Lunar Foundation Model: How NASA and IBM Are Using AI to Unlock the Moon’s Secrets

The Moon has been watched from above for decades, generating an enormous archive of images, maps and scientific measurements. Now, artificial intelligence is giving scientists a new way to make sense of that information.

NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, an open-source AI model designed specifically for lunar science. Trained primarily on 17 years of observations from NASA’s Lunar Reconnaissance Orbiter (LRO), the model brings together different types of lunar data to help researchers study the Moon’s surface more efficiently.

What Is the Lunar Foundation Model?

Unlike a conventional machine-learning system built for just one task, the Lunar Foundation Model is designed as a reusable scientific foundation. It learns patterns from large amounts of unlabeled lunar data and can then be adapted for different research applications with comparatively small amounts of task-specific data.

The model was trained using nearly 2 million co-registered lunar data bundles, covering 11 modalities and two spatial scales. The dataset combines high-resolution imagery, multispectral information, terrain and other measurements gathered from multiple lunar missions, giving researchers a more connected view of the Moon.

AI Could Make Mapping the Moon Much Faster

One of the biggest advantages of the NASA IBM Lunar Foundation Model is its ability to process information that would otherwise require scientists to examine huge numbers of maps and images manually.

The model can help identify and map lunar craters, including features that may not yet be catalogued. Since crater counts are important for estimating the age of lunar surfaces and understanding the history of the Solar System, faster crater mapping could give scientists more time to focus on interpreting what those features mean.

The technology can also combine information from different instruments rather than treating each observation separately. This multimodal approach allows researchers to look for relationships between terrain, imagery, illumination, temperature and other characteristics of the lunar environment.

Finding Ice Could Be One of Its Most Important Jobs

Perhaps the most exciting application of the lunar AI model is the search for water ice near the Moon’s poles.

Some permanently shadowed regions receive little or no direct sunlight and can remain extremely cold. Scientists believe these locations can preserve ice for very long periods. Identifying where ice is most likely to be stable could therefore become important not only for scientific research but also for future lunar exploration.

The model performed particularly well in tests involving potential lunar ice deposits. IBM and NASA reported that it reduced prediction error by up to 22% compared with a SwinV2-B baseline in their testing. Reuters separately reported that the model achieved improvements of up to 23% in identifying key lunar features compared with existing methods, depending on the benchmark.

Why does that matter? Water found on the Moon could potentially support astronauts and, after processing, provide oxygen and hydrogen for future spacecraft propulsion. That makes Moon ice detection an important part of planning for a more sustained human presence beyond Earth.

The Model Can Also Study Lunar Volcanism

The Moon may appear geologically quiet today, but its surface preserves evidence of a much more active past. Researchers are particularly interested in unusual formations called Irregular Mare Patches, which may provide clues about the Moon’s volcanic history and thermal evolution.

The Lunar Foundation Model can help identify and segment these formations across large areas of the lunar surface. By speeding up the process of locating unusual geological structures, AI could help scientists investigate questions about when and how the Moon’s interior cooled.

Why Combining Different Lunar Data Matters

The Moon does not look the same under every observation. Shadows, illumination angles, viewing geometry and differences in resolution can dramatically change how a surface feature appears.

The NASA-IBM model addresses this by incorporating observation geometry and by learning from both metre-scale and broader 100-metre-scale information. The approach helps the model connect tiny surface details with larger geological patterns instead of examining each observation in isolation.

This is particularly useful because lunar missions have collected data using instruments with very different capabilities. LRO, GRAIL, Lunar Prospector and Japan’s Kaguya/SELENE mission all contributed complementary information to the broader dataset used in the project.

Open Source Could Accelerate Moon Science

Another important part of the project is that it is open source. NASA and IBM have made the model, machine-learning-ready datasets and benchmark resources available to the wider research community, with the model integrated into the open-source TerraTorch toolkit.

That means researchers outside NASA and IBM can experiment with the model, adapt it to new lunar science problems and compare their results with existing benchmarks. It could also help researchers develop specialised tools without having to build an entire lunar AI system from the beginning.

From Studying the Moon to Preparing for Future Missions

The development of the NASA-IBM Lunar Foundation Model comes as space agencies increasingly look beyond simply visiting the Moon. NASA's Artemis programme is focused on returning humans to the lunar surface and developing technologies for longer-term exploration.

Better maps and a deeper understanding of lunar resources could become increasingly important as future missions consider where astronauts should land, what areas are scientifically valuable and where resources such as water ice may be found.

The model is not a replacement for scientific instruments or direct measurements. Instead, it works as a powerful analysis tool that can help researchers navigate the enormous amount of information already collected about the Moon.

A New Era for AI and Moon Science

The Lunar Foundation Model represents a shift in how scientists can approach lunar data. Instead of building a separate AI system every time they want to identify a crater, investigate volcanic terrain or search for ice, researchers can start with a common model and adapt it to the problem at hand.

The Moon has already given humanity generations of scientific discoveries. With NASA and IBM bringing AI into that exploration, the next breakthroughs may come not only from new missions, but also from finding new meaning in the data we have already collected.

As humanity prepares for a new era of lunar exploration, the ability to understand the Moon faster and more comprehensively could prove just as important as the rockets that take us there.