IBM and NASA jointly launch AI model for lunar exploration

IBM and NASA jointly released an open source lunar basic model to support the lunar return program

In preparation for human return to the lunar surface, researchers can now use the open source AI model jointly developed by IBM and NASA-NASA-IBM Lunar Foundation Model. The model is capable of processing decades of lunar observation data to identify ice, craters and volcanic terrain. According to NASA's science department, the system has been available for free and open download on the Hugging Face platform, and its complete code base has also been released on GitHub.

Breaking through the bottleneck of traditional research

For a long time, lunar surface research has been a relatively slow process, relying mainly on two traditional methods: one is to manually consult maps and images, and the other is to Train dedicated narrow-domain, low-resolution machine learning models for each specific need. IBM pointed out in the announcement that these two methods are not only costly to run, but are often difficult to capture the key details scientists need in research.

The release of the basic model effectively solved this problem. Instead of rebuilding models for different geological features, researchers can directly use pre-trained models to respond to new tasks. The lunar model is another important member of the IBM "Prithvi" series of open source scientific models, which has been widely used in fields such as geospatial analysis, meteorology and helio-spherical physics.

"The new system provides scientists with a means to analyze lunar observation data on a large scale, allowing them to identify patterns that are difficult to detect when studying data sets alone," said Juan Bernabe-Moreno, director of IBM's Europe, UK and Ireland Research Institute. He added: "The model provides scientists with the basis for exploring the moon. By connecting observation data from different instruments, it reveals patterns that are difficult to detect from an isolated perspective."

Performance exceeds benchmark

NASA and IBM tested the model in comparison with SwinV2-B. SwinV2-B is a vision system trained by Microsoft and is widely used for baseline assessments of image analysis. The results showed that although the lunar model used only half of the training data, it had a 23% lower error rate in locating ice sediments and was 19% higher than SwinV2-B in discovering and classifying craters. IBM's technical paper noted that when identifying volcanic features known as "Irregular Mare Patches", model performance improved by 3% and required less fine-tuning work.

On August 5, the model passed field testing. At the time, IBM entered an image into the model showing the impact point left behind by the SpaceX Falcon 9 rocket after it hit the moon. Although the new impact trace covered an existing crater almost directly, the model correctly identified it as a newly formed crater.

Goal: Establish a sustainable lunar presence

The ultimate goal is to achieve a sustained human presence on the moon. Currently, mapping permanent shadow areas of the moon's polar regions remains a major challenge because these areas are the most difficult places on the moon to observe but may contain underground ice deposits. IBM said the ice could provide water and oxygen for future lunar bases, and could also be used as a raw material to produce rocket fuel to support Mars missions.

Datasets may be valuable beyond the model itself

In addition, both sides have released what they call the first unified lunar dataset suitable for machine learning. The dataset integrates more than 30 spatially aligned layers collected by nine instruments from four independent lunar missions. It includes data from NASA's Lunar Reconnaissance Orbiter (LRO) and Gravity Recovery and Interior Laboratory (GRAIL) gravity missions, and adds data from Japan's SELENE/Kaguya orbiter.

Bernabe-Moreno said that these approximately two million co-registered data points may be a contribution of more lasting value than the model itself. "Data is the core element that really spawns the AI model industry," he emphasized.

The release of the AI model is based on a partnership spanning more than 50 years, dating back to the Apollo era.