NASA and IBM have released an open-source artificial-intelligence model designed to analyze the Moon’s surface and help researchers identify features including water ice, impact craters and volcanic terrain. The Lunar Foundation Model brings together observations that are normally studied separately, giving scientists a common system that can be adapted to several mapping tasks.
The model was trained on more than 30 data layers collected by nine instruments across four NASA missions, according to IBM and reporting by Reuters. Those inputs include imagery and measurements from the Lunar Reconnaissance Orbiter, which has surveyed the Moon since 2009 and produced a detailed record of topography, temperature and surface composition.
Foundation models learn broad patterns from large datasets before being tuned for a more specific task. In this case, researchers can adapt the same underlying model to classify terrain, detect craters or examine regions where permanently shadowed conditions may preserve water ice. IBM said tests showed gains of up to 23 percent in accuracy compared with existing approaches, although performance will vary by dataset and application.
The open-source release is strategically significant because lunar data is abundant but difficult to combine. Instruments observe at different resolutions, wavelengths and times, and many research groups lack the computing resources to train a large model from the beginning. A reusable foundation model can reduce that barrier and make experiments easier to reproduce.
Water ice is among the most important targets. Deposits near the lunar poles could support crews and, if processed, provide oxygen and hydrogen. They are also scientifically valuable records of how volatile material reached and evolved on the Moon. Reliable maps are therefore relevant to mission safety, landing-site selection and the longer-term debate over access to lunar resources.
Crater detection provides another test. The number and distribution of craters help scientists estimate the relative age of surface regions, but manual counting is slow and can produce inconsistent results. Automated systems can process larger areas, leaving researchers to examine uncertain or scientifically unusual cases.
The model also supports analysis of volcanic structures that reveal how the Moon cooled and differentiated. Recognizing subtle features across large datasets could guide follow-up observations or identify locations that deserve closer study by robotic and crewed missions.
NASA and IBM have previously worked on the Prithvi family of geospatial foundation models for Earth observation. Extending that approach to the Moon shows how space agencies are using AI as a layer between raw remote-sensing data and scientific decisions, rather than treating it solely as an autonomous mission system.
The release comes as NASA prepares further Artemis missions and aims to return astronauts to the lunar surface. AI-generated maps will not replace instrument teams, geologists or mission assurance, and their outputs must be checked against the biases and limits of the training data. Their value lies in rapidly narrowing a vast search space.
By publishing the model openly, NASA and IBM are inviting universities, space agencies and commercial teams to test it against new questions. That broader scrutiny may prove as important as any single accuracy result, especially as competition around the lunar south pole makes transparent, comparable analysis increasingly consequential.
