On September 10, 2024, NASA and IBM released the Lunar Foundation Model on Hugging Face under an Apache‑2.0 license. The model is intended to simplify combining and interpreting data from different lunar instruments and modalities.
Training data and compute
The model was pretrained on roughly two million co‑registered lunar tiles covering 11 data modalities. Those tiles come from nine instruments across four missions. The teams report that the pretraining run consumed about 1,100 GPU‑hours.
Why this matters
Raw lunar data itself was not the scarcest resource; the harder problem was aligning, fusing and reading across datasets produced by instruments built for very different purposes. Building that cross‑instrument expertise required years of work. The Lunar Foundation Model encapsulates part of that workflow, allowing researchers to point a single model at the archive instead of manually sifting and aligning decades of maps.
Practical implications
The model can assist researchers searching for lunar ice in permanently shadowed regions or compiling crater catalogs, among other tasks. This does not render domain experts obsolete: human expertise remains necessary, but some of the previously scarce components of that expertise are now packaged in a downloadable model.
Licensing and access
Released under Apache‑2.0 on Hugging Face, the model is available for broad research and development use and integration.
Conclusion
By compressing parts of cross‑instrument data processing into a pretrained model, NASA and IBM have lowered a practical barrier to lunar research. The Moon itself is unchanged, but the cost and effort required to study it have been reduced.
