![\"The](\"https:\/\/i0.wp.com\/mereja.com\/index\/wp-content\/uploads\/2026\/09\/moon-1280.jpg?resize=780%2C439&ssl=1\")
<\/p>\nNASA and IBM Research released the NASA-IBM Lunar Foundation Model on 10 September, an open-source AI system built mainly on 17 years of Lunar Reconnaissance Orbiter imagery and posted on Hugging Face with code on GitHub. The model was trained on roughly two million image tiles, more than a million high-resolution camera frames at metre scale plus nearly a million multispectral scenes, and pulled in supporting data from GRAIL, Lunar Prospector, and JAXA\u2019s SELENE. Unlike one-off detectors that start from scratch for each task, the foundation approach is meant to be fine-tuned quickly for crater mapping, irregular mare patches linked to volcanic history, and estimates of where polar ice may stay stable in permanently shadowed craters.<\/p>\n
<\/p>\n
Kevin Murphy, NASA\u2019s chief science data officer, cast the release as a way to make petabytes of lunar data usable rather than merely archived. Early benchmarks say the model matched or beat strong baselines on crater work and irregular mare segmentation, with a clearer edge on ice prospectivity. NASA already fields related Earth and Sun foundation models (Prithvi and Surya) through the same IBM partnership. For Artemis-era planning, faster ice and hazard maps matter because polar cold traps can hold water ice that crews might one day turn into drinking water, oxygen, or propellant.<\/p>\n
Open weights do not replace field geology or orbital confirmation. They do change who can run lunar science experiments without a national lab cluster. Universities and smaller space agencies can fine-tune the same backbone NASA used, then publish comparable results. That is the quieter story under the press release: lunar AI is joining weather and Earth-observation foundation models as shared infrastructure, not a closed mission tool.<\/p>\n
Sources:<\/strong><\/p>\n