NASA and IBM have released an open-source artificial intelligence model designed specifically to study the Moon. Announced on September 10, 2026, the NASA-IBM Lunar Foundation Model was trained on a huge archive of lunar images and geophysical measurements, much of it collected during 17 years of observations by NASA’s Lunar Reconnaissance Orbiter.
The idea is straightforward but powerful: instead of asking scientists to build a new machine-learning system from scratch every time they want to count craters, investigate unusual volcanic terrain, or estimate where polar ice could remain stable, the new foundation model already has a broad statistical understanding of lunar data. Researchers can then adapt it to a particular scientific task using a much smaller amount of labeled training data.
This does not mean an AI is independently “discovering the Moon,” and it certainly does not mean it can prove that water ice exists simply by looking at a picture. What it can do is help scientists search enormous datasets faster, combine different kinds of observations, and identify places or patterns worth investigating more closely. That could become increasingly valuable as robotic and human missions return to the lunar surface.
What exactly is the NASA-IBM Lunar Foundation Model?
A foundation model is an AI system pre-trained on a large dataset so that it can later be adapted to many related tasks. A familiar general-purpose AI model may learn patterns from text or images. This one was built around scientific observations of the Moon.
NASA says the model was trained on roughly two million image tiles. These include more than one million high-resolution camera images at about one-meter resolution and nearly 964,000 multispectral images at roughly 100-meter resolution. The training material came primarily from the Lunar Reconnaissance Orbiter, but data from NASA’s GRAIL and Lunar Prospector missions and Japan’s SELENE mission were also incorporated.
That combination matters because no single instrument tells scientists everything about the Moon. A camera reveals surface shapes and shadows. Spectral observations provide clues about composition. Temperature, topography, gravity and other measurements reveal different parts of the geological story. The model is intended to learn useful relationships across these different datasets rather than treating each observation as an isolated picture.
NASA and IBM have made the model public, along with machine-learning-ready datasets and benchmark material. It is integrated with the open-source TerraTorch toolkit. That means researchers outside NASA and IBM can test it, reproduce experiments, fine-tune it and compare their own methods against it.
How can AI find craters that humans might overlook?
The Moon is covered with impact craters, but those scars are more than dramatic scenery. Scientists use the number, size and distribution of craters to estimate the relative ages of different surfaces. In general, terrain that has been exposed longer has accumulated more impacts, although interpreting crater populations requires careful geological context.
Cataloging craters manually across a world-sized dataset is slow. Automated detection can scan far more terrain, especially for smaller features that have not yet been cataloged. According to IBM Research, the lunar foundation model could identify craters at one-meter resolution about as accurately as a strong task-specific comparison model. At a coarser 100-meter-per-pixel scale, IBM reported that it outperformed the comparison model by nearly 19% while using about half as much training data.
The model was also tested on a real surface change. NASA showed it images of the area near Einstein crater before and after a rocket-body impact. The post-impact image had not been included during pre-training. After being adapted for the task, the model highlighted the newly formed crater while recognizing previously existing craters around it.
That does not eliminate the need for planetary geologists. Lighting angle can change how small craters appear, and shadows can make similar terrain look surprisingly different from one orbit to another. AI detections still need validation. The benefit is that software can perform the first enormous sweep, leaving scientists more time to interpret what the changes mean.
Could the model really help scientists search for water ice?
This may be the model’s most attention-grabbing application, but it also needs the most careful wording. The AI does not directly see underground reservoirs of water. Instead, it can estimate ice prospectivity: locations where environmental conditions make stable ice more plausible based on the data supplied to it.
The lunar poles contain permanently shadowed regions where sunlight may not reach the surface for extremely long periods. Some of these places remain cold enough for volatile materials, including water ice, to survive for geological timescales. Confirming where that ice exists, how much is present and how accessible it is remains a major scientific and exploration challenge.
NASA says the foundation model reproduced fine-scale patterns in reference maps of ice prospectivity. IBM reported that, in one test involving unfamiliar polar terrain, the model reduced error by about 22% compared with a state-of-the-art transformer trained specifically for ice prospecting.
Why is that important? Lunar water is scientifically valuable because it can preserve information about the Moon’s history and the delivery of volatile materials through the Solar System. It could also become a practical resource. In principle, future explorers might process lunar ice into drinking water, oxygen, and hydrogen and oxygen propellants. But a promising AI map is not the same as a confirmed resource deposit. Spacecraft measurements, drilling and direct sampling would still be needed before mission planners could rely on any particular location.
What can AI reveal about the Moon’s volcanic past?
The Moon looks geologically quiet today, yet its surface records a far more active past. Vast dark plains called maria were created by ancient lava flows, and the timing of lunar volcanism helps scientists reconstruct how the Moon cooled internally.
One puzzle involves small structures known as irregular mare patches. Some appear unusually fresh compared with the surrounding terrain, raising questions about whether volcanic activity continued more recently than traditional timelines suggest. Determining their ages is difficult, and researchers first need reliable maps showing where these features are and how large they are.
The NASA-IBM model can be fine-tuned to identify and outline irregular mare patches. NASA says its performance was comparable to or better than strong baseline models across evaluated lunar tasks, while IBM reported an improvement of roughly 3% over a task-specific comparison model for mapping these volcanic features.
The important part is not the percentage by itself. A reusable foundation model can potentially perform several jobs without researchers having to construct an entirely separate AI pipeline for each scientific question. The same underlying system can be adapted to craters, polar environments, volcanic landforms and surface-change detection. That flexibility is one of the main reasons foundation models are attracting attention in science.
Will AI choose where astronauts land on the Moon?
Not by itself. Landing-site selection is a high-stakes engineering and scientific decision involving terrain hazards, lighting, communications, thermal conditions, spacecraft performance, mission goals and many other constraints. No responsible mission would simply ask an AI model for coordinates and send astronauts there.
But tools like the Lunar Foundation Model could become one layer in a much larger decision system. Imagine scientists examining thousands of square kilometers around the lunar south pole. AI could rapidly flag crater fields, unusual geology, possible surface changes or regions with conditions favorable for ice stability. Human teams could then combine those results with engineering maps, orbital observations and mission requirements.
The model may be especially useful because the volume of lunar data is growing faster than any individual research team can inspect manually. NASA notes that LRO’s dataset is larger than that of all other NASA planetary missions combined. Future orbiters, landers, rovers and astronauts will add even more observations.
There is also a broader lesson here. NASA and IBM have already developed foundation models for Earth observation and solar science. The lunar project shows how the same basic idea can be applied to a planetary body: train an AI on a deep scientific archive, release it as an adaptable research tool, and let specialists use it to ask narrower questions.
The most interesting possibility is therefore not that AI will replace the scientists exploring the Moon. It is that it may change which questions scientists can realistically ask. A dataset too large to examine manually can become searchable. Measurements collected by different instruments can be analyzed together. Small changes or unusual features can be flagged for human attention.
For the next era of lunar exploration, that may be exactly what researchers need. The Moon has been observed for centuries and mapped by spacecraft for decades, yet it still contains enormous amounts of information waiting to be connected. The NASA-IBM Lunar Foundation Model is an attempt to make those connections faster — while leaving the scientific judgment, verification and final decisions where they belong: with humans.
Post a Comment