IBM and NASA open up an AI model to read the Moon: lunar resources start with the data

By Zendoric · September 13, 2026.

IBM and NASA have unveiled the NASA-IBM Lunar Foundation Model, an open-source artificial intelligence system trained on decades of lunar observations. It is, according to both institutions, one of the first "foundation models" —a model trained on large volumes of raw data that is then adapted to specific tasks, rather than built from scratch for each one— designed specifically for lunar science.

The training combines more than 30 spatially aligned data layers, drawn from nine instruments and four missions: NASA's Lunar Reconnaissance Orbiter and GRAIL mission, plus data from the Japanese SELENE/Kaguya mission, run by the JAXA space agency. According to the technical paper published by IBM and NASA, the model cuts the error in identifying areas with high ice potential by up to 22% compared with the SwinV2-B reference model, improves 100-meter-scale crater mapping by almost 19% while using only half the training data, and sharpens by 3% the detection of Irregular Mare Patches, formations that document the Moon's past volcanic activity.

The goal is not abstract. The permanently shadowed regions of the lunar poles, some of the hardest areas of the satellite to observe, could hold ice beneath the surface: water and oxygen to sustain a human base, and potentially rocket fuel for later hops toward Mars. Precise crater mapping, in turn, is what allows NASA to choose safe landing zones and plan where to build long-term infrastructure. Kevin Murphy, NASA's acting chief data and AI officer, summed it up this way: "collecting data is only part of the job; we must also make it easier for scientists to explore and use the data." Juan Bernabe-Moreno, of IBM Research, speaks of an "open platform on which the global scientific community can keep building."

Alongside the model, both institutions have also released the first unified, publicly accessible lunar dataset of its kind: tens of thousands of images and maps combined into a format ready for training machine learning models, something that did not exist until now despite decades of accumulated observation.

As sector context, IBM and NASA do not arrive at this from nowhere: the two have long collaborated on geospatial foundation models for observing Earth —climate, land use, deforestation— published openly and widely reused by outside researchers. What changes here is the object of study: instead of monitoring our own planet, the same approach is applied to preparing the exploration of another celestial body. It is a sign of a pattern repeating across the industry: the AI frontier is no longer just the chatbot or the coding assistant, but foundation models specialized in very specific scientific domains —climate, materials, biology, and now lunar geology— that compress petabytes of observation into manageable tools.

Our reading is that this kind of announcement fits Zendoric's underlying thesis better than almost any other: AI as an engine of abundance, not merely of automating office tasks. There is no job displacement or market disruption here; there is an instrument that accelerates scientific discovery and, ultimately, the responsible use of resources —water, oxygen, energy— beyond Earth. That the model and the dataset are published openly also fits another thesis we have been maintaining: open science and technology are the fastest route for these advances to benefit the entire research community, not only whoever funds them.

It is still worth tempering the enthusiasm with both feet on the ground. A model that better locates ice on a map is not the same as an operational lunar base: between the laboratory data and real infrastructure lie engineering, funding and years of robotic and crewed missions, with the Artemis program as NASA's main vehicle for getting there. And it should not be forgotten that the Moon, and in particular its water-ice resources, has also become the object of growing geopolitical competition between the United States, China and India: open science coexists, uncomfortably, with the strategic interest in controlling those same deposits. Publishing the model openly is good news for the global scientific community; it does not remove that underlying tension, but it does show that, even in the middle of a space race, there is still room for knowledge to be shared before it is fought over.