NASA and IBM Release Open-Source AI Model for Lunar Research

NASA and IBM have jointly released an open-source AI model designed to advance lunar research. The model, trained on 17 years of data from NASA’s Lunar Reconnaissance Orbiter (LRO) and other lunar missions, integrates diverse observations into a single AI system. Researchers can use it to map craters, study volcanic features, and estimate stable water ice near the Moon’s poles.

The model was primarily trained on roughly 2 million image tiles, including over 1 million high-resolution images and nearly 964,000 multispectral images. It also incorporates data from NASA’s GRAIL and Lunar Prospector missions, as well as Japan’s SELENE mission. This approach allows scientists to compare datasets captured under varying conditions without building new models from scratch.

Initial applications focus on crater mapping, volcanic research, and polar ice prospecting. The AI model has shown an advantage in estimating polar ice stability and improved accuracy in identifying key lunar features by up to 23% in benchmark testing. However, it does not independently confirm the presence of lunar ice or determine safe landing sites.

The model is available through Hugging Face, with its codebase on GitHub. NASA and IBM have also released pre-training datasets and benchmark collections, integrated into IBM’s open-source TerraTorch toolkit. This open approach aims to broaden lunar research, allowing scientists worldwide to adapt the model for various scientific tasks and prepare for future missions.