under an open licence because both IBM and NASA contribute to it from publicly funded research. Making it freely available means scientists at universities, space agencies, and private companies can use, evaluate, and improve it without licensing friction. TerraTorch, the IBM framework it runs on, handles the underlying geospatial data pipeline.

The 22% improvement figure is a benchmark result — a comparison against other models on a standardised test dataset, not a measurement from a live operational mission. Whether the Lunar Foundation Model will improve actual mission decision-making depends on how well benchmark conditions reflect real lunar observation data, which varies with orbital geometry, instrument calibration, and solar illumination angles. Independent validation by groups not affiliated with IBM or NASA has not yet been published.

The Lunar Foundation Model is available on Hugging Face and GitHub. IBM and NASA have not set a timeline for integrating the model into active Artemis mission planning.

Tags: artificial intelligence, nasa, open-source AI, IBM, Lunar Foundation Model, moon exploration