On September 10, NASA and IBM Research released the NASA-IBM Lunar Foundation Model, with weights hosted on Hugging Face and code on GitHub. The Lunar Reconnaissance Orbiter (LRO) has circled the Moon for 17 years and produced more data than all of NASA's other planetary missions combined. The model was trained on approximately 2 million image patches—over 1 million at 1-meter resolution from the Narrow Angle Camera and nearly 964,000 at 100-meter multispectral resolution. These are the first open-source foundation models built specifically for lunar science.
The data-alignment problem
NASA's lunar archives come from multiple missions and instruments with incompatible formats, resolutions, and physical units. GRAIL mapped the Moon's interior at roughly 20 km per pixel; LRO's cameras resolve individual boulders and crater rims at 1 meter. Aligning nine instruments across four missions (LRO, GRAIL, Lunar Prospector, and JAXA's SELENE/Kaguya) previously could consume months of a PhD student's time.
The solution was a dataset: SomBench, aggregating 30+ spatially aligned data layers into ~2 million registered patches covering the lunar surface and subsurface. The layers include imagery, elevation, UV reflectance, thermal behavior, mineral composition, and gravity anomalies.
Training with shadows in mind
The model uses masked-token learning: patches are partially masked and the model predicts the missing content. The key difference is conditional input—solar elevation angle and spacecraft position are fed in as conditions so the model learns to separate terrain from illumination-driven shadows. On the Moon, with no atmospheric scattering, the contrast between permanently shadowed regions and noon highlands is extreme; without explicit lighting modeling, models tend to mistake shadows for terrain. A multi-scale design combines regional context (mare vs. highland) with fine detail (crater rim locations).
Benchmark results
| Task | Baseline | Result | |---|---|---| | Polar ice prospects (RMSE) | SwinV2-B | Up to 22% lower error | | Crater mapping (100 m context) | SwinV2-B | Nearly 19% better, with half the labels | | Crater mapping (meter scale) | Best current detector | Comparable | | Irregular Mare Patch segmentation | Specialized models | Comparable or better |
- Polar ice. Permanently shadowed craters near the south pole can reach −246 °C (−410 °F); assessing subsurface ice requires fusing thermal, radar, gravity, and topographic data. NASA lunar terrain expert Michael Barker said the model has internalized the underlying physics of the lunar surface.
- Craters. Both the Moon's most abundant surface feature and a key geological clock—the older the terrain, the more craters.
- Irregular Mare Patches (IMPs). Apparently young features that challenge existing lunar cooling timelines; a census could sharpen the Moon's thermal evolution history.
- Prithvi: pretrained on Earth-observation data for flood mapping and disaster monitoring
- Surya: a heliophysics model trained on solar observations for flare and space weather prediction
Out-of-distribution generalization
In one notable test, the model detected a newly formed impact crater from a SpaceX rocket-body strike, even though the post-impact image was not in the training data. This single result suggests the model learned general features of fresh craters rather than memorizing images—useful for long-term lunar monitoring, though one data point cannot support stronger conclusions.
Open release: weights, code, and data
Beyond weights and code, the team released the machine-learning-ready pretraining dataset and benchmark suites, integrated the model into the open-source TerraTorch toolkit, and published a companion paper on Hugging Face. Reproducibility requires that downstream researchers can compare fine-tunes on the same data—weights without benchmarks would leave the model a black box. NASA Chief Science Data Officer Kevin Murphy noted that collecting data is only part of the job; scientists also need easier ways to explore and use it.
What it can and cannot do
NASA is explicit: the model is a pattern-recognition assistant, not a substitute for physical measurement. It compresses the time spent finding features, not the time spent interpreting them. Geological interpretation of crater counts, drill validation of ice maps, and thermal-evolution inference for IMPs still require humans. Label scarcity on the lunar surface is another constraint—which is why achieving better results with half the annotations is a highlighted result.
Part of a family
The lunar model belongs to NASA's "AI for Science & Research" strategy under the Office of the Chief Science Data Officer. The same NASA-IBM collaboration previously produced:
Why now
Artemis plans to send astronauts near the lunar south pole, where water ice—if usable—could become drinking water, breathable oxygen, and rocket fuel. NASA estimates the lunar north pole may hold 600 million metric tons of water ice, with IBM Research noting the south pole is expected to hold comparable or higher concentrations. These are estimates, not measured reserves. The model's real contribution is reducing the time cost of going from archive to usable product, freeing researchers to focus on hypothesis testing rather than scanning petabytes of imagery.
References
1. NASA Science, "NASA, IBM Launch AI Foundation Model for Lunar Science," 2026-09-10 — https://science.nasa.gov/science-research/artificial-intelligence-lunar-foundation-model/ 2. Space.com — https://www.space.com/astronomy/moon/nasa-ibm-launch-new-ai-model-for-studying-the-moon 3. Tech Times, 2026-09-11 — https://techtimes.com/articles/327302/20260911/nasa-ibm-open-source-lunar-ai-that-pinpoints-moon-ice-22-more-accurately.htm 4. Science Times, 2026-09-11 — https://www.sciencetimes.com/articles/62585/20260911/nasa-ibm-unveil-new-ai-model-designed-map-moon-greater-detail.htm 5. Complete AI Training — https://completeaitraining.com/news/nasa-releases-open-source-ai-model-trained-on-lunar-orbiter