The Moon has been photographed in extraordinary detail, yet turning that vast image archive into useful scientific knowledge remains a formidable task. Every crater rim, boulder field, landslide and permanently shadowed slope may contain information about lunar history—or affect where a future spacecraft can safely land. NASA and IBM are applying a new kind of artificial intelligence to that challenge: a Lunar Foundation Model designed to learn the visual language of the Moon.
Rather than building a separate algorithm for every mapping job, the NASA IBM lunar AI model can serve as a reusable base for multiple scientific tasks. After learning patterns from large collections of lunar observations, it can be adapted to classify terrain, outline craters, flag unusual surface changes and support searches for scientifically important features. That combination makes AI Moon mapping more than an exercise in image recognition. It could become part of the digital infrastructure behind Artemis and future robotic missions.
The most compelling capability is also the easiest to understand: the model can identify and map features across the lunar surface much faster than researchers could annotate them manually. Its value, however, depends on what it has learned, how scientists validate its results and whether the wider research community can build on it.
What Is the NASA IBM Lunar Foundation Model?
A foundation model is an AI system pretrained on a broad dataset so that it can later be adapted to many related applications. Large language models apply this concept to text. Geospatial foundation models apply it to satellite and aerial observations. The Lunar Foundation Model brings the same approach to imagery and data from the Moon.
The project extends NASA and IBM’s wider collaboration on scientific foundation models. Instead of treating each orbital image as an isolated picture, the lunar AI model learns recurring structures, textures and spatial relationships. These may include the circular geometry of impact craters, the sharp contrast around young ejecta, the subtle flow patterns associated with mass wasting and the differences between smooth maria and rugged highlands.
This does not mean the NASA AI model understands the Moon in the way a planetary geologist does. It learns numerical representations that capture useful patterns in lunar data. Scientists can then fine-tune or evaluate those representations for a defined task, such as crater segmentation or terrain classification.
Much of the foundational observation record comes from instruments aboard lunar spacecraft, including NASA’s Lunar Reconnaissance Orbiter. LRO has spent years collecting high-resolution images, elevation measurements, temperature data and other observations. Foundation-model techniques offer a way to extract more value from that expanding archive.
How a Lunar AI Model Learns to Read the Moon
Traditional supervised AI requires large numbers of examples labeled by people. For Moon crater mapping, researchers might need to draw boundaries around thousands of craters before a model can reliably recognize them. That process is expensive, slow and sometimes subjective, especially when features are eroded, overlapping or partly hidden by shadows.
The NASA IBM Lunar Foundation Model uses the more scalable logic of self-supervised pretraining. During this phase, portions of an image can be masked, transformed or compared with neighboring observations. The model learns to reconstruct missing information or recognize relationships without requiring a human label for every pixel. In effect, the imagery itself supplies the training exercise.
Several characteristics make lunar training unusually demanding:
- Extreme lighting: Low Sun angles can create long shadows that resemble geological boundaries, while polar darkness can obscure surface details.
- Different image resolutions: Orbital instruments capture the Moon at varying scales, so a model must distinguish a small nearby crater from a larger distant one.
- Overlapping geological features: New impacts, old basins, ejecta and volcanic deposits can occupy the same area.
- Limited ground truth: Only a tiny fraction of the Moon has been directly examined at the surface.
- Sensor differences: Optical images, radar, thermal measurements and elevation models describe different physical properties.
Once pretrained, the lunar foundation model can be fine-tuned with a smaller labeled dataset. A research team could teach it to recognize a particular terrain type without starting from an untrained network. This transfer-learning approach is one reason foundation models are attracting attention across AI space exploration.
What NASA and IBM’s AI Can Find on the Lunar Surface
The central promise of NASA AI lunar mapping is automated feature identification at planetary scale. A model can scan consistent grids of orbital data, propose boundaries and rank detections for expert review. That does not replace planetary scientists; it gives them a faster way to find the locations that deserve attention.
Impact craters and crater boundaries
Craters are essential records of lunar history. Their size, density and condition help researchers estimate the relative age of a surface. Moon crater mapping is difficult because old craters may be eroded, filled by later material or cut by newer impacts. AI can detect circular and partially circular patterns while also using surrounding texture to separate crater rims from unrelated shadows.
A well-adapted model could produce more consistent crater catalogs, locate small craters missed in earlier surveys and compare images captured at different times. Change detection may also help identify fresh impact sites, which reveal information about present-day impact rates and expose material beneath the weathered surface.
Terrain, slopes and landing hazards
Future landers need more than a photograph of their destination. Mission teams must understand slope, surface roughness, boulder distribution and the geometry of nearby craters. IBM lunar AI could assist by segmenting an area into terrain classes and highlighting zones that warrant higher-resolution analysis.
When combined with digital elevation models, AI lunar mapping can help distinguish relatively level ground from scarps, steep crater walls or irregular ejecta. It can also help prioritize candidate landing ellipses. Final site certification still requires engineering analysis and multiple datasets, but automated screening can narrow an enormous search area.
Landslides and other unusual formations
Lunar regolith moves downhill through impacts, moonquakes and repeated temperature changes. These movements can leave flow-like patterns on crater walls and slopes. Because a foundation model learns broad visual representations rather than one rigid shape, it may be adapted to identify landslides, rockfalls, channels and other uncommon formations.
Anomaly detection is particularly valuable. Instead of asking the model to find only a known category, scientists can use it to flag areas that look statistically different from their surroundings. Some alerts will be imaging artifacts or ordinary geology, but the best candidates can direct experts toward overlooked features.
Could Lunar Ice Detection AI Find Water on the Moon?
Water ice is one of the most consequential targets in lunar exploration. Deposits may preserve a record of material delivered by comets and asteroids. In the longer term, accessible water could support life-support systems or be separated into hydrogen and oxygen. The most promising regions are permanently shadowed areas near the lunar poles, where temperatures can remain low enough for volatile compounds to survive.
However, an optical AI model cannot simply look into a dark crater and confirm that it contains ice. Reliable assessment requires evidence from several sources, potentially including radar signatures, neutron measurements, thermal data, illumination models and surface imagery. This is where a multimodal lunar foundation model could become especially useful.
Lunar ice detection AI can combine spatial clues, rank locations with ice-compatible conditions and identify relationships that may be difficult to see in one dataset. Its output should be understood as a probability map or exploration priority—not proof of a deposit. Confirmation requires targeted remote sensing or direct measurements by a lander, rover or drill.
How AI Moon Mapping Could Support Artemis
NASA’s Artemis campaign is focused on sustained exploration, with significant attention on the Moon’s south polar region. That area is scientifically compelling but operationally difficult. The Sun stays close to the horizon, shadows shift across rugged terrain and communications may be obstructed by local topography.
Artemis AI tools built from a lunar foundation model could help mission planners compare landing zones, map traversable corridors and identify targets for astronauts or robotic scouts. Faster processing would also make it easier to update maps when new orbital observations arrive.
Potential applications include:
- Ranking candidate landing areas using terrain and hazard indicators
- Mapping crater rims, boulders, slopes and shadow boundaries
- Identifying locations where multiple datasets suggest possible volatiles
- Planning rover routes between scientific targets and illuminated areas
- Comparing repeat images to detect surface changes
- Compressing or prioritizing observations for transmission to Earth
The distinction between decision support and autonomous control is important. As of September 2026, a general lunar AI model should not be assumed to be a flight-certified navigation system or an independent mission controller. Operational software must meet strict reliability, testing and cybersecurity standards. The nearer-term opportunity is to strengthen the maps and analysis that humans use to make mission decisions.
Why Open-Source Lunar AI Matters
An open source lunar AI model can lower the barrier to advanced planetary research. Universities, startups and space agencies would not need to reproduce the cost of large-scale pretraining before testing a new scientific idea. They could adapt shared model weights, code and evaluation methods to regional mapping, landing-site studies or instrument-specific data.
