IBM and NASA Release Open-Source AI Model to Support Lunar Exploration

Details

IBM and NASA have publicly released the NASA-IBM Lunar Foundation Model, one of the first foundation models built specifically for Moon exploration. Trained on an extensive multi-instrument lunar dataset curated by IBM and NASA researchers, the model can process petabytes of sensor data across different types and resolutions simultaneously — something previous task-specific models couldn't do. The model's technical performance is measured against the SwinV2-B ImageNet benchmark: it reduces error (RMSE) by up to 22% in predicting lunar ice locations in permanently shadowed regions. These regions are priority targets for future Moon bases because ice indicates the presence of water and oxygen — resources needed for life support and producing rocket fuel for Mars missions. The model also identifies volcanic features called Irregular Mare Patches better than the benchmark by 3%, using imperfect labels with comparable accuracy and greater efficiency. The multimodal, multi-resolution architecture lets researchers combine different instrument types — cameras, spectrometers, radar — in a single inference pass, replacing the need to run separate task-specific models for each data type. The open-source release makes the model freely available for the broader scientific community to adapt and build on. This is the second major NASA collaboration model from IBM, following earlier Earth observation foundation model work. The pattern suggests a repeatable framework for applying foundation models to large-scale scientific datasets in space agencies.