NASA and IBM have released an open-source AI model and dataset designed to help researchers map lunar craters, volcanic formations and areas with potential ice deposits. The NASA-IBM Lunar Foundation Model processes observations collected by multiple instruments at different resolutions, giving scientists a shared starting point for specialized lunar-research tools instead of requiring them to build systems from scratch.
The release provides a pretrained model and machine-learning-ready dataset that could reduce the work required to assemble data and train a lunar-mapping system. For research teams working with limited computing resources, this lowers the barrier to building specialized analysis tools.
What the model does
Sensors and instruments have generated petabytes of information about the Moon. Researchers often examine those materials manually or use machine-learning models developed for individual tasks, according to IBM's announcement. The foundation model was trained to identify relationships across different types and resolutions of lunar data, allowing researchers to adapt it to particular scientific questions instead of developing a separate system for every project.
"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer, said in the announcement.
IBM and NASA also released what they describe as the first open-source lunar dataset of its kind to combine multimodal and multiresolution observations in a machine-learning-ready framework. The dataset contains more than 30 spatially aligned layers from nine instruments across four missions. Its sources include NASA's Lunar Reconnaissance Orbiter, Gravity Recovery and Interior Laboratory and Lunar Prospector missions, along with observations from the Japan Aerospace Exploration Agency's SELENE mission, also known as Kaguya.
Targeting ice, craters and volcanic features
One potential application is identifying areas where ice may exist beneath the lunar surface. Permanently shadowed regions are difficult to observe, but lunar ice could provide water and oxygen for future bases. It could also potentially be used to produce rocket fuel for missions traveling farther into space.
In a technical paper authored by IBM and NASA researchers, the model reduced root mean square error by as much as 22% compared with the SwinV2-B image model when predicting areas with high potential for lunar ice. The system was also tested on Irregular Mare Patches, unusual formations that scientists study to understand the Moon's volcanic and thermal history. When working with imperfect labels, the model improved the identification of the formations' extent by 3% compared with SwinV2-B.
Crater detection represents another use. At meter-scale resolution, the model delivered accuracy comparable to current specialized methods while offering greater efficiency and lower fine-tuning costs. At a broader contextual resolution of approximately 100 meters, it outperformed SwinV2-B by nearly 19% while using half as much training data, according to the paper.
Mapping craters can help scientists estimate the age and composition of lunar terrain. It can also help mission planners identify slopes, boulders and other hazards as NASA prepares for future surface missions following the Artemis II crewed flight around the Moon.
What this means for researchers
The Lunar Foundation Model joins IBM's Prithvi family of open models, which covers weather, geospatial analysis and heliophysics. The aim is to provide reusable AI foundations that researchers can fine-tune for individual scientific tasks. Research teams can begin with a model trained on a large collection of aligned lunar observations instead of assembling every data source independently.
Open access also allows developers to inspect the system, reproduce its reported results and create tools using the accompanying dataset. However, the results come from research conducted by the organizations behind the model. Independent researchers will need to test it across additional datasets and use cases before its broader performance becomes clear. For now, it should be treated as a research foundation rather than a replacement for scientific observations or mission-specific validation.
For scientists interested in applying foundation models to their own domains, this release fits into a broader pattern of open AI models built for scientific work. Researchers can explore an AI Learning Path for Research Scientists to understand how these tools fit into research workflows. The approach also aligns with wider developments in AI for Science & Research.
Why this matters for science and research professionals
The Lunar Foundation Model demonstrates a practical shift in how scientific data gets reused. Instead of each research group cleaning, aligning and training on raw observations independently, teams can start from a shared pretrained baseline. That could accelerate small-scale lunar studies and make specialized analysis feasible for labs without extensive machine-learning infrastructure. The key caveat for researchers: treat the reported performance metrics as preliminary until independent replication confirms them across varied terrain types, lighting conditions and instrument sources.
Your membership also unlocks:
