NASA and IBM Research launched an open-source artificial intelligence model on September 10, 2026, designed to process nearly two decades of lunar data and assist scientists with future crewed missions to the Moon. The NASA-IBM Lunar Foundation Model helps researchers map impact craters, locate potential ice deposits, and identify rare geological features using petabytes of archived planetary information.
Seventeen Years of Lunar Data Mapped by NASA
The artificial intelligence model relies primarily on observations collected by NASA’s Lunar Reconnaissance Orbiter, which has spent more than 17 years mapping the lunar surface since 2009. The spacecraft’s dataset exceeds the volume of all other NASA planetary missions combined, forming the baseline for modern lunar cartography. Developers trained the system on two million Lunar Reconnaissance Orbiter image tiles, merging one million high-resolution one-meter captures with 964,000 multispectral frames at 100-meter resolution.
Developers also integrated terrain data and imagery from NASA’s GRAIL and Lunar Prospector missions alongside Japan’s SELENE, also known as the Kaguya spacecraft. These supplementary datasets supply the model with gravity signals and compositional metrics, allowing the system to interpret subsurface geological processes rather than relying solely on visual surface appearance. Rather than building custom machine-learning architectures for individual queries, researchers can adapt the pre-trained foundation model using smaller, specialized datasets.
Identifying Ice, Craters, and Volcanic Patches
One primary application involves searching for potential water ice near the lunar poles, where permanently shadowed regions maintain temperatures cold enough to preserve deposits for billions of years. The model estimates surface and subsurface ice stability to pinpoint target areas for subsequent observation. While these calculations do not confirm physical presence, locating prospective resources supports mission goals by identifying potential sources for drinking water, oxygen, and rocket fuel.
The system also automates impact crater mapping, a task researchers use to determine surface ages and reconstruct the geological history of the inner solar system. In benchmark tests, the AI matched or outperformed existing baseline systems in detecting and measuring craters, sparing scientists weeks of manual image labeling. The model also detects irregular mare patches, which are relatively young volcanic formations that provide insight into the Moon's thermal evolution and help establish precise timelines for past volcanic activity.
Open-Source Tools for Global Researchers
NASA hosted the model on Hugging Face and published the full source code on GitHub, integrating the software into the open-source TerraTorch toolkit to standardize geoscience benchmarking. The project builds on previous collaborations between NASA and IBM that produced Earth-focused Prithvi models and the Surya space-weather model. Within NASA, the Impact AI team at Marshall Space Flight Center partnered with the Planetary Science Division, Goddard Space Flight Center, and Ames Research Center to develop the platform.
External academic contributors included the Universities Space Research Association, the SETI Institute, the University of Maryland, Baltimore County, and Howard University. The release allows universities, space agencies, and private commercial ventures to test hypotheses on identical data and contribute improvements back to the repository. Kevin Murphy, NASA’s chief science data officer, stated that the project demonstrates how artificial intelligence turns large-scale data archives into active discoveries.
Frequently Asked Questions About the Lunar Foundation Model
What specific spacecraft data trained the NASA-IBM model?
The model was trained primarily on observations from NASA’s Lunar Reconnaissance Orbiter, supplemented by data from NASA’s GRAIL and Lunar Prospector missions as well as Japan’s SELENE spacecraft.
Where can researchers access the source code and documentation?
NASA published the full source code on GitHub and hosted the model, companion paper, documentation, and example notebooks on Hugging Face.
What are the primary scientific uses for the model?
Scientists use the tool to estimate water ice stability near the lunar poles, automate impact crater mapping, and identify young volcanic formations known as irregular mare patches.
Who collaborated with NASA on the project?
IBM Research co-developed the model alongside academic institutions including the Universities Space Research Association, the SETI Institute, the University of Maryland, Baltimore County, and Howard University.
Unresolved Questions Surrounding Lunar Ice Predictions
While the foundation model generates predictive maps for potential polar ice and geological features, those stability calculations remain unconfirmed by physical samples. Researchers have not yet verified the model’s subsurface ice estimates against ground-truth data collected directly by rovers or human astronauts.
Keep reading