NASA and IBM Research have released an open-source foundation model for lunar science. The model is trained on 17 years of data from the Lunar Reconnaissance Orbiter, alongside information from the GRAIL, Lunar Prospector, and JAXA Kaguya missions. The dataset, known as SomBench, includes nearly 2 million tile bundles across 11 modalities.
The model architecture is based on the TerraMind Earth observation system. It processes high-resolution and coarse imagery simultaneously, using FlexiViT to adapt to different patch sizes. The training data is split geographically to prevent leakage, incorporating explicit context like illumination angles and sun position to account for lunar lighting conditions.
Performance tests show the model reduces ice deposit prediction error by up to 22 percent compared to the SwinV2-B baseline. In crater detection at a 100-meter scale, it outperformed the same baseline by nearly 19 percent using half the training data. The model is available on Hugging Face, with code hosted on GitHub and integrated into the TerraTorch toolkit.
This project is part of the NASA-IBM AI for Science collaboration, active since early 2022. While the model excels at pattern recognition, researchers note it is not suitable for absolute geodetic positioning. The team previously released the Prithvi model for Earth observation in August 2023, which was integrated into the IBM Watsonx platform.
