NASA and IBM open source lunar AI model

NASA and IBM have released the Lunar Foundation Model as open source software, giving scientists and mission teams a shared AI system trained on geospatial observations of the Moon. Built as a multimodal model based on a visual transformer, it can connect imagery with other lunar data such as topography, temperature and lighting conditions, reducing the need to build separate algorithms for each research question.

One major use case is the search for water ice near the lunar poles, where permanently shadowed regions may preserve ancient deposits. The model combines surface temperature, ice stability depth, terrain morphology and shadow maps to identify promising areas for follow-up investigation. In team tests, the method reduced prediction error by 23% compared to a reference model, though it does not detect ice directly beneath the surface.

The model can also support crater detection, mission planning and studies of volcanic features known as Irregular Mare Patches. It can outline craters with a resolution down to about a meter, and at a scale of roughly 100 meters its performance was nearly 19% better than the baseline. IBM and NASA have made the model and related data available through Hugging Face, aiming to broaden testing and accelerate lunar research.