NASA and IBM Release Open-Source Lunar Foundation Model to Support Moon Research

The model was pretrained from scratch using SomBench, a multimodal lunar dataset containing nearly two million co-registered data bundles across 11 modalities and two spatial scales. This dataset includes a wide range of information about the lunar surface, such as imagery, topography, illumination geometry, thermophysical properties, mineralogy, radar, gravity, and other geologic and environmental data.

The development involved significant contributions from the Universities Space Research Association (USRA), which provided planetary science expertise, dataset development, and scientific evaluation. USRA's involvement ensured that the model development and applications were aligned with lunar science priorities and the physical characteristics of planetary datasets.

The NASA-IBM Lunar Foundation Model was evaluated using three downstream benchmarks: crater detection at regional and meter scales, segmentation of irregular mare patches (IMPs), and regression of lunar polar ice prospectivity. In these tests, the model matched or outperformed comparison models based on ImageNet pretraining, as well as identical models initialized without lunar pretraining. The study also noted high label efficiency in crater detection, suggesting that lunar pretraining can reduce the amount of task-specific labeled data required for certain applications.