NASA and IBM have released the NASA-IBM Lunar Foundation Model, one of the first publicly available foundation models developed specifically for scientific exploration of the Moon. Trained on a large, multi-instrument lunar dataset assembled by researchers from both organizations, the model is designed to extract new insights from decades of observations and support scientific and operational planning for a sustained human presence on the lunar surface.
The release addresses a longstanding data-processing challenge in lunar science. Spacecraft have accumulated petabytes of observations covering the Moon’s surface and subsurface, but researchers have often had to examine maps and images manually or use lower-resolution machine-learning systems developed for individual tasks. Those approaches can require substantial computing resources and may not provide the accuracy or flexibility needed to identify and analyze complex geological features.
Until now, researchers also lacked a publicly available, unified dataset that aligned multimodal lunar observations at different resolutions within a common framework suitable for modern machine learning.
NASA and IBM developed such a dataset alongside the foundation model. It contains more than 30 spatially aligned data layers derived from nine instruments aboard four lunar missions. The collection brings together tens of thousands of images and maps, including observations from NASA’s Lunar Reconnaissance Orbiter and Gravity Recovery and Interior Laboratory mission, as well as complementary data from the Japan Aerospace Exploration Agency’s SELENE, also known as Kaguya.
Combining these observations allows the model to examine relationships that may be difficult to detect when individual datasets are analyzed separately. Potential applications include locating areas that may contain lunar ice, mapping craters, studying the Moon’s volcanic history and investigating connections between surface geology and subsurface structure.
Technical results reported by NASA and IBM indicate that the model can outperform a general-purpose computer-vision model while requiring less task-specific training data. Compared with the SwinV2-B model trained on ImageNet, the Lunar Foundation Model reduced errors in identifying areas with high lunar-ice potential by as much as 22%.
In tests involving Irregular Mare Patches—unusual volcanic landforms that may represent relatively recent geological activity—the model mapped their extent 3% more accurately despite being trained with imperfect labels.
For crater detection at a contextual spatial resolution of approximately 100 meters, the lunar model outperformed SwinV2-B by nearly 19% while using half as much training data. It can also identify and classify craters in meter-scale imagery, giving researchers a tool that can operate across different observation scales.
Those capabilities have implications beyond geological research. Accurate crater maps can contribute to landing-site assessment, surface-navigation planning and the placement of long-term lunar infrastructure. Improved identification of potential water-ice deposits could help prioritize areas for further orbital observations or surface investigations.
Water ice is a particularly important target for future lunar exploration. If accessible in sufficient quantities, it could supply water for crews and potentially be processed into oxygen and hydrogen for life-support systems and propellant production. The model does not replace direct measurements or surface verification, but it could help researchers narrow the search area and identify locations that merit more detailed investigation.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job,” said Kevin Murphy, NASA’s chief science data officer and acting chief data and artificial intelligence officer. “We also have to make data easier for scientists to explore and use.”
Juan Bernabe-Moreno, director of IBM Research Europe for the United Kingdom and Ireland, said the model provides a foundation for examining the Moon at scale by connecting observations from different instruments and revealing patterns that may remain hidden when datasets are studied independently.
The project extends the Prithvi family of open foundation models developed by IBM and NASA. Previous members of the family have supported research involving geospatial observations, weather and heliophysics. The lunar model applies the same broader approach to Moon science: researchers begin with a pretrained model and adapt it to specific investigations instead of constructing a separate machine-learning system for every scientific question.
NASA and IBM have made both the model and its accompanying dataset available to the research community. Scientists can now adapt the system for additional lunar-analysis tasks, test its performance against other models and incorporate new observations as international and commercial missions expand the volume of data collected around and on the Moon.
