NASA and IBM have released the Lunar Foundation Model, an open-source artificial intelligence system trained on more than 2 million co-registered data points from three lunar missions, now available on Hugging Face. The model unifies imagery, gravity, topography, and thermal data into a single machine-learning-ready dataset designed to help scientists map lunar ice deposits and support crewed Artemis missions planning where astronauts will live and work on the Moon.

Why Finding Water Ice on the Moon Requires AI?

Locating water ice on the Moon is not a treasure hunt; it is a survival calculation for every crewed mission NASA plans. Water ice determines where astronauts establish habitats, where rovers can safely drive, and whether in-situ water extraction becomes viable as fuel and drinking supply. The Lunar Foundation Model was built to solve this exact problem by analyzing data from NASA's Lunar Reconnaissance Orbiter, the Gravity Recovery and Interior Laboratory mission, and Japan's SELENE/Kaguya spacecraft.

The model ingests high-resolution imagery, topography, gravity measurements, and thermal readings simultaneously. Every pixel of imaging corresponds to every other data type in the same grid location, giving the AI a unified picture of the lunar surface rather than disconnected instrument snapshots. This unified approach operates at both meter-scale resolution for fine surface detail and roughly 100-meter context scale for broader geological patterns.

What Made Training an AI on Lunar Data So Difficult?

Training a computer vision model on Earth is straightforward because our planet has diffuse sunlight that softens shadows and preserves detail across a scene. The Moon presents a radically different challenge. With no atmosphere, lunar shadows are, as IBM researcher Dr. Juan Bernabé-Moreno described them, "knife-edged and pitch black," carrying zero information and radically altering how a crater looks depending on the sun angle at imaging time.

Standard computer vision training compounds this problem. On the Moon, where thousands of craters look nearly identical from orbit, IBM found that asking a model to reconstruct masked image regions was "a complete disaster." The solution required rethinking the entire training approach from the ground up.

How IBM Solved the Lunar Training Challenge

• Spatial Partitioning Strategy: IBM divided the Moon into wedges, like an orange, and completely separated training zones from testing zones to prevent the model from memorizing crater patterns.
• Multi-Modal Data Integration: The model processes imagery, gravity data, topography, and thermal readings in a unified framework rather than treating each data source independently.
• Vision Transformer Architecture: IBM built the system as a Vision Transformer encoder-decoder trained from scratch on the co-registered dataset, enabling both fine-scale and broad-scale geological analysis.

"We took the Moon and divided it into wedges, like an orange, and we completely separated the training wedges from the testing wedges. This way, we could give the model a lot of consistency," explained Dr. Juan Bernabé-Moreno, Director of IBM Research Europe, UK and Ireland.

Dr. Juan Bernabé-Moreno, Director of IBM Research Europe, UK and Ireland

How Does the Model Perform Compared to Existing Systems?

The Lunar Foundation Model outperforms a leading Microsoft-trained vision baseline on ice mapping and crater detection. Against SwinV2-B, a widely used Microsoft-trained vision baseline, the Lunar Foundation Model reduced error on ice prospectivity mapping by approximately 22 to 23 percent. On crater detection at context-scale resolution, it outperformed SwinV2-B by roughly 19 percent using only half the training data.

The model passed an unplanned real-world test in August when a SpaceX Falcon 9 rocket stage impacted the Moon on August 5. IBM researchers fed an image of the impact site to the model, and it correctly identified the result as a new crater even though the impact closely overlapped an existing one, on the first attempt. This real-world validation demonstrates the model's ability to handle novel scenarios it was never explicitly trained on.

Why the Dataset Matters More Than the Model Itself?

NASA and IBM argue that the co-registered dataset, not the model itself, is the most durable contribution of this release. Models get superseded by newer versions over time, but datasets, when built carefully, become infrastructure that supports an entire ecosystem of research. The dataset contains more than 2 million spatially aligned data points from three separate lunar missions, eliminating the need for other research teams to reconstruct data pipelines from scratch.

"That alone is a massive scientific contribution, because put the model aside: the community now has a co-registered dataset with more than two million data points," noted Dr. Juan Bernabé-Moreno.

Dr. Juan Bernabé-Moreno, Director of IBM Research Europe, UK and Ireland

NASA and IBM describe the dataset as a platform for an entire industry of AI models. Research teams worldwide can use it to build and benchmark specialized tools for lunar science without reconstructing data pipelines from scratch. Both the model and dataset are open-source on Hugging Face, allowing planetary scientists to adapt them for lunar research tasks ranging from crater segmentation to ice prospectivity mapping.

How Will This Support Future Moon Missions?

NASA's Artemis program is preparing to send crews back to the Moon, with subsequent missions needing precise answers about where ice exists in permanently shadowed craters, which terrain is safe for landing, and where a habitat can be sited without geological risk. The Lunar Foundation Model is designed to support that planning work by providing accurate, AI-powered analysis of lunar surface conditions. This accessibility positions the release as a shared foundation for lunar science well beyond any single mission, enabling the entire planetary science community to contribute to humanity's return to the Moon.