IBM and NASA have open-sourced the NASA-IBM Lunar Foundation Model, an AI system that consolidates decades of multi-modal lunar data from US and Japanese missions into a single representation. Built on IBM's TerraMind architecture, the model integrates observations at varying scales and viewing angles to address challenges like complex lighting and data resolution.
The model is initially prioritized for mapping uncatalogued craters, investigating volcanic history, and locating ice in polar regions. It uses lightweight low-rank adapters (LoRAs) to fine-tune the base model while keeping 90% of weights frozen. In testing, it reduced error rates by 22% compared to a SwinV2 transformer for ice prospecting and outperformed task-specific models in crater detection and volcanic feature mapping.
This open-source release aims to help scientists explore the lunar landscape and assist future astronauts in identifying resources like water and helium-3 for the Artemis program.