NASA and IBM release open-source AI trained on 2 million lunar images

NASA and IBM this week released an open-source artificial intelligence model trained on roughly 2 million images and data layers of the lunar surface, the same week NASA published the results of a separate effort to extend its data-sharing standards to all 71 countries that have signed the Artemis Accords. Taken together, the two announcements mark one of the clearest examples yet of how open scientific infrastructure and AI are becoming intertwined in the renewed race to explore the Moon.

The model, called the NASA-IBM Lunar Foundation Model, was announced on September 10 and is freely downloadable on Hugging Face under an open-source license, with fine-tuning code posted to a NASA-affiliated GitHub repository. A day later, NASA said in a separate release that it had wrapped up a two-part virtual workshop series, held between July 28 and September 8, that gave the 71 Artemis Accords signatory nations a working reference architecture for organizing and sharing their own lunar science data.

Millions of images, a handful of percentage points

IBM said the foundation model was trained on more than 30 spatially aligned data layers gathered by nine instruments across four NASA and international missions, including the Lunar Reconnaissance Orbiter, the GRAIL gravity-mapping mission, Lunar Prospector and Japan's SELENE spacecraft. NASA's own description of the training set put the volume at close to 2 million image tiles, more than a million of them one-meter-resolution camera frames.

On the tasks NASA and IBM tested, the model cut error rates in identifying potential ice deposits near the lunar poles by roughly 22 percent compared with a standard computer-vision baseline, and it matched or exceeded that baseline on crater mapping and on spotting young volcanic features known as irregular mare patches, according to the model's release notes described in IBM's announcement. At a coarser, 100-meter resolution, the model showed roughly a 19 percent improvement over the baseline while training on half as much labeled data, a detail NASA's science division highlighted as evidence that future missions could get useful results with far less manual data-labeling work.

An old data-loss problem, and a new standard to fix it

The open-data push behind both announcements traces back more than four decades. A 1982 National Academy of Sciences review found that scientific data from NASA's earliest planetary missions was becoming unusable or was being lost outright, because there was no common format or documentation standard. NASA created the Planetary Data System in 1989 in response, and its current version, known as PDS4, requires missions to describe every dataset with a standardized, non-proprietary metadata format so that, in principle, a rover image from one country's mission can be cross-referenced against an orbital instrument reading from another's.

That compatibility is what let IBM and NASA combine data from four different missions, run by two different space agencies, into a single training set for the new model. It is also the model NASA is now asking its 71 Artemis Accords partners to adopt for their own lunar data. "NASA is committed to leading by example when it comes to open science," said Andrew Mitchell, deputy chief science data officer for NASA's Science Mission Directorate, in describing the workshops that walked partner nations through the PDS framework. Jacob Bleacher, NASA's chief exploration scientist, said the goal was to let scientific "discovery through transparency, collaboration, and accessibility" keep pace with the Artemis program's return to the Moon.

Two coalitions, two incompatible systems

The Artemis Accords, a set of principles for civil space exploration that the United States and seven other founding nations established in 2020, now count 71 signatories, and the Republic of Djibouti is due to become the 72nd on Monday at a ceremony NASA has scheduled at its Washington headquarters, according to a NASA media advisory. China and Russia are not among the signatories; the two countries are instead building their own lunar effort, the International Lunar Research Station, and, according to reporting on the rollout, the data architecture behind that rival coalition is not designed to interoperate with the PDS4 standard NASA is distributing to Artemis partners.

The practical effect, at least for now, is a lunar science community split into two separate data ecosystems that cannot easily exchange results, even as missions from Japan's space agency, India's ISRO and the United Arab Emirates approach the Moon on the Artemis-aligned side of that divide. NASA's pitch to those partners is that adopting a common standard multiplies the value of each individual mission, since a crater map built from one country's camera can be laid directly over another's spectrometer readings, or folded into shared AI tools like the new foundation model. Smaller space agencies and universities without the budget to build their own data-processing pipelines stand to gain the most from that arrangement, since they can plug directly into an already-built archive rather than developing equivalent infrastructure on their own.

"The NASA-IBM Lunar Foundation Model shows what's possible when we bring AI to NASA's petabytes of scientific data," said Kevin Murphy, NASA's chief science data officer, describing the model as a demonstration of what standardized, openly shared archives can enable.

What happens next

Monday's signing ceremony will add Djibouti as the 72nd Artemis Accords nation, and NASA officials have signaled they intend to keep running similar open-data workshops as more countries prepare their own lunar and Mars missions. For researchers outside government space programs, the more immediate change is that the lunar foundation model, its training data and its underlying code are already publicly posted, available for any scientist, company or student to download, adapt and build on without a licensing agreement or a NASA affiliation.