NASA and IBM Unveil Open-Source Lunar AI Foundation Model

NASA and IBM released the NASA-IBM Lunar Foundation Model, an open-source AI system for Moon exploration, plus a large co-registered dataset from LRO, GRAIL, and SELENE on Hugging Face. The model identifies ice-rich regions and craters, beating the SwinV2-B baseline by 23% in ice-area mapping and 19% in crater classification with half the training data. It correctly identified a new crater from a SpaceX Moon impact in August. Training on lunar data was difficult due to knife-edged shadows, prompting a wedge-based train/test split to maintain consistency. The release includes tens of thousands of images and instrument data and over two million co-registered data points, intended to accelerate future lunar AI research.