NASA and IBM Launch Open-Source Lunar AI Model to Map Moon’s Ice, Craters and Volcanic Features

NASA-IBM Lunar Foundation Model Turns Decades of Moon Data Into an Open AI Research Platform

This makes the release different from conventional consumer AI tools. Instead of generating text, images or software code, the model is designed to extract useful scientific information from remote-sensing data.

NASA says the project demonstrates how AI can make its large scientific datasets easier for researchers to explore and use.

What Is the NASA-IBM Lunar Foundation Model?

The NASA-IBM Lunar Foundation Model is a multimodal, multi-resolution AI foundation model for lunar remote sensing.

It was developed through collaboration between NASA, IBM Research and academic research organizations. NASA describes it as one of the first open-source AI models specifically built for lunar science.

The model was primarily trained using observations collected by NASA's Lunar Reconnaissance Orbiter, along with data from other lunar missions.

The training data contains roughly 2 million image tiles, including more than 1 million high-resolution camera images at approximately 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. Data from missions including NASA's GRAIL and Lunar Prospector and Japan's SELENE/Kaguya also contribute to the broader dataset.

IBM and NASA also created a unified machine-learning-ready lunar dataset containing more than 30 spatially aligned layers from nine instruments across four missions.

That multi-source approach is one of the most important aspects of the project.

Why Combining Lunar Data Matters

Scientists have collected Moon observations for decades, but those observations can come from different instruments, missions, resolutions and measurement systems.

Analyzing each dataset separately can make it harder to see relationships between different observations.

The NASA-IBM model attempts to create a common representation of these different measurements.

IBM describes the system as a way to consolidate observations across modalities, viewing angles and spatial scales so researchers can reuse the model for different scientific tasks.

This is where foundation models become particularly useful.

Rather than creating an entirely new machine-learning system for every research question, scientists can start with a pretrained model and adapt it to a specific task.

What Can the Lunar AI Model Do?

NASA and IBM initially highlight several important applications for the model.

Detecting Previously Uncatalogued Craters

The Moon's surface contains an enormous number of impact craters.

Some are already catalogued, while smaller or difficult-to-identify formations can require extensive analysis.

The model can help researchers detect and map lunar craters more efficiently.

According to the model's public documentation, it can be adapted for crater-detection tasks using lunar remote-sensing imagery.

This could help scientists build more detailed maps of the lunar surface and improve understanding of its geological history.

Studying Volcanic Features

The Moon also contains evidence of ancient volcanic activity.

Researchers can use the model to investigate volcanic formations and geological structures, helping scientists understand how the lunar surface evolved over time.

IBM says studying the Moon's volcanic history is one of the initial priorities for the model.

AI can potentially accelerate this work by allowing researchers to analyze large areas of imagery and identify patterns that deserve closer scientific examination.

Searching for Potential Lunar Ice

One of the most strategically important applications involves lunar ice.

Scientists are particularly interested in permanently shadowed regions near the lunar poles because water resources could be valuable for future human exploration.

Water could potentially support drinking-water supplies, oxygen production and fuel-related processes after further processing.

The NASA-IBM model includes an ice-prospectivity application designed to help researchers analyze areas associated with potential ice resources.

However, an important limitation should be understood: the model's ice-prospectivity output is not a direct measurement of lunar ice.

The public model documentation explicitly says it should not be treated as a scientific-grade measurement or a replacement for instruments and other geodetic methods.

That distinction is important because AI predictions can help scientists prioritize locations for further investigation, but they do not replace physical measurements.

NASA-IBM Model Shows Up to 23% Higher Accuracy in Key Tasks

IBM and NASA say the model exceeded widely used methods by up to 23% when identifying important geographic features on the Moon, including potential ice deposits, craters and volcanic formations.

Reuters also reported that benchmark testing showed the model identifying key lunar features with up to 23% higher accuracy than existing methods.

The figure should not be interpreted as meaning that the model is 23% more accurate at every possible lunar-science task.

The actual performance depends on the task, dataset and evaluation methodology.

The public model card provides more detailed benchmark results across crater detection, segmentation and ice-prospectivity tasks.

The Model Is Open Source and Available to Researchers

One of the most significant parts of the release is its openness.

NASA says the model is publicly hosted on Hugging Face, while its complete codebase is available through GitHub. The project also includes machine-learning-ready datasets and benchmark collections.

The model is released under the Apache 2.0 license according to its Hugging Face model card.

