NASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science

Lunar ice deposit predictions become up to 22 percent more accurate, potentially guiding future resource extraction.

NASA and IBM Research, collaborating with academic institutions, have launched the NASA-IBM Lunar Foundation Model, an open-source tool for lunar science. This model is pretrained on a vast dataset of unlabeled lunar observation data, allowing it to be adapted for specific tasks with minimal labeled examples. The foundation model was trained on SomBench, the largest co-registered multimodal lunar corpus to date, comprising nearly 2 million tile bundles from 17 years of Lunar Reconnaissance Orbiter (LRO) data, supplemented by information from GRAIL, Lunar Prospector, and JAXA's Kaguya/SELENE probes. A key innovation is providing lighting geometry as explicit input, rather than requiring the model to infer it from pixels, which significantly aids in tasks like predicting ice deposits and detecting craters. The model shows substantial improvements, particularly in identifying potential ice distribution at the lunar poles, reducing prediction error by up to 22 percent compared to leading baselines. While effective for analysis, the model is not a substitute for physical measurements and has limitations in precise geodetic positioning.

This release signifies a pivotal step in democratizing access to advanced lunar data analysis, moving beyond specialized algorithms to a more flexible foundation model approach. By leveraging a massive, multimodal dataset and explicitly incorporating illumination geometry, the model overcomes key limitations in current lunar research, particularly where labeled data is scarce. The open-source nature and public availability on Hugging Face suggest a broader impact, enabling a wider community of researchers and developers to build upon this work for new scientific discoveries and potential resource utilization strategies.

The successful development and application of the NASA-IBM Lunar Foundation Model indicate a growing trend toward large, foundational AI models tailored for specific scientific domains, similar to IBM's Prithvi model for Earth observation and Google Deepmind's AlphaEarth Foundations. This approach promises to accelerate scientific breakthroughs by making complex datasets more accessible and analytical capabilities more robust. Researchers and space agencies should anticipate more domain-specific foundation models emerging, which could streamline the analysis of vast amounts of scientific data across various fields.