NASA and IBM have introduced an open-source Lunar Foundation Model designed to bring decades of lunar observations together and make the Moon easier to map, study and plan for future exploration.

The AI model has been trained on more than 30 layers of data collected by nine instruments across four lunar missions, including NASA’s Lunar Reconnaissance Orbiter and the Gravity Recovery and Interior Laboratory (GRAIL) missions. By combining different types of observations in one system, the model is intended to help researchers examine features across the lunar surface without treating each dataset in isolation.

A new way to read the lunar surface

Lunar exploration generates enormous amounts of data, but those observations come from different instruments and missions. Bringing them together can make it easier to study the Moon at a broader scale.

The NASA-IBM model uses a multimodal approach to analyse several kinds of lunar information at once. It can be used to identify craters, map volcanic features and assess areas that may contain deposits of ice.

Testing showed improvements over the SwinV2 baseline across several tasks. For ice prospectivity, the model reduced the root mean square error (RMSE) by 22%. Its mapping of irregular mare patches recorded a 3% improvement in intersection over union (IoU). For crater detection at 100-metre resolution, it achieved a 19% improvement in accuracy and mean average precision (mAP) while using half the training data.

That last result is particularly relevant for organisations that may not have access to large computing resources.

Why lunar ice matters

Ice on the Moon is more than a scientific curiosity. Water could potentially be separated into hydrogen and oxygen, making it relevant to future exploration and resource planning.

Areas near the lunar poles, particularly permanently shadowed regions, are considered important places to investigate because extremely low temperatures can preserve ice. However, these regions are difficult to study using conventional observations alone.

The Lunar Foundation Model brings together thermal, topographic and multispectral information to identify areas with greater potential for ice. Its reported 22% reduction in error on ice-prospectivity tasks could help researchers narrow down locations for closer investigation.

The aim is not simply to produce another lunar map. Better identification of potentially useful terrain could eventually help missions decide where to focus their instruments, robotic exploration and landing studies.

Mapping craters and the Moon’s volcanic past

The model also has applications beyond resource exploration.

Crater detection can help researchers understand the Moon’s geological history while providing information about terrain that may matter when selecting landing areas. Another task involves mapping irregular mare patches, geological formations associated with the Moon’s volcanic history.

At 100-metre resolution, the model delivered a 19% improvement over the SwinV2 baseline while using half the training data. That suggests the system could extract useful information without requiring the same volume of labelled training data.

IBM Research Europe Director Juan Bernabé-Moreno compared the process to understanding unfamiliar terrain before travelling through it.

The broader idea is straightforward: researchers and future astronauts can benefit from having a detailed understanding of the lunar landscape before missions reach the surface.

Built for an open research community

NASA and IBM are also making the Lunar Foundation Model and its unified dataset publicly available through Hugging Face and GitHub.

The open-source approach means researchers, universities, space organisations and startups can build on the model rather than developing a lunar AI system entirely from the beginning. It also gives scientists an opportunity to adapt the technology to questions beyond the tasks used in its initial evaluation.

The model follows the open-science approach associated with the Prithvi family of foundation models, extending that philosophy to lunar exploration.

AI could become part of the next phase of lunar exploration

The release comes as lunar exploration increasingly focuses not only on reaching the Moon but also on understanding where to land, what resources may be available and how surface conditions could affect future missions.

NASA’s Artemis programme is working toward sustained lunar exploration, making detailed surface information increasingly important for mission planning. An AI system capable of combining observations from multiple missions could give researchers another way to turn existing lunar data into usable maps and scientific insights.

For now, the Lunar Foundation Model is primarily a research tool. Its significance lies in bringing multiple forms of lunar data into a single open framework and making that capability available to a much wider research community.

If researchers can continue adapting it for new lunar questions, the same data already collected from the Moon could become useful in ways that were not possible when those observations were studied separately.

Also Read: Why De-dollarisation Could Make Gold the Standout Commodity in a Changing Trade System