NASA released the NASA-IBM Lunar Foundation Model, one of the first open-source AI models built specifically for lunar science, through a collaboration with IBM Research and several academic institutions. The model, trained primarily on 17 years of data from NASA's Lunar Reconnaissance Orbiter (LRO), is publicly available on Hugging Face with its complete codebase on GitHub, giving planetary scientists a tool to analyze the Moon's surface faster than manual methods allow.
"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," said Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters in Washington. "We also have to make data easier for scientists to explore and use. The NASA-IBM Lunar Foundation Model shows what's possible when we bring AI to NASA's petabytes of scientific data. That's a real opportunity we see with AI: turning large-scale data into new discoveries."
What the model was trained on
The foundation model was pre-trained on roughly 2 million image tiles from LRO data, which covers most of the lunar surface and exceeds the volume of all other NASA planetary missions combined. The dataset includes more than 1 million high-resolution camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. Training also incorporated high-resolution Moon imagery and terrain data from NASA's GRAIL mission, NASA's Lunar Prospector, and JAXA's Selenological and Engineering Explorer.
Unlike traditional models that require building specialized algorithms from scratch, foundation models are pre-trained on vast, unlabeled datasets. The broad knowledge acquired through pre-training lets them generalize across multiple scientific domains through quick fine-tuning, which makes them efficient for accelerating research. For lunar scientists, this means adapting the model to specific tasks using only small amounts of labeled data.
Mapping ice, volcanoes, and craters
The model can help researchers estimate where ice patches are likely to be stable near the lunar poles, both on and below the surface. Dark areas like the Moon's permanently shadowed regions remain cold enough to trap and preserve ice for up to billions of years. Studying these areas offers insight into the Moon's history and presents an opportunity to map potentially usable resources for future space exploration.
For researchers studying lunar volcanism, the model accelerates identification of unusual volcanic features called irregular mare patches. These structures appear relatively young and challenge established timelines for lunar cooling. Mapping them could help scientists piece together a more accurate understanding of the Moon's thermal evolution. The model also maps surface features such as craters more efficiently than manual methods. Every crater forms from an impact, making crater counts essential for dating the lunar surface and reconstructing solar system history. The foundation model speeds up identification and measurement, allowing scientists to focus on interpreting findings.
In testing, the model matched or exceeded the performance of several strong baseline models across all evaluated tasks. It achieved comparable results on crater mapping and segmentation of irregular mare patches while showing a clear advantage on estimating polar ice stability. One test demonstrated the model's ability to detect a newly formed impact crater from a SpaceX rocket body strike, even though the post-impact image was excluded from pre-training.
A growing family of open science models
The lunar model is part of the Office of the Chief Science Data Officer's strategy for AI for Science & Research. It joins other models developed through the NASA-IBM partnership, including the Prithvi Models, pre-trained on Earth observation data for applications such as disaster monitoring and flood mapping, and the Surya Model, a heliophysics model trained on solar observation data to predict space weather events like solar flares.
The project brought together the Impact AI team at NASA's Marshall Space Flight Center, scientists in the Science Mission Directorate Planetary Science Division, NASA's Goddard Space Flight Center, and NASA's Ames Research Center. The team released machine learning-ready pre-training datasets and benchmark collections alongside the model, which is integrated into the open-source TerraTorch toolkit. A companion paper on Hugging Face supports reproducible research and equips scientists worldwide to build, compare, and refine AI models for lunar exploration. This approach to AI Data Analysis turns raw planetary data into a resource the global research community can use immediately.
Why this matters for science and research professionals
The model removes a bottleneck that lunar scientists have faced for years: manually sifting through petabytes of imagery to identify features of interest. By handling the initial detection work, the foundation model lets researchers spend their time on interpretation and hypothesis testing rather than counting craters. For teams working on lunar resource mapping, surface dating, or thermal history reconstruction, this means faster iteration between observation and insight. The open-source release and pre-packaged benchmark datasets also lower the barrier for smaller research groups that lack the compute resources to train foundation models from scratch.
Your membership also unlocks:
