Finding water ice on the Moon is not a treasure hunt. It is a survival calculation for every crewed mission NASA plans to land there, determining where astronauts live, where rovers drive, and whether in-situ water becomes viable fuel and a drinking supply.
NASA and IBM released the Lunar Foundation Model on September 10, an open-source AI system trained on more than two million co-registered data points drawn from three lunar missions, now publicly available on Hugging Face.
One Model, Decades of Lunar Data
A single AI backbone now unifies imagery, gravity, topography, and thermal data from three separate lunar missions into one machine-learning-ready dataset.
The model ingests high-resolution imagery, topography, gravity measurements, and thermal readings simultaneously. All data points are spatially aligned across NASA’s Lunar Reconnaissance Orbiter, the Gravity Recovery and Interior Laboratory mission, and Japan’s SELENE/Kaguya spacecraft.
Every pixel of imaging corresponds to every other data type in the same grid location. That alignment gives the AI a unified picture of the surface rather than disconnected instrument snapshots.
IBM built the system as a Vision Transformer encoder-decoder trained from scratch on this co-registered dataset. It operates at both meter-scale resolution for fine surface detail and roughly 100-meter context scale for broader geological patterns.
Why Training an AI on the Moon Is Harder Than It Looks
Lunar lighting and near-identical craters forced IBM to rethink standard computer-vision training from the ground up.
Earth-observation models benefit from diffuse sunlight that softens shadows and preserves detail across a scene. The Moon has no atmosphere, and shadows there are, in Dr. Juan BernabĂ©-Moreno’s words, “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 the 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,” according to BernabĂ©-Moreno.
IBM’s solution was spatial partitioning: slice the Moon into sections and keep training zones entirely separate from testing zones.
“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.” , Dr. Juan BernabĂ©-Moreno, Director of IBM Research Europe, UK and Ireland
What the Benchmarks Show
The model outperforms a leading Microsoft-trained vision baseline on ice mapping and crater detection, and it passed an unplanned real-world test in August.
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.
A real-world test arrived unexpectedly 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.
The Dataset Outlasts the Model
NASA and IBM argue that the co-registered dataset, not the model itself, is the most durable contribution of this release.
Models get superseded. Datasets, when built carefully, become infrastructure. Bernabé-Moreno made that distinction explicit.
“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.” , Dr. Juan BernabĂ©-Moreno
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 without reconstructing data pipelines from scratch.
What This Means for Artemis
The model is designed to support the precise terrain and resource mapping that upcoming Artemis missions will require.
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.
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. That accessibility positions the release as a shared foundation for lunar science well beyond any single mission.
