NASA and IBM Launch Open-Source Lunar Foundation Model Trained on 17 Years of Orbiter Data

NASA and IBM have released the Lunar Foundation Model, an open-source AI system built to support lunar science and one of the first foundation models aimed specifically at the Moon. The model draws on nearly 2 million tile bundles, most of them collected over 17 years by NASA's Lunar Reconnaissance Orbiter, and is available for researchers to download and build on. NASA and IBM say the model reduces error in predicting polar ice deposits by as much as 22 percent compared with the strongest competing model it was tested against.

The training data traces back to the Lunar Reconnaissance Orbiter, which has been circling the Moon since 2009 and returning imagery and instrument readings ever since. Rather than training a narrow model for a single task, the team assembled those observations into tile bundles that let the system learn general representations of lunar terrain. That approach mirrors the foundation-model playbook used on Earth, where large pretrained systems are adapted to many downstream jobs instead of being rebuilt from scratch each time.

The headline number is the 22 percent error reduction on polar ice prediction, a task that matters because water ice near the lunar poles is a leading target for future crewed and robotic missions. Locating those deposits accurately affects where landers touch down and where resources might eventually be extracted. A model that sharpens that prediction from orbital data could change how mission planners prioritize candidate sites.

Open sourcing the model matters as much as the accuracy gain. Lunar datasets are expensive to gather and often siloed across missions and institutions, so a shared pretrained base gives smaller research groups and university labs a starting point they could not afford to build alone. It also sets up a common benchmark that different teams can fine-tune against, which tends to speed up comparison and replication across the field.

NASA and IBM have not said how often the model will be updated or whether it will fold in data from newer lunar missions, such as the increasing flow of commercial lander imagery. The open question is whether a foundation model trained largely on one orbiter's 17-year record will generalize to the different sensors and resolutions those newer missions bring. Researchers can already test that themselves, since the weights are public.

The release also lands amid a broader push to apply Earth-style AI infrastructure to space science, from Mars terrain analysis to satellite imagery processing. If the lunar model holds up on independent benchmarks, it could become a default starting point for polar ice mapping and surface characterization work. For now, the practical test is simple: whether outside teams adopt it and report results that match the 22 percent figure.