IBM and NASA Release Open-Source AI Model for Lunar Exploration

For decades, NASA's instruments have been watching the Moon—mapping its craters, measuring its thermal signatures, scanning for ice in permanently shadowed regions. The data accumulated in petabytes, layered across nine different instruments on four separate missions. But having the data and being able to use it are different things. Scientists either spent weeks manually sifting through maps and images or relied on narrow, task-specific machine learning models that were computationally expensive and often lacked the precision needed for serious scientific work. The bottleneck was not observation; it was translation.

IBM and NASA have now released an open-source foundation model trained on this accumulated lunar record, designed to help researchers surface patterns across decades of multi-instrument observations at a scale no single tool has previously offered. The NASA-IBM Lunar Foundation Model represents the first publicly available foundation model built specifically for scientific exploration of the Moon. It works by learning from a unified dataset that aggregates more than 30 spatially-aligned data layers from instruments aboard NASA's Lunar Reconnaissance Orbiter and GRAIL mission, supplemented by complementary observations from Japan's SELENE/Kaguya spacecraft. The result is a machine learning-ready dataset that brings together tens of thousands of images and maps showing different geophysical properties of the lunar surface and subsurface—a common framework where previously no such unified resource existed.

The practical applications are immediate and concrete. Researchers can use the model to identify potential ice deposits in the Moon's permanently shadowed regions, where water and oxygen lie buried beneath the surface. These resources are considered essential for establishing a sustained human presence on the Moon and for producing rocket fuel for future missions to Mars. In testing, the model reduced error in identifying areas with high potential for lunar ice by 22 percent compared to existing state-of-the-art models. The model also helps scientists study the Moon's volcanic history by identifying Irregular Mare Patches—volcanic features that reveal information about the Moon's thermal evolution and that matter strategically for future surface operations. It captures the extent of these features 3 percent more accurately than comparable models while requiring lower fine-tuning costs. For crater detection, which is essential for understanding the Moon's age, geology, and chemical composition, and for selecting safe landing sites and planning long-term infrastructure, the model achieves comparable accuracy to leading alternatives while offering greater efficiency. At context-scale resolution of roughly 100 meters, it outperforms existing models by nearly 19 percent using only half the training data.

Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer, framed the release as a fundamental shift in how the agency approaches its scientific archive. "NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," he said. "We also have to make data easier for scientists to explore and use." Juan Bernabe-Moreno, director of IBM Research Europe, UK and Ireland, emphasized the collaborative dimension: "The NASA-IBM Lunar Foundation Model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation, and providing an open platform the global research community can build on."

The release extends a broader partnership between IBM and NASA that has produced a family of open foundation models spanning geospatial data, weather, heliophysics, and now lunar science. The strategy reflects a shift in how scientific AI is being built: instead of constructing a new algorithmic system for each research question, scientists can now start from a shared, pre-trained model and adapt it to their specific tasks. By open-sourcing the lunar model, the agencies are inviting the global research community to build on the same foundation, accelerating discovery across domains and democratizing access to tools that were previously available only to well-resourced institutions. The model is available now, and the unified lunar dataset it was trained on is open to researchers worldwide.