The NASA-IBM Lunar Foundation Model was trained primarily on data collected by NASA’s Lunar Reconnaissance Orbiter, which has mapped almost the entire lunar surface in high resolution over the past 17 years. The model was trained on roughly two million image tiles, including more than one million high-resolution camera images and nearly 964,000 multispectral images.
Researchers can fine-tune the model for tasks such as mapping craters, identifying volcanic features, and estimating the location and stability of ice near the Moon’s polar regions. It can also help detect new surface changes, including impact craters, across large volumes of lunar imagery.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job,” said Kevin Murphy, NASA’s chief science data officer and acting chief data and AI officer. He said the model could help turn NASA’s large scientific datasets into new discoveries.
Unlike conventional AI systems developed for a single task, foundation models are pre-trained on large datasets and can be adapted to multiple research applications with comparatively small amounts of labelled data.
The model was also trained using imagery and terrain data from NASA’s GRAIL and Lunar Prospector missions, as well as Japan’s Selenological and Engineering Explorer mission. NASA said it performed as well as or better than several baseline models across the evaluated tasks, with a notable advantage in estimating polar ice stability.
The NASA-IBM Lunar Foundation Model, along with its codebase, datasets and benchmark collections, has been made publicly available through Hugging Face, GitHub and the open-source TerraTorch toolkit. NASA said the release is intended to support reproducible research and enable scientists worldwide to test and improve the model.
The initiative is part of NASA and IBM’s wider collaboration on AI for science, which also includes the Prithvi models for Earth-observation applications and the Surya model for space-weather prediction.
