NASA and IBM Release AI Model for Lunar Science

NASA and IBM have released an open-source artificial intelligence model built specifically to study the Moon, giving researchers a reusable way to analyze lunar observations that previously had to be processed instrument by instrument. The NASA-IBM Lunar Foundation Model was released on September 10, 2026, and is designed to identify features such as craters, volcanic formations and potential ice deposits. The interesting part is not simply that AI is being used for lunar imagery; the model combines multiple types of lunar observations so researchers can build specialized tools from the same foundation.

The NASA-IBM Lunar Foundation Model is built for lunar data

The model is a multimodal foundation model, meaning it can learn from several different kinds of data rather than treating every observation as an isolated image. Its training data contains roughly 2 million co-registered lunar data bundles covering 11 modalities at two major image scales. Those inputs include visible and ultraviolet observations, high-resolution imagery, terrain information, slope and aspect measurements, plus additional scientific context gathered from lunar instruments.

The project is part of the broader Prithvi family of open foundation models developed by IBM and NASA for scientific and geospatial applications. Instead of creating a separate machine-learning system for every lunar task, researchers can start with the pretrained model and adapt it to a particular problem. That approach matters because preparing a useful model from scratch for every new scientific question requires substantially more labeled data and training work.

NASA and IBM trained it on decades of observations

The model draws primarily on observations from NASA's Lunar Reconnaissance Orbiter and combines data associated with multiple lunar instruments and missions. NASA describes the release as one of the first open-source foundation models specifically built for lunar science. The accompanying datasets, benchmark collections and software are also being released so researchers can reproduce experiments and develop their own adaptations.

The training setup is unusually specialized. The model was trained using both Lunar Reconnaissance Orbiter Narrow Angle Camera imagery at roughly 1 meter per pixel and Wide Angle Camera imagery at roughly 100 meters per pixel. That is a 100-fold difference in spatial scale, yet the same pretrained system is designed to work across both families of observations. The researchers also feed acquisition information such as illumination geometry into the model because the appearance of lunar terrain can change dramatically depending on where the Sun is relative to the surface.

The strongest results are not equally strong everywhere

IBM and NASA say the model can outperform widely used methods by up to 23% on key lunar-feature identification tasks. That number should not be read as a blanket 23% improvement across every task, however. The released technical evaluation compares the lunar model with several established computer-vision backbones across crater detection, volcanic-feature segmentation and polar ice prospectivity, and the size of the advantage varies substantially between benchmarks.

The detailed benchmark results show why that distinction matters. On one wide-angle crater benchmark using all available training data, the best NASA-IBM configuration reached a mean average precision of 0.2581 compared with 0.2420 for the strongest listed baseline. On the meter-scale crater benchmark, however, the best NASA-IBM result was 0.1543 compared with 0.1552 for the baseline, effectively leaving the two approaches within the uncertainty of the experiment. The model therefore looks more compelling as a reusable lunar foundation than as a system that automatically wins every individual vision task.

Potential ice detection is one of the most useful applications

One of the project's most practical targets is identifying areas where water ice may exist near the lunar poles. Permanently shadowed regions receive little or no direct sunlight, making them difficult to study and especially interesting because water-related resources could matter for future human missions.

The model does not directly measure ice. Its released documentation makes clear that the ice-prospectivity output is a knowledge-driven prospectivity map rather than a direct measurement of lunar ice. That distinction is critical: the AI can help researchers prioritize places for further investigation, but it cannot replace an instrument measurement or establish that a particular location contains usable ice.

Researchers can turn the model into specialized tools

The open release is arguably more significant than any single benchmark because researchers can fine-tune the model for their own scientific tasks. Fine-tuning means taking a pretrained model and training it further on a narrower dataset so that it becomes better suited to a particular job. The NASA-IBM release supports tasks including object detection, image segmentation and dense regression, allowing the same underlying representation to be adapted to different questions about lunar terrain.

The project also supports Low-Rank Adaptation, commonly called LoRA, which updates a relatively small set of additional parameters instead of retraining the entire model. The published experiments found LoRA to be a sensible default for several tasks, while full fine-tuning retained advantages on some smaller benchmarks. That gives researchers a practical choice between a lighter adaptation process and more extensive retraining when the task justifies it.

The model is not ready to certify a landing site

There is a significant limitation buried beneath the excitement around an AI system for lunar exploration: the model is not a replacement for scientific instruments or operational navigation systems. Its documentation explicitly says that generated outputs are not calibrated scientific predictions and that the model does not maintain a geodetic reference frame, which means it cannot reliably provide absolute geographic values required for precision operational decisions.

The researchers also warn that the model has not been validated for operational decisions such as landing-site certification or hazard clearance. Its ice output is a prospectivity estimate rather than measured ice, and its evaluation is limited to the Moon and the products represented in its training and benchmark data. Those restrictions do not make the model less useful; they define the point at which an AI result has to be handed back to conventional scientific measurement.

Why an open lunar model matters beyond this release

Scientific AI has often faced a different problem from consumer AI: researchers may have enormous quantities of data but relatively small amounts of carefully labeled examples for a specific task. A foundation model can reduce that bottleneck by learning broad representations before researchers adapt it to a narrower problem. The NASA-IBM project applies that strategy to a domain where data comes from instruments with different resolutions, wavelengths, viewing conditions and scientific purposes.

That makes the release potentially useful beyond crater maps or ice prospectivity. A researcher studying volcanic terrain, surface composition or another lunar feature can begin from a model that already understands relationships across several forms of lunar observation instead of starting with an untrained vision model. The real test will come as independent teams use the open model on new datasets and tasks, where its limitations and generalization ability can be measured outside the development benchmarks.

The next test is whether researchers can build on it

NASA and IBM have released the model, code, datasets and benchmark material openly enough for researchers to inspect and extend the work rather than treating it as a closed scientific service. That makes adoption and independent validation the next important milestones. If researchers can repeatedly adapt the model to new lunar datasets with less labeled data and lower training cost, its value will extend well beyond the initial benchmark results.

For now, the most defensible description is narrower than the headline suggests: this is an open foundation for lunar remote-sensing research, not an autonomous lunar scientist or a landing-site decision system. Its strongest contribution may be giving researchers a common starting point for combining observations that were previously harder to analyze together. As Artemis missions push toward longer-duration lunar operations, that ability to turn large collections of scientific observations into task-specific tools could become more important than any single accuracy figure.