NASA and IBM Research have open-sourced the NASA-IBM Lunar Foundation Model, the first open-source foundation model built specifically for lunar science. The weights are on Hugging Face and the code is on GitHub (released September 10). The training corpus comprises roughly 2 million image tiles—more than 1 million at 1-meter resolution from LRO's Narrow Angle Camera, and nearly 964,000 multispectral tiles at 100-meter resolution.
The Lunar Reconnaissance Orbiter has circled the Moon for 17 years, producing more data than all other NASA planetary missions combined.
The "data exists but can't be used" problem
NASA's lunar archive spans multiple missions and instruments with mismatched formats, resolutions, and physical units. GRAIL mapped the Moon's interior structure at roughly 20 km per pixel; LRO's cameras can resolve individual boulders and crater rims at 1 meter. Aligning these datasets used to require months of manual registration and unit conversion by researchers.
IBM and NASA's solution starts with the dataset: SomBench, which aggregates more than 30 spatially aligned data layers from nine instruments into about 2 million registered tiles, covering diversity of the lunar surface and subsurface. The layers include imagery, elevation, UV reflectance, thermal behavior, mineral composition, and gravity anomalies, drawing on LRO, GRAIL, Lunar Prospector, and JAXA's SELENE/Kaguya missions.
Learning to ignore shadows
The model uses masked-token learning: parts of each tile are hidden and the model predicts them. The key difference is conditioning—solar elevation angle and spacecraft position are fed as conditional inputs, so the model learns to separate terrain itself from illumination-caused shadows. On the Moon, with no atmospheric scattering, the contrast between permanently shadowed regions and noon highlands is extreme; without explicit illumination modeling, models easily confuse shadow for terrain.
A multi-scale design combines regional context with fine detail, letting the model see both "is this mare or highland?" and "where is this crater rim?"
Three benchmarks, three numbers
All three initial science targets connect directly to Artemis landing-site planning:
1. Polar ice prospecting. Permanently shadowed regions near the lunar south pole reach surface temperatures as low as −246 °C (−410 °F). Judging subsurface ice requires fusing thermal readings, radar returns, gravity, and topography. The model reduced RMSE by up to 22% versus the SwinV2-B baseline (an ImageNet-pretrained transformer commonly used as a reference). 2. Crater mapping. Craters are the Moon's most abundant surface features, the most dangerous landing hazards, and a geological clock. At meter-scale resolution the model matches the best existing crater detectors; in 100-meter contexts it beats the same baseline by nearly 19% while using only about half the labeled training data. NASA lunar terrain expert Michael Barker says the model has internalized the underlying physics of the lunar surface. 3. Irregular Mare Patches (IMP). These apparently young structures challenge existing lunar cooling timelines; cataloging them helps build a more accurate thermal evolution history.
| Task | Baseline | Result | |---|---|---| | Polar ice prospecting (RMSE) | SwinV2-B | Up to 22% error reduction | | Crater mapping (100 m context) | SwinV2-B | ~19% better, half the labels | | Crater mapping (meter-scale) | Best existing detector | Comparable | | IMP segmentation | Dedicated models | Comparable or better |
An unseen crater
In one notable test, the model detected a newly formed impact crater from a SpaceX rocket-body impact—even though the post-impact image was not in the pretraining data. This is a small sample of out-of-distribution generalization: the model appears to have learned general features of what a fresh crater looks like, rather than memorizing images. Such change detection could be valuable for long-term lunar monitoring, though a single result supports only limited conclusions.
Open means dataset included
The release goes beyond weights and code: an ML-ready pretraining dataset, benchmark collections, integration into the open-source TerraTorch toolkit, and an accompanying paper on Hugging Face. This matters because foundation models require that downstream researchers can reproduce and compare results on the same data—weights alone without benchmarks would be a black-box API.
NASA Chief Science Data Officer Kevin Murphy noted that NASA has spent decades building a remarkable lunar science record, but 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 that the model does not replace direct measurement; it 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 behind crater counts, drilling validation behind ice prospecting maps, and thermal evolution inference behind IMP studies all still require humans. A second limitation is labels: ground-truth labels for lunar surface features are scarce—which is exactly why the "better results with half the labels" result is significant.
Part of a family, and timely
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 has already produced:
- Prithvi: pretrained on Earth observation data for flood mapping and disaster monitoring.
- Surya: a heliophysics model trained on solar observations for predicting flares and other space weather events.
The timing is tied to Artemis, which will send astronauts near the lunar south pole—where water ice may exist as drinking water, breathable oxygen, and rocket fuel. NASA estimates the lunar north pole may hold 600 million metric tons of water ice; IBM Research's blog notes the south pole is expected to contain comparable or higher concentrations. These are estimates, not measured reserves.
The model's real contribution is cutting the time cost of going from archive to usable data—not changing scientific conclusions. It frees researchers from paging through petabyte-scale imagery and lets them spend time on hypothesis testing instead.
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, "NASA, IBM launch new AI model for studying the moon" — https://www.space.com/astronomy/moon/nasa-ibm-launch-new-ai-model-for-studying-the-moon 3. Tech Times, "NASA and IBM Open-Source Lunar AI That Pinpoints Moon Ice 22% More Accurately," 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, "NASA and IBM Unveil New AI Model Designed to Map the Moon in Greater Detail," 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, "NASA releases open-source AI model trained on lunar orbiter data" — https://completeaitraining.com/news/nasa-releases-open-source-ai-model-trained-on-lunar-orbiter