IBM and NASA have released an open-source AI model designed to accelerate lunar surface research. The NASA-IBM Lunar Foundation Model is available for free download on Hugging Face, with full code published on GitHub.
The model addresses a longstanding challenge in lunar research. Traditionally, scientists either manually examined maps and images or trained narrow, low-resolution machine-learning models for individual tasks—both approaches expensive and prone to missing fine details. The foundation model allows researchers to apply a pre-trained system to new tasks rather than building separate models for different geologic features.
Juan Bernabe-Moreno, director of IBM Research Europe, UK and Ireland, described the capability: "The model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation."
Performance Benchmarks
NASA and IBM tested the model against SwinV2-B, a Microsoft-developed vision system commonly used as a baseline for image analysis. The lunar model reduced errors by 23 percent when locating ice deposits and outperformed SwinV2-B by 19 percent in identifying and classifying craters, despite training on half the data. It also showed a 3 percent improvement in identifying volcanic features called Irregular Mare Patches while requiring less fine-tuning.
The model was tested operationally on August 5, when IBM analyzed an image of an impact site created by a SpaceX Falcon 9 rocket striking the Moon. The model correctly identified the newly formed crater despite its proximity to an existing crater.
Research Applications
The model supports long-term lunar exploration objectives, particularly mapping permanently shadowed regions near the lunar poles. These areas, among the most challenging to observe on the Moon, may contain subsurface ice deposits. Ice could supply water and oxygen for future lunar bases and provide raw materials for rocket fuel to support missions to Mars.
The partnership also released what was described as the first unified, machine-learning-ready lunar dataset. This dataset combines more than 30 spatially aligned data layers collected by nine instruments across four separate lunar missions, drawing imagery from NASA's Lunar Reconnaissance Orbiter and GRAIL gravity mission, as well as data from Japan's SELENE/Kaguya orbiter. The dataset comprises roughly two million co-registered data points.
Bernabe-Moreno noted the dataset's significance: "The data is what really creates the industry of AI models." The model represents the latest addition to IBM's Prithvi family of open-science models, which spans geospatial analysis, weather, and heliophysics applications.
