Built primarily using observations from NASA’s Lunar Reconnaissance Orbiter (LRO), the model is publicly available for researchers to test, adapt and develop for new lunar science applications. The Lunar Foundation Model has been trained on around two million image tiles covering much of the Moon, including more than one million high-resolution camera images and almost 964,000 multispectral images. Additional lunar imagery and terrain data from missions including NASA’s GRAIL and Lunar Prospector, alongside JAXA’s SELENE mission, were also incorporated into its training.
Initial testing found that the model matched or outperformed several established AI models across the evaluated tasks. It delivered comparable results for crater mapping and identifying irregular mare patches, while showing a particularly strong advantage when estimating the stability of ice deposits near the lunar poles. Speaking on the groundbreaking model, chief science data officer and acting chief data and AI officer at NASA Headquarters, Kevin Murphy said: “NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job.
“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.” The project brings together NASA’s Impact AI team at Marshall Space Flight Center with researchers from NASA’s Planetary Science Division, Goddard Space Flight Center and Ames Research Center, as well as IBM Research and academic institutions.
The AI model draws heavily on data gathered by LRO during more than 17 years of lunar observations. The spacecraft has produced an exceptionally detailed record of the Moon, covering almost its entire surface and generating more data than NASA’s other planetary missions combined. That scale makes the dataset particularly valuable for training a foundation model. Rather than developing a separate machine-learning system for every scientific problem, researchers can start with an AI model that has already learned patterns from a large collection of lunar observations and fine-tune it for individual applications.
The approach could substantially reduce the amount of labelled data and development time required for new research projects. Using mapping craters and ancient volcanic activity, the model is identifying and measuring lunar craters. Because craters form from impacts, their distribution and characteristics give scientists an important way to estimate the age of different parts of the lunar surface and investigate the history of the Solar System.
Traditionally, analysing huge numbers of craters can require extensive manual work. The Lunar Foundation Model can automate much of the mapping process, allowing researchers to spend more time interpreting the resulting data rather than identifying individual features. The model can also help scientists locate irregular mare patches, unusual volcanic formations that appear relatively young compared with much of the Moon’s surface. Although the Moon is not considered volcanically active today, it experienced substantial volcanic activity earlier in its history.