NASA’s open-science ecosystem, including repositories maintained through NASA IMPACT, shows how public code and reproducible workflows can broaden participation in Earth and space science. For the Lunar Foundation Model, useful openness includes more than publishing code. Researchers also need documentation, training-data descriptions, benchmark tasks, model limitations and clear licenses.
Shared benchmarks are particularly important. If teams test models on different regions using different definitions of a crater or hazard, performance numbers are difficult to compare. Common datasets can reveal whether an apparent improvement works across lighting conditions, geological regions and image resolutions.
Foundation Models Are Becoming Tools for Space Science
The NASA IBM lunar AI project reflects a larger shift from single-purpose algorithms toward reusable scientific models. Similar techniques are being developed for Earth observation, solar activity, weather, astronomy and spacecraft telemetry. Each domain has vast datasets that are too large for purely manual analysis but too specialized for general consumer AI.
Future AI for space exploration is also likely to be multimodal. A single system may learn relationships among images, elevation, radar, temperature, mineral spectra and scientific text. Smaller versions could eventually run on spacecraft or surface robots, allowing them to prioritize observations when bandwidth is limited.
The key advantage is not that one model will answer every question. It is that researchers can begin with a representation already grounded in the physical patterns of a planetary environment. That can reduce training requirements and make specialist tools available to more teams.
Limits Scientists Must Solve Before Trusting Lunar AI
Lunar imagery contains traps for machine learning. Shadows can be mistaken for crater edges, image seams can appear geological and a model trained heavily on one region may perform poorly elsewhere. Rare but mission-critical hazards may also be underrepresented in training data.
For those reasons, NASA AI lunar mapping requires uncertainty estimates and expert validation. Model predictions should be compared with independent observations, tested across different illumination conditions and reviewed for geographic bias. Scientists also need ways to understand why a location was classified as a crater, hazard or possible volatile-rich area.
Human oversight is not a weakness of the approach. It is the practical model for scientific AI: machines search and organize at scale, while experts test hypotheses, reject false detections and determine what the evidence means.
Frequently Asked Questions
What is the NASA IBM Lunar Foundation Model?
It is a foundation-model approach for learning patterns from large collections of lunar data. After broad pretraining, the model can be adapted to tasks such as crater detection, terrain classification, feature segmentation and surface-change analysis. Its reusable design distinguishes it from an algorithm built for only one mapping job.
Can the lunar AI model detect water ice directly?
Not from ordinary surface images alone. AI can combine or compare radar, thermal, neutron, topographic and optical evidence to identify locations with conditions consistent with ice. Those predictions guide investigation, but direct detection requires suitable instruments and scientific confirmation.
Will AI choose where Artemis astronauts land?
AI may help experts evaluate terrain, hazards and scientific opportunities, but landing-site selection is a multidisciplinary human decision. Mission teams must consider spacecraft performance, lighting, communications, navigation, safety and program goals in addition to AI-generated maps.
Why make a lunar AI model open source?
Open access allows researchers to reproduce results, inspect limitations and adapt the model without repeating expensive pretraining. It can also encourage shared benchmarks and new applications from universities, commercial space companies and international science teams.
Is the Lunar Foundation Model useful beyond the Moon?
The specific learned features are lunar, but its methods can inform models for Mars, asteroids and other planetary bodies. More broadly, NASA and IBM’s work demonstrates how domain-specific foundation models can organize large scientific archives and turn observations into usable maps.
A New Layer in the Moon’s Digital Map
The Lunar Foundation Model will not eliminate the need for geologists, mission engineers or dedicated instruments. Its importance lies in scale. It can examine more data, apply consistent criteria across wide areas and reveal candidates that humans can investigate more closely.
If the model remains open, transparent and carefully validated, it could become a shared layer in the technology stack for Moon exploration. Better crater catalogs, terrain maps, hazard assessments and ice-prospecting priorities would benefit robotic missions and Artemis alike. NASA and IBM are ultimately teaching AI to recognize the Moon’s visual patterns—and giving scientists a faster way to decide where to look next.
