NASA and IBM's lunar model, between uncalibrated estimates and operational limits
On 10 September 2026, NASA and IBM Research announced the open source release of the NASA-IBM Lunar Foundation Model, developed together with institutions including the Universities Space Research Association, the SETI Institute, the University of Maryland Baltimore County and Howard University. The model was trained on SomBench, a dataset hosted on AWS under a CC BY 4.0 licence and made up of roughly 2 million bundles of co-registered lunar tiles. The underlying material draws on 17 years of Lunar Reconnaissance Orbiter data — including 38 TB of Wide Angle Camera imagery (963,609 tiles) and 1.4 TB from the Narrow Angle Camera (1,000,113 tiles) — alongside surveys from the GRAIL and Lunar Prospector missions and JAXA's SELENE/Kaguya probe. The weights sit on Hugging Face under the nasa-ibm-ai4science organisation, while the code, built on the TerraTorch toolkit, lives in the nasa-impact GitHub repository; both carry an Apache-2.0 licence. Commenting on the initiative, NASA's Chief Science Data Officer, Kevin Murphy, said: “The NASA-IBM Lunar Foundation Model shows what's possible when we bring AI to NASA's petabytes of scientific data”.
The model uses an encoder-decoder ViT-B architecture (dimension 768, 12 layers, 12 heads), but the documentation gives neither the exact parameter count nor the compute hours involved. In the benchmarks reported in the release materials and analysed by Unite.AI, the polar ice prospectivity estimate records an RMSE of 0.0293 against 0.0377 for the SwinV2-B baseline, an error reduction of about 22%. On crater detection at a context scale of roughly 100 metres the system beat the baseline by almost 19%, using around half the training labels. NASA summed the results up more cautiously, saying the model had “matched or exceeded” existing references. The technical report is described as a companion paper and no peer review is indicated; no independent third-party evaluations have appeared so far.
The usage limits declared by the developers themselves, however, draw a sharp line between data analysis and exploration in the field. The official model card states explicitly that the system “is not a scientific-grade generative product” and specifies that “Generated fields are not calibrated predictions and are no substitute for instruments”. The model produces a fuzzy overlap map for polar ice, which the model card describes as uncalibrated — not measured ice. It has no absolute geodetic reference system, and the dataset carries a geolocation uncertainty on the order of tens of metres. As of 12 September 2026, the initiative's GitHub repository shows 3 commits in total, while the Hugging Face page reports 120 downloads over the past month.
Turning petabytes of orbital archives into a visual model is a useful exercise in synthesis for steering scientific research, as long as statistical correlations are not mistaken for the consistency of the ground. When it comes to certifying landing sites, even the creators of the AI remind us that an algorithm is no replacement for a calibrated sensor.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official announcement on science.nasa.gov (date, partners, dataset, the Kevin Murphy quote, the cautious wording of the results), the model card on Hugging Face (Apache-2.0 licence, ViT-B architecture, usage limits, downloads) and the nasa-impact repository on GitHub (licence, links to the weights, 3 commits). As independent confirmation, the Unite.AI article with the precise benchmark figures and the breakdown of SomBench; the release is also confirmed by The Next Web and Technode Global. I separately checked that the ibm-nasa-geospatial organisation on Hugging Face does not list this model: I put that among the open questions rather than pass over it. No rumours, no leaks — everything here comes from an institutional announcement, a model card or a public repository.
- Incertezze
- Parameter count and compute hours have not been published: the architecture is stated (ViT-B), the exact count is not. The precise benchmark numbers (19% on craters, RMSE 0.0293 against 0.0377 on ice, 3% on IMPs) come from the release documentation as reported by Unite.AI; NASA summarises them qualitatively, and no independent evaluations exist. The model card cites a reference dated 15 April 2025 (arXiv 2504.11171) that clearly points to earlier work (TerraTorch/Prithvi): I could not open the paper to confirm, so I do not use that date. The model sits under the nasa-ibm-ai4science organisation, not under ibm-nasa-geospatial, which hosts the Prithvi family. The technical report is a companion paper, with no peer review indicated.
- Perché pubblicarla
- This is an open source release by a public agency — not a company talking about itself — with Apache-2.0 weights, a CC BY 4.0 dataset and code that can be checked today. It matters twice over: foundation models are moving from Earth observation to other celestial bodies, and this is a textbook case of limits stated honestly, since the official model card says in plain words that the ice maps are not measured ice and cannot certify a landing site. The gap between the marketing numbers doing the rounds (“up to 23% better”) and NASA's own phrasing (“matched or exceeded”) is exactly the distinction this site exists to draw.