NASA and IBM release open-source AI model to map lunar surface

In partnership with IBM, NASA unveiled the Lunar Foundation Model, an open-source AI tool for lunar image analysis that is downloadable from Hugging Face. The system excels at spotting ice-prone regions and classifying craters, reducing errors by 23 percent compared with Microsoft’s SwinV2-B and achieving a 19-percent gain in crater identification while needing half the training data. A test using an image of SpaceX’s Falcon 9 crash on August 5 confirmed the model’s ability to recognize new impact sites.

Developers also received a novel, fully co-registered dataset of over two million data points drawn from the Lunar Reconnaissance Orbiter, GRAIL and SELENE missions. Training the model required novel techniques to handle the Moon’s stark shadows and avoid traditional image-reconstruction methods. IBM’s Juan Bernabé-Moreno highlighted the dataset itself as a lasting contribution for future AI research on lunar science.