NASA and IBM Release Open-Source Lunar AI Model for Scientific Mapping

NASA and IBM have released an open-source AI model that researchers can adapt for mapping lunar craters, studying volcanic features and estimating where polar ice may be stable. The NASA-IBM Lunar Foundation Model comes with public code, datasets and benchmarks, creating a shared starting point for lunar-science analysis.

A foundation model is trained broadly before being adapted to a narrower task with relatively small amounts of labeled data. NASA says the approach could spare planetary scientists from building a specialized model from scratch for each new question. Its core training material came from the Lunar Reconnaissance Orbiter, whose 17 years of observations cover most of the Moon’s surface in detail.

One training base, several views of the Moon

The model was trained on roughly 2 million image tiles. The set included more than 1 million high-resolution camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. It also drew on imagery and terrain data from GRAIL, Lunar Prospector and Japan’s SELENE mission.

Roughly 2 millionTraining image tiles

NASA says the model was trained on roughly 2 million lunar image tiles.

More than 30Aligned dataset layers

IBM says the accompanying dataset combines more than 30 spatially aligned layers from nine instruments across four missions.

The open package is more than model weights

The model is hosted on Hugging Face, and its complete codebase is available on GitHub. NASA also released machine-learning-ready pretraining datasets and benchmark collections, while the model is integrated into the open-source TerraTorch toolkit. IBM says the unified dataset puts observations from different instruments and resolutions into one machine-learning-ready framework.

Three research targets

• Crater mapping: Researchers can adapt the model to identify and measure impact craters, which help date lunar terrain.
• Volcanic features: It can be fine-tuned to find irregular mare patches, unusual formations relevant to the Moon’s thermal history.
• Polar ice: It can estimate where ice may be stable near the lunar poles, including on or below the surface.

Reported results favor ice-stability estimates

NASA says the model matched or exceeded several strong baselines across its evaluated tasks, with its clearest advantage in estimating polar ice stability. IBM cites a NASA-IBM technical paper that reported up to a 22% reduction in error for identifying areas with high ice potential against the SwinV2-B ImageNet model. For crater identification at about 100-meter context-scale resolution, the cited paper reported nearly 19% better performance than SwinV2-B using half the training data.

NASA also demonstrated fine-tuning the system to spot a new impact crater in an image excluded from pretraining. The agency notes that changing orbital lighting can make smaller craters harder to see, a practical limit for surface-change detection.