NASA and IBM Release Lunar Foundation Model Trained on 17 Years of LRO Data

NASA and IBM launch an open-source AI model trained on 17 years of lunar data

NASA, in collaboration with IBM Research and several academic partners, has released the NASA-IBM Lunar Foundation Model — an open-source artificial intelligence model trained primarily on 17 years of data from NASA’s Lunar Reconnaissance Orbiter (LRO). The model, designed specifically for lunar science, is publicly hosted on Hugging Face with the complete codebase available on GitHub for testing and experimentation.

The model aims to help researchers rapidly analyze the moon’s surface by turning petabytes of LRO observations into actionable outputs for mapping, geological interpretation and mission planning. LRO has collected multi-instrument datasets since its 2009 arrival at the moon; NASA says those 17 years of coverage made the dataset particularly well suited to pre-training a foundation model that can generalize across many lunar science tasks.

Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters in Washington, is quoted in NASA’s announcement saying: “NASA has spent decades building an extraordinary scientific record of the moon, but collecting data is only part of the job. We also have to make data easier for scientists to explore and use. The NASA-IBM Lunar Foundation Model shows what's possible when we bring AI to NASA's petabytes of scientific data.”

Foundation models are pre-trained on large, mostly unlabeled datasets to acquire broad knowledge that can be fine-tuned quickly for specific applications. NASA and IBM present this architecture as an efficient way to accelerate research across multiple domains of lunar science without building new task-specific algorithms from scratch each time.

According to the announcements, the model will support tasks such as mapping the distribution of surface water ice, identifying craters and volcanic features, and generally speeding up analyses that previously required bespoke software and painstaking manual work. The LRO instruments and long-duration record provide imagery and other measurements that cover most of the lunar surface, which the team used to expose the model to a wide variety of terrains and lighting conditions.

Both the NASA Science release and the Phys.org report emphasize that the model is intended to be an open scientific tool. The Hugging Face deployment and GitHub repository make it possible for researchers, educators and the public to inspect, test and fine-tune the model for their own studies. NASA frames this as part of an effort to make its scientific data more accessible and usable.

NASA’s announcement highlights the pragmatic motivations behind the release: accelerating the pace at which large lunar datasets can be turned into insight and operational products. By reducing the time needed to develop task-specific algorithms, the foundation model approach can help researchers respond faster to questions about the moon's geology and support planning for future missions that will rely on detailed surface information.

Contextually, the release comes as lunar exploration is experiencing renewed activity worldwide, with many agencies and commercial partners preparing missions that require detailed surface knowledge — for science, resource prospecting and landing site selection. While the NASA-IBM model does not replace targeted scientific analysis, the partners say it can streamline many of the preliminary mapping and classification steps that once consumed large amounts of researcher time.

There are limits and distinctions to note: foundation models are powerful generalists but are not a substitute for hypothesis-driven, peer-reviewed research. NASA’s write-up describes the model as a tool to make data easier to explore and use, not as a source of definitive scientific claims. The announcements underline that further validation and fine-tuning by domain experts remain necessary when using AI-generated products for scientific conclusions or mission-critical decisions.

Making the model open source also invites scrutiny and improvement from the broader community. Hosting on Hugging Face and publishing code on GitHub permits reproducibility and collaborative refinement, aligning with NASA’s stated goal of boosting open science and data sharing. Researchers will be able to adapt the model to niche tasks — for example, more precise detection of permanently shadowed regions, improved crater catalogs or focused volcanic mapping — using smaller labeled datasets for fine-tuning.

Looking ahead, the NASA-IBM Lunar Foundation Model represents a practical step toward integrating modern AI tools with long-standing planetary datasets. If widely adopted, it could reduce barriers to entry for lunar research and help teams convert the LRO archive into refined maps and products needed by upcoming missions. As NASA notes, collecting data is only part of the scientific enterprise; making that data easier to use is the next task the agency and its partners are addressing with this release.

NASA Science: “NASA, IBM Launch AI Foundation Model for Lunar Science” (official NASA release); Phys.org: “Release of open AI model trained on 17 years of lunar data maps ice, craters and volcanoes” (reporting on NASA release).