Artificial intelligence is increasingly moving beyond chatbots and into scientific discovery. Its latest destination is the Moon.
NASA and IBM have released the NASA-IBM Lunar Foundation Model, an open-source AI system designed to help scientists analyse decades of lunar observations and identify features including potential ice deposits, craters and volcanic formations.
The model brings together more than 30 spatially aligned data layers from nine instruments across four lunar missions. Rather than forcing researchers to study each source of lunar data separately, the system is designed to find relationships across multiple types and resolutions of observations.
That may sound like a highly specialised research project, but the implications are surprisingly practical. Better maps of lunar ice and terrain could eventually help determine where astronauts land, where long-term infrastructure is built and which areas of the Moon deserve closer investigation.
What NASA and IBM have actually built
The new model is part of the Prithvi family of open foundation models developed through IBM and NASA’s collaboration on scientific AI.
NASA has accumulated enormous quantities of lunar observations over decades. The challenge is no longer simply collecting information; it is making sense of datasets produced by different instruments, at different resolutions and for different scientific purposes.
IBM and NASA say their new machine-learning-ready lunar dataset combines tens of thousands of images and maps, including information from NASA’s Lunar Reconnaissance Orbiter and GRAIL mission as well as complementary observations from Japan’s SELENE/Kaguya mission.
A foundation model can learn broader patterns across this information and then be adapted to specific scientific tasks, potentially reducing the need to build a completely new machine-learning system for every question researchers want to investigate.
AI could help scientists search for lunar ice
One of the most interesting applications is the search for water ice on the Moon.
Permanently shadowed lunar regions are notoriously difficult to observe, but some may contain ice below or around the surface. Finding and understanding those deposits is important because water is more than something future astronauts could drink.
Water can also provide oxygen and, when separated into hydrogen and oxygen, potentially contribute to producing propellant for future space missions. That makes accessible lunar resources an important part of the longer-term discussion around sustained exploration of the Moon and eventual missions deeper into the solar system.
In testing described by IBM and NASA researchers, the Lunar Foundation Model reduced error in identifying areas with high potential for lunar ice by as much as 22% compared with the referenced SwinV2-B model.
That does not mean AI has suddenly discovered a hidden reservoir of water. The more meaningful development is that researchers now have another tool for deciding where to investigate.
Safer landing sites are another potential use
The model can also help researchers analyse lunar craters.
Crater mapping is scientifically valuable because craters reveal clues about the Moon’s geological history. But accurate terrain information also matters when choosing potential landing sites and planning long-term surface operations.
At roughly 100-metre context-scale resolution, IBM says the model outperformed the comparison SwinV2-B system by nearly 19% while using half the training data. At finer resolutions it can identify and classify craters down to metre scale.
For future missions, understanding where steep slopes, boulders, craters and other hazards are located could contribute to safer planning.
Why this AI model matters beyond the Moon
The most interesting part of the announcement may not be any single benchmark.
It is what the project says about where AI is heading.
Much of the public AI conversation has focused on systems that generate text, images, video and software. Scientific foundation models represent a different direction: using machine learning to organise enormous datasets that would be difficult for individual researchers to examine manually.
Instead of asking an AI system to write something, scientists can use one to help identify patterns worth investigating.
That distinction matters. AI does not replace the scientific process or prove that a particular patch of lunar terrain contains ice. It can narrow an enormous search space and help researchers decide where limited human attention and future observations should go next.
An open model researchers can build on
NASA and IBM are releasing both the model and its associated lunar dataset openly, allowing researchers to adapt the technology for additional lunar-science tasks.
The approach follows earlier Prithvi projects covering areas including Earth observation, weather and heliophysics.
That openness could ultimately be as important as the model itself. Lunar exploration is becoming increasingly international, and giving researchers a common AI foundation provides a starting point for experiments that NASA and IBM may never have designed themselves.
AI may help decide where humanity goes next
There is an appealing irony to the project. Some of humanity’s most advanced AI technology is being applied to observations of an object people have studied for thousands of years.
But the problem has changed. Modern spacecraft produce vastly more information than previous generations of scientists could have imagined.
The next breakthrough in lunar exploration may therefore come not only from a new spacecraft or sensor, but from finding better ways to connect observations we already have.
If AI can help researchers identify promising resources, understand hazards and choose where to look next, systems such as the NASA-IBM Lunar Foundation Model could quietly influence where future astronauts explore — and perhaps where humanity eventually builds its first lasting presence beyond Earth.
Sources: IBM and NASA’s September 10, 2026 announcement of the NASA-IBM Lunar Foundation Model; IBM Research technical overview; independent reporting by Reuters.
Published September 11, 2026.
