The open-source NASA-IBM Lunar Foundation Model, dedicated to lunar science, has been launched through an ongoing collaboration between NASA, IBM Research and various academic institutions. Primarily trained on data from NASA’s Lunar Reconnaissance Orbiter (LRO), the model has been made available on Hugging Face; the complete code is available on GitHub.
The model aims to enable the rapid analysis of large datasets concerning the lunar surface, helping to map the surface, examine its geological history and plan future research. Unlike algorithms that are traditionally trained from scratch for each mission, foundation models are pre-trained on extensive unlabeled datasets and can then be adapted to different scientific tasks using a small amount of labeled data.
Data collected by LRO over the past 17 years was used for training because it covers most of the lunar surface in detail. The data produced by this mission is larger than that from all of NASA’s other planetary missions combined and provides an almost continuous, high-resolution mosaic of the Moon. The model was trained on approximately 2 million image tiles from a dataset consisting of more than 1 million high-resolution camera images at 1-meter resolution and approximately 964,000 multispectral images at 100-meter resolution. Images and terrain data from NASA’s GRAIL and Lunar Prospector missions and JAXA’s Selenological and Engineering Explorer mission were also used. The model can be adapted for mapping craters, identifying young volcanic formations and estimating the location of ice near the poles.
Why it matters
This study aims to transform lunar data from archives processed solely for individual missions into a shared research infrastructure that can be adapted to different scientific questions. Its open-source approach could enable lunar researchers to build their own analyses on an existing foundation and evaluate the same data for different purposes. This could reshape the data-processing process, particularly for teams working on surface mapping, geological studies and research planning. However, how successful the model will be in different scientific tasks and how well it can be adapted to situations beyond the data used in training remain questions that will need to be answered as its applications become more widespread.
Background
NASA is not a new name in the FikirPilot archive: we have published 29 news articles mentioning it in the past 90 days; the most recent was dated September 11, 2026.