NASA and IBM Research have released the open-source NASA-IBM Lunar Foundation Model, transforming 17 years of lunar observation data into a versatile foundation for machine learning and lunar science. The model is designed to solve the challenge of scarce data labels in planetary research by leveraging pretrained capabilities on vast amounts of unlabeled data, specifically excelling at predicting polar ice deposits and detecting craters.
The model was trained from scratch using SomBench, the largest co-registered multimodal lunar corpus to date, which includes nearly 2 million tile bundles gathered from nine instruments and four missions. This dataset integrates 17 years of observations from the Lunar Reconnaissance Orbiter alongside data from the GRAIL, Lunar Prospector, and Kaguya missions. By incorporating explicit imaging geometry—such as sun position and illumination angles—as direct input, the model accurately accounts for extreme lunar lighting conditions.
In performance evaluations, the pretrained model matched or outperformed existing baselines and control models across tasks like crater detection and volcanic feature segmentation. The most significant improvement occurred in polar ice prediction, where the model reduced error rates by up to 22 percent compared to the SwinV2-B baseline. This open-source foundation model matters because it makes decades of complex planetary observation data easily usable for scientists studying water resources and surface features on the Moon.