This gives researchers and developers an opportunity to experiment with the technology rather than simply consuming it through a closed API.

The model can be used with the open-source TerraTorch toolkit for fine-tuning and downstream adaptation.

Researchers Can Adapt the Model for Different Tasks

The model was built to provide reusable representations of lunar remote-sensing data.

Its documentation describes applications including:

Crater detection
Image segmentation
Dense regression
Ice-prospectivity analysis
Multimodal lunar-data research

The model also supports different modality combinations during adaptation, making it possible for researchers to experiment with different forms of lunar information.

This flexibility could be particularly useful for universities and research organizations that want to build specialized lunar-analysis systems without starting from zero.

How the Lunar Model Differs From General AI Models

The NASA-IBM model represents a different direction for foundation-model development.

Most consumer-facing AI models focus on language, images, video, audio or software development.

The lunar model instead focuses on scientific observations.

It is designed around the specific characteristics of remote-sensing data and lunar geography.

That specialization matters because scientific AI often requires models to understand relationships between different measurements rather than simply predict the next word or generate an image.

The project also shows how foundation models are expanding into specialized scientific fields.

NASA has already explored AI foundation models for Earth observation, and the lunar model extends this approach to another planetary environment.

What This Means for Future Moon Missions

The model could eventually become part of a larger scientific workflow supporting future lunar exploration.

Better maps of craters and terrain could help scientists understand the Moon's surface and identify areas requiring additional investigation.

Potential ice mapping could also help researchers study lunar resources.

NASA's broader Artemis strategy makes these capabilities particularly relevant because future missions require increasingly detailed knowledge of the lunar environment.

Reuters reported that NASA's Artemis plans target a return of astronauts to the Moon in 2028, with longer-term goals involving sustained lunar exploration and future Mars missions.

AI cannot solve all of the challenges associated with those missions, but it can help researchers process the enormous amount of scientific information required to plan them.

The Project Is Part of a Larger Shift Toward Scientific AI

The NASA-IBM release also reflects a broader change in the AI industry.

Foundation models are increasingly being developed for specialized domains rather than only general-purpose chatbots.

TheInfoBytes has already covered other open-model developments, including IFM's K2 Horizon open AI model family and Tencent's Hy4 open-source AI model.

The NASA-IBM project takes that trend into scientific research.

Instead of optimizing an AI system primarily for conversations or coding, researchers are building a reusable foundation for analyzing a specific scientific environment.

That could become an increasingly important model-development strategy as organizations apply AI to biology, climate science, astronomy, materials research and planetary science.

What Developers and Researchers Can Access

Researchers interested in experimenting with the model can access the public model through Hugging Face.

The project provides TerraTorch integration, pretrained model files, documentation and downstream adaptation resources.

NASA also says the complete codebase and supporting resources are available publicly for experimentation and research.

This makes the release more useful than a conventional research announcement because outside researchers can inspect the model, reproduce experiments and build their own adaptations.

Important Limitations to Keep in Mind

Despite the promising benchmark results, the NASA-IBM Lunar Foundation Model is not a replacement for scientific instruments.

Its documentation states that it is not validated for operational decisions such as landing-site certification or hazard clearance. It also does not maintain a geodetic reference frame and should not be treated as producing calibrated scientific measurements.

That means the model should be viewed as a research and analysis tool.

It can help scientists identify patterns, prioritize areas and develop new research workflows, but important mission decisions still require validated scientific measurements and other specialized systems.

This distinction will become increasingly important as AI becomes more common in scientific research.

Why the NASA-IBM Lunar AI Model Matters

The biggest importance of the NASA-IBM Lunar Foundation Model is not simply that NASA is using AI to study the Moon.

It is the combination of open AI, scientific data and reusable foundation-model technology.

The project turns millions of lunar observations into a machine-learning resource that researchers can adapt to multiple problems.

Its ability to work across different data types and resolutions could make it easier to investigate questions that would otherwise require separate specialized models.

The open release also gives the wider scientific community an opportunity to improve the technology and create new applications.

For AI developers, the project is another example of how foundation models are moving beyond general-purpose chatbots.

For space researchers, it provides a new way to work with decades of lunar observations.

And for the broader AI industry, it demonstrates how specialized foundation models could become important tools for scientific discovery.

Frequently Asked Questions

What is the NASA-IBM Lunar Foundation Model?

It is an open-source multimodal AI foundation model developed by NASA and IBM for analyzing lunar remote-sensing data and supporting scientific research on the Moon.