NASA and IBM released the NASA-IBM Lunar Foundation Model on September 10, 2026, as an open-source foundation model for lunar remote-sensing research. Trained primarily on 17 years of Lunar Reconnaissance Orbiter observations, it can be adapted to estimate ice prospectivity, map craters, track surface changes and study irregular mare patches. The important caveat is also the most important fact: an ice-prospectivity estimate is not direct evidence that frozen water is present.
NASA and IBM release an open model for lunar science
A foundation model is a reusable machine-learning system that can be adapted to several related tasks instead of being built from scratch for just one job. In this case, NASA and IBM designed the model to work across different kinds of lunar remote-sensing data.
That makes the release more than a single-purpose crater detector. Researchers can fine-tune the model with labeled examples for particular questions, including where ice may be stable near the lunar poles, how craters are distributed, whether the surface has changed and where unusual volcanic features appear.
The release also includes machine-learning-ready datasets, benchmark collections, a companion technical report and integration with the open-source TerraTorch toolkit. Those pieces matter because a model is far more useful to researchers when they can adapt it to their own questions rather than treat it as a sealed demonstration.
The Moon-sized training archive
The model’s main training source is the Lunar Reconnaissance Orbiter, which has produced observations of the Moon across 17 years. NASA describes a corpus of roughly 2 million image tiles, but those tiles do not all represent the same kind of measurement or level of detail.
More than 1 million images have an approximate resolution of 1 meter per pixel. Nearly 964,000 multispectral images have an approximate resolution of 100 meters per pixel. Multispectral data records information across multiple wavelength bands, giving researchers more than a conventional grayscale view of the terrain.
Additional lunar data came from GRAIL, Lunar Prospector and the Japan Aerospace Exploration Agency’s Selenological and Engineering Explorer, better known as Kaguya. Combining these sources gives the model different views of the Moon’s surface and interior-related properties, while keeping the model’s tasks tied to scientific mapping rather than general-purpose chatbot behavior.
What “ice prospectivity” really means
The model does not directly detect lunar ice. It estimates where ice may be likely to occur or remain stable, including at or below the surface. In other words, it produces a map of prospectivity: areas that match conditions associated with a greater possibility of stable ice.
That distinction is easy to lose when a false-color map looks authoritative. A value on a prospectivity scale is an analytical output, not a drill core, sample or direct confirmation of frozen water. Researchers would still need other observations and measurements to establish what is physically present.
The polar comparison uses a prospectivity scale from 0.0 to 1.0, north and south orientation markers, model labels and a 10 km scale bar. Those visual elements explain how researchers compare mapped patterns; they do not turn the highlighted regions into confirmed ice deposits.
Teaching the model to spot craters and surface change
NASA also describes a fine-tuned demonstration using before-and-after Lunar Reconnaissance Orbiter imagery near Einstein crater. The images show a new impact feature associated with a SpaceX rocket-body impact, alongside existing craters.
This is a useful example of what adaptation can achieve. A model trained on broad lunar observations can be tuned to distinguish a newly formed surface feature from the much older terrain around it. But the demonstration should be read at its actual scale: it shows a supported research workflow for analyzing surface change, not a guarantee of real-time crater detection or mission readiness.
The cyan boxes mark detected crater-like features across the paired lunar views, while a red outline singles out one feature in the right-hand panel. The graphic makes the input-to-analysis relationship clear without establishing a benchmark percentage on its own.
Why young volcanic features matter
Another supported use is mapping irregular mare patches, unusual volcanic features that may be relatively young compared with much of the Moon’s surface. Their distribution can help researchers investigate the Moon’s volcanic and thermal history.
This is where a reusable model becomes especially interesting. The same broad training foundation can be adapted to geological features that are not simply “another crater.” Instead of asking only what shape appears in an image, researchers can use the model to organize large areas of terrain around different scientific questions.
The result is not a machine-generated final answer about the Moon’s past. It is a way to search and compare patterns across a huge archive, giving scientists a faster starting point for deeper investigation.
Useful research infrastructure, not a landing-site clearance
The NASA-IBM Lunar Foundation Model is best understood as an open research and mapping tool. Its practical value lies in making a large, varied lunar archive easier to reuse across tasks:
- estimating polar ice prospectivity;
- mapping craters at different spatial scales;
- detecting surface changes after fine-tuning;
- mapping irregular mare patches and other volcanic features; and
- combining lunar observations from several missions in adaptable workflows.
IBM also reports task-specific comparisons against SwinV2-B. The reported results vary by task and operating conditions: one crater comparison concerns approximately 100-meter-per-pixel data and uses half the training data, while another comparison concerns irregular mare patches. Those figures should not be converted into a blanket claim that the NASA-IBM model outperforms every alternative at every lunar-analysis task.
Nor does the model certify a safe landing site or clear operational hazards. NASA presents it in the context of lunar science and mapping. That boundary matters: a system that highlights promising terrain or a possible new crater is not the same thing as a system authorized to make a flight-safety decision.
For researchers, the release points toward a more reusable way to work with lunar archives. For everyone else, the lesson is refreshingly simple: the model can help scientists decide where to look and what patterns to compare. It cannot, by itself, turn a probability-like map into ice, or a research result into a landing clearance.
