Reading the Moon with AI: NASA and IBM’s new model hunts for hidden ice

NASA and IBM have released a free, open-source artificial-intelligence model built to study the Moon, the NASA-IBM Lunar Foundation Model. Trained on nearly two decades of orbital data, it can point to where ice may be hiding in the Moon’s darkest craters, map craters to help choose safe landing sites, and flag volcanic terrain, just as the United States prepares to send humans back to the surface. But there is an important caveat about what such a tool can, and cannot, actually tell us.

A model trained on the Moon

The tool is a foundation model, the same broad, general-purpose kind of AI behind modern language systems, except trained on lunar data instead of text. It ingested more than 30 layers of measurements gathered by nine instruments across four NASA missions, including the long-running Lunar Reconnaissance Orbiter, which has been circling the Moon for some 17 years. It joins IBM and NASA’s Prithvi family of open models, which already spans Earth’s geography, weather and space weather, and its trained weights are free for anyone to download.

Finding ice in the dark

The headline task is hunting for water ice near the lunar south pole. There, some crater floors are so deep that sunlight never reaches them; these permanently shadowed regions can sit at around minus 246 degrees Celsius, cold enough to trap ice for billions of years. That is precisely what makes them so hard to study, since ordinary cameras see only blackness. So the model fuses other clues, thermal readings, radar returns, gravity data and topography, to estimate where ice is most likely to lurk, both on the surface and below it. In the developers’ own benchmark tests, it flagged high-potential ice areas with roughly a fifth less error than the best existing tools, and mapped craters more accurately while using only half the training data.

What it can and cannot say

Here rigor matters. The model does not detect ice; it predicts where ice is likely, and where such deposits would remain stable. Those are informed estimates, not confirmations of anything physically present. Their real value is to narrow the search, telling scientists and mission planners where to aim rovers, instruments and, eventually, astronauts, instead of sifting through maps and images by hand. Confirming the ice itself still requires going there and looking.

Why the Moon’s ice is worth the hunt

The stakes explain the effort. Water on the Moon would be a transformative resource: it can be split into hydrogen and oxygen for breathable air and rocket fuel, which means a future base might not have to haul everything up from Earth. The same model also speeds up crater mapping, a chore that does double duty, since counting and measuring craters both dates lunar terrain, and by extension helps reconstruct the history of the inner solar system, and reveals which slopes are safe for a lander. By automating work that once took researchers weeks or months, the model is less a discovery in itself than a new pair of eyes, one that could help decide where humanity next sets foot on the Moon.

© 2026 Homer Dávila Gutiérrez, FRAS — SKYCR.ORG. All rights reserved. Source: NASA / IBM Research — NASA-IBM Lunar Foundation Model (2026); model weights on Hugging Face, code on GitHub. Trained on data from NASA’s Lunar Reconnaissance Orbiter and other missions.