NASA and IBM Release Open-Source Lunar Foundation Model Trained on 30 Data Layers

NASA and IBM have released an open-source Lunar Foundation Model designed to pull decades of lunar observations into one system and make the Moon easier to map, study and plan for. The model, its unified dataset and related code are being made publicly available through Hugging Face and GitHub, so universities, space organisations and startups can build on it instead of developing a lunar AI system from scratch.

The scale of the data is the headline. The model was trained on more than 30 layers of observations collected by nine instruments across four lunar missions, including NASA's Lunar Reconnaissance Orbiter and the Gravity Recovery and Interior Laboratory, or GRAIL. Those measurements previously sat in separate missions and instruments, and researchers often had to work with one dataset in isolation; folding them into a single multimodal system makes cross-source analysis of surface features possible.

On capabilities, the model analyses several kinds of lunar information at once. It can identify craters, map volcanic features and assess areas that may hold deposits of ice. Lunar ice is more than a scientific curiosity: water can be split into hydrogen and oxygen, which ties directly to exploration and resource planning. Permanently shadowed regions near the poles are the most promising places to look because extremely low temperatures can preserve ice, but they are hard to study with conventional observations, so the model combines thermal, topographic and multispectral information to flag terrain with higher potential.

The benchmark results are the most concrete part of the release. Against the SwinV2 baseline, the model reduced root mean square error by 22% on ice prospectivity, improved intersection over union by 3% when mapping irregular mare patches, and lifted accuracy and mean average precision by 19% on crater detection at 100-metre resolution, while using only half the training data.

That last result matters for organisations without large computing budgets: it suggests useful information can be extracted without an equally large volume of labelled data. For universities, smaller space organisations and startup teams, that lowers the cost of building lunar AI capability of their own.

Juan Bernabe-Moreno, director of IBM Research Europe, compared the process to understanding unfamiliar terrain before travelling through it. The analogy captures the model's role: the aim is not another static map of the Moon but a way for missions to decide where to point instruments, robotic exploration and landing studies.

The model also reaches beyond resource prospecting. Crater detection helps researchers read the Moon's geological history and supplies terrain information relevant to landing-site selection, while mapping irregular mare patches connects to the Moon's volcanic past. Tasks that once needed dedicated models are now folded into a shared representation.

Methodologically, the release extends the open-science approach associated with the Prithvi family of foundation models from Earth observation to lunar exploration. As NASA's Artemis programme pushes toward sustained lunar exploration, turning existing observations into usable maps and scientific insight is becoming as important as reaching the surface. The next thing to watch is which problems outside the original evaluation the open-source community adapts the model to.