NASA and IBM Release Open-Source AI Model for Lunar Science
The Lunar Foundation Model, trained on 17 years of Lunar Reconnaissance Orbiter data, improves polar ice prediction by up to 22 percent and is freely available to researchers.
NASA and IBM have released the Lunar Foundation Model, one of the first open-source artificial intelligence models dedicated to lunar science. The model is designed to serve as a foundation for a wide range of research tasks, from mapping surface features to locating water ice, and it is being made freely available to the scientific community.
The system was trained on nearly 2 million tile bundles, the vast majority of which come from 17 years of observations by NASA's Lunar Reconnaissance Orbiter. That spacecraft has been circling the Moon since 2009, collecting high-resolution imagery and other data that have become a cornerstone of modern lunar research. By turning that archive into a trainable dataset, the team behind the model aims to give scientists a powerful new tool for interpreting the Moon's surface and subsurface.
One of the most significant results so far concerns the search for polar ice deposits. In tests, the Lunar Foundation Model reduced the error in predicting these deposits by up to 22 percent compared with the strongest competing model it was evaluated against. That improvement matters because water ice at the lunar poles is considered a critical resource for future exploration, potentially providing drinking water, breathable oxygen, and rocket propellant. More accurate maps of where that ice is located could directly shape mission planning and landing site selection.
The release also reflects a broader shift in how space agencies and technology companies collaborate. NASA has increasingly turned to commercial partners and open-source frameworks to accelerate scientific discovery, while IBM has been expanding its footprint in geospatial and scientific AI. By making the model open source, the two organizations are inviting outside researchers to fine-tune it for their own purposes, whether that means studying crater formation, analyzing mineral composition, or monitoring changes in the lunar surface over time.
For lunar scientists, the appeal of a foundation model lies in its flexibility. Rather than building a separate algorithm for every task, researchers can start with a system that has already absorbed a massive amount of orbital data and then adapt it to specific questions. This approach has proven effective in other domains, such as Earth observation and medical imaging, and its application to the Moon signals a growing maturity in planetary data science.
The model's reliance on the Lunar Reconnaissance Orbiter archive also highlights the enduring value of long-running missions. Data collected over nearly two decades are now being repurposed through machine learning, allowing scientists to extract insights that may not have been possible when the observations were first made. As NASA and its partners prepare for a new era of lunar exploration, including crewed missions under the Artemis program, tools like the Lunar Foundation Model could help ensure that those efforts are guided by the best available science.
The Lunar Foundation Model is available now, and its developers expect it to be used both for pure research and for applied projects related to resource mapping and mission support. Its release marks a step toward a more collaborative and data-driven approach to understanding Earth's nearest neighbor.
3
