NASA and IBM Release Open AI Model Mapping Lunar Ice, Craters and Volcanoes

NASA, in partnership with IBM Research and multiple academic institutions, has unveiled the NASA-IBM Lunar Foundation Model, an open-source artificial intelligence system designed to accelerate lunar surface analysis. Hosted publicly on Hugging Face with its codebase available on GitHub, the model represents a strategic shift in planetary science by leveraging pre-trained AI to process NASA’s extensive archival data. Kevin Murphy, NASA’s acting chief data and AI officer, emphasized that the initiative transforms petabytes of scientific records into actionable intelligence, enabling researchers to explore lunar geology with unprecedented speed and precision. The foundation model was trained on seventeen years of high-resolution imagery and terrain data, primarily from NASA’s Lunar Reconnaissance Orbiter, which captured a seamless mosaic of the entire lunar surface. The training dataset comprises approximately two million image tiles, including one million high-resolution camera images at one-meter resolution and nearly one million multispectral images at one-hundred-meter resolution. Supplementary data from the GRAIL, Lunar Prospector, and Japan’s Selenological and Engineering Explorer missions further enriched the model’s baseline knowledge. Unlike traditional task-specific algorithms, this foundation model acquires broad geological patterns during pre-training, allowing scientists to fine-tune it for specific objectives using minimal labeled data. Researchers can deploy the model to efficiently map impact craters, identify young volcanic formations known as irregular mare patches, and estimate the stability of polar ice deposits. The model demonstrated performance comparable to or exceeding established baselines across all evaluated tasks, with particular strength in predicting ice retention in permanently shadowed regions. By automating crater counting and segmentation, the system significantly reduces manual processing time, allowing scientists to focus on interpreting geological timelines and solar system history. Its ability to detect recently formed volcanic structures also challenges existing models of lunar thermal evolution, offering new insights into the moon’s geological activity. Developed under NASA’s Office of the Chief Science Data Officer and the broader AI for Science collaboration with IBM, the project integrates experts from Marshall Space Flight Center, Goddard Space Flight Center, Ames Research Center, and academic partners including the SETI Institute, Universities Space Research Association, and Howard University. The release includes machine-learning-ready pre-training datasets, benchmark collections, and integration with the open-source TerraTorch toolkit, ensuring reproducible research and community-driven model refinement. This initiative underscores a commitment to open science, providing global researchers with standardized tools to advance future lunar exploration and resource utilization strategies.