NASA and IBM have released a free, open-source artificial intelligence model built to help scientists analyze the Moon's surface, including estimating where water ice could remain stable near the lunar poles. The NASA-IBM Lunar Foundation Model, announced Sept. 10, is available on Hugging Face, with its full code on GitHub. NASA describes it as among the first open-source AI models built specifically for lunar science.

The model's most practical use is resource scouting. IBM said lunar ice indicates the presence of water and oxygen, resources it described as essential for a future Moon base and for producing rocket fuel. Knowing where ice is likely to persist can help planners prioritize landing sites and instruments before crews arrive.

The release also adds to a family of NASA and IBM science models that already cover Earth observation and space weather, work that reaches everyday life through flood mapping, crop monitoring and solar flare forecasting.

Trained on Nearly Two Decades of Lunar Orbiter Data

According to NASA, the model was trained primarily on data from the Lunar Reconnaissance Orbiter, which has mapped most of the Moon in detail over 17 years and whose data volume is larger than that of all other NASA planetary missions combined. Training used roughly 2 million image tiles, including more than 1 million camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution.

The team also drew on imagery and terrain data from NASA's GRAIL and Lunar Prospector missions and from JAXA's SELENE, also called Kaguya. IBM said the researchers built a unified, machine learning-ready dataset of more than 30 spatially aligned layers from nine instruments across four missions, which the company described as the first open-source lunar dataset of its kind.

"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," said Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer. "We also have to make data easier for scientists to explore and use."

Foundation models are pre-trained on large amounts of unlabeled data, so researchers can adapt them to new tasks with relatively small labeled datasets. That can save time compared with building a new algorithm for every scientific question.

Polar Ice, Young Volcanic Features and Fresh Craters

Permanently shadowed regions near the lunar poles stay cold enough to trap and preserve ice for up to billions of years, according to NASA. The model can help estimate where ice patches are likely to be stable on and below the surface. NASA's comparison images of four locations near Mons Mouton, close to the south pole, show the model preserving many fine-scale patterns in a reference ice prospectivity map.

The model can also speed identification of irregular mare patches, unusual volcanic features that appear relatively young and challenge established timelines for how the Moon cooled. It can map craters, which scientists measure and count to date lunar surfaces and which NASA uses to avoid hazards when selecting landing sites.

In one test, the model highlighted a newly formed crater near Einstein Crater after a SpaceX rocket body impact, using a post-impact image that was excluded from pre-training. NASA said the approach could help scientists detect natural impacts and surface changes, though varying lighting conditions between orbits may affect how visible smaller craters are.

"The NASA-IBM Lunar Foundation Model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation," said Juan Bernabe-Moreno, director of IBM Research Europe for the U.K. and Ireland.

Evidence Check on the Performance Claims

The performance figures come from a technical report by the development team posted with the model. IBM said the model reduced error in identifying areas with high potential for lunar ice by up to 22 percent compared with a SwinV2-B model, and outperformed that model by nearly 19 percent at crater mapping at roughly 100-meter resolution while using half the training data. NASA said the model matched or exceeded several strong baseline models across all evaluated tasks, with a clear advantage in estimating polar ice stability and comparable results on crater mapping and irregular mare patch mapping.

Readers should weigh those results carefully. The report was written by IBM and NASA researchers, and neither organization has described it as independently peer-reviewed. An ice prospectivity map estimates where ice could be stable. It is not a detection of ice, and confirming deposits still requires orbital instruments, landers, or rovers. The model also inherits the limits of the data it was trained on.

Independent evaluations by lunar scientists outside the development team had not been published as of this report. Wider testing, which the open release is designed to allow, will show how well the model performs beyond its benchmarks.

Open Tools for Researchers, Students and Moon Watchers

The model is free to download, and the team also released pre-training datasets and benchmark collections, with the model integrated into the open-source TerraTorch toolkit. Using it still requires machine learning skills and computing resources, so the most immediate users are researchers, university teams, and advanced students. NASA's Impact AI team at Marshall Space Flight Center built the model with NASA's Goddard and Ames centers and a science team that included the Universities Space Research Association, the SETI Institute, the University of Maryland, Baltimore County, and Howard University.

NASA lists the related Prithvi models, which support Earth applications such as flood mapping, crop yield prediction and hurricane prediction, and the Surya model, which targets space weather such as solar flares that can disrupt power grids and satellite operations. Those tools are where AI science models most directly reach households, through better hazard maps and space weather warnings.

For anyone curious about the Moon itself, International Observe the Moon Night is Sept. 19, and NASA lists ways to take part. Readers do not need special software to follow the science, since NASA and IBM posted explanations and sample maps with the release.

What comes next depends on how researchers use the model. Nature World News will follow independent evaluations and any new ice prospectivity maps produced with it.

The bottom line: NASA and IBM have made a lunar mapping tool freely available, with developer-reported gains in estimating where ice could last near the Moon's poles. Those results still need outside testing, and an AI map of where ice might persist is a guide for exploration, not proof that ice is there.

What Readers Want to Know

What did NASA and IBM release? The NASA-IBM Lunar Foundation Model, a free, open-source AI model for analyzing the Moon's surface, was announced Sept. 10 and is available on Hugging Face and GitHub.

What data trained it? Mainly Lunar Reconnaissance Orbiter imagery, plus data from NASA's GRAIL and Lunar Prospector and JAXA's SELENE. IBM said the dataset includes more than 30 layers from nine instruments across four missions.

Can it find water ice on the Moon? It estimates where ice is likely to be stable, especially in permanently shadowed polar regions. It does not directly detect ice.

How accurate is it? IBM reported up to 22 percent lower error than a common baseline model on ice prospectivity, based on a technical report by the developers. Independent evaluations have not yet been published.

What else can it do? Map craters, identify young-looking volcanic features called irregular mare patches, and help spot new impact craters between observations.

Who can use it? Anyone can download it, but using it effectively requires machine learning experience and computing resources.

How does this connect to life on Earth? It belongs to a family of NASA and IBM models that includes Prithvi, used for flood and crop mapping, and Surya, aimed at forecasting solar flares that can affect power grids and satellites.