NASA and IBM Research have made a new lunar AI model publicly available. The NASA‑IBM Lunar Foundation Model is open‑source and can be downloaded from Hugging Face, with its code on GitHub. It was trained on 17 years of data from the Lunar Reconnaissance Orbiter and other missions, turning that archive into a tool that scientists can adapt with only a few labeled examples.
What the model can do
The model works with high‑resolution (about 1 meter per pixel) and coarse (100 m per pixel) images at the same time. It also takes lighting angles and sun position as explicit inputs, which helps it separate real surface features from shadows. In tests, it cut error on polar‑ice predictions by up to 22 % compared with the best baseline model. For crater detection at coarse scales it was about 19 % more accurate, even though it used only half the labeled data.
How it was built
Researchers assembled a data set called SomBench, the largest co‑registered multimodal lunar corpus to date. It contains nearly 2 million “tile bundles” from 11 different data types, covering more than 30 aligned layers from nine instruments across four missions. The model was trained from scratch using a technique called FlexiViT, which lets the same model handle different image patch sizes without retraining.
Why it matters
For planetary scientists, the open model provides a ready‑made AI that can spot ice and craters with fewer hand‑labeled examples, speeding up research on lunar resources. For anyone following future Moon missions, better ice maps could inform where to land or mine water for rocket fuel. However, the model is not a substitute for direct measurements; its geographic coordinates can still be off by dozens of degrees.
