IBM and NASA open-source a lunar foundation model: up to 23% better at spotting ice, volcanoes and craters, with the dataset released too

Start with the three specific metrics in the official release, because that is where the value of this story sits. First, potential lunar ice: permanently shadowed regions are among the Moon's most difficult environments to observe, yet they may hold lunar ice below the surface. Ice means water and oxygen — resources considered essential for a future Moon base and for producing rocket fuel for missions to Mars. A technical paper authored by IBM and NASA researchers shows the model reduced error (RMSE) in identifying areas with high potential for lunar ice by up to 22% compared with the SwinV2-B (ImageNet) model. Second, volcanic history: scientists study lunar volcanic features called Irregular Mare Patches to understand the Moon's volcanic history and thermal evolution, and detecting these changing regions is also strategic for future surface operations. Using imperfect labels, the model captures the extent of volcanic features 3% better than SwinV2-B, with comparable accuracy, greater efficiency and lower fine-tuning costs. Third, crater detection: the model can identify, contextualise and classify craters at metre-scale resolution with accuracy comparable to state-of-the-art models like SwinV2-B while offering greater efficiency and lower fine-tuning costs. At context-scale resolution (~100 m) it outperforms SwinV2-B by nearly 19% using just half the training data.

Why this is hard in itself is stated plainly in IBM's release: despite the wealth of lunar data, no publicly available, unified dataset existed that brings multi-modal, multi-resolution data into a common framework suitable for modern machine learning. So they open-sourced the dataset too. It is the first open-source lunar dataset of its kind, aggregating over 30 spatially aligned layers from nine instruments across four missions — tens of thousands of images and maps showing unique geophysical properties of the lunar surface. The sources include NASA's Lunar Reconnaissance Orbiter (LRO) and the GRAIL mission, with complementary lunar data from the Japan Aerospace Exploration Agency's SELENE/Kaguya for a rich, multi-modal view of the lunar surface and subsurface. That detail is especially relevant to anyone working with data: model performance gains often come not from a more complicated architecture but from genuinely aligning layers that were scattered across different coordinates, resolutions and instrument conventions onto one spatial reference. Getting 30-plus layers spatially aligned is serious upstream engineering.

There is one line about methodology worth quoting on its own. Juan Bernabe-Moreno, Director of IBM Research Europe, UK and Ireland, said the model gives scientists a foundation to explore the Moon at scale — connecting observations across instruments, revealing patterns that are difficult to see in isolation, and providing an open platform the global research community can build on. Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters, points at the more general problem: NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job, and the data also has to become easier for scientists to explore and use. That sentence explains the real value of this class of foundation model for research institutions — not replacing scientists' judgment with AI, but lowering the entry barrier to cross-instrument data. The operational significance is explicit too: crater mapping helps NASA select safe landing sites, avoid hazards such as steep slopes and boulders, and plan locations for long-term lunar infrastructure, all of which are concrete requirements under the Artemis programme.

Finally, the positioning: this model joins the Prithvi family of open foundation models, which until now spanned geospatial, weather and heliophysics, and now the Moon. IBM's framing of the family is that instead of building a new algorithmic system for every scientific question, researchers can start from a shared model and adapt it to new tasks to accelerate discovery across domains. That logic will be familiar to developers — it is the same move software engineering made from rebuilding wheels to reusing shared components. The difference is that in science the data is far more heterogeneous and the labels are far worse than in most commercial settings, so whether a shared model transfers is a real question rather than a slogan. The technical paper's approach here — training with imperfect labels and reporting error rates — is more useful than a single flattering number. The model, the dataset and the technical report are all directly available on Hugging Face, which makes reproduction straightforward, and for an open-source community reproducibility matters more than a claimed benchmark.