According to NASA's official website and IBM Research's official social media, the two organizations have released and open-sourced the first lunar model, with model weights and code open to researchers worldwide for download, fine-tuning, and experimentation. This is one of the first publicly available intelligent analysis tools specifically designed for lunar science, aiming to transform decades of scattered lunar observational data into reusable analytical tools for crater mapping, volcanic landform identification, and assessment of polar water ice potential.

The "NASA-IBM model" reproduces patterns of lunar ice detection (from blue to yellow), with the model retaining many fine-scale exploration patterns in the reference data. Image source: NASA/IBM Research

Lunar research has long faced the contradiction of abundant data but difficult integration. The training data for this model primarily comes from NASA's Lunar Reconnaissance Orbiter. LRO has been in orbit for approximately 17 years, covering most of the lunar surface, and its data volume exceeds that of all other NASA planetary missions combined. The model uses approximately 2 million lunar image tiles, including over 1 million 1-meter resolution Narrow Angle Camera images and nearly 964,000 100-meter resolution multispectral images, and integrates terrain and remote sensing data from GRAIL gravity, Lunar Prospector, and Japan's JAXA SELENE missions. The accompanying dataset integrates 9 instruments, 4 missions, and more than 30 spatially aligned data layers, solving the problem of difficult unified access across different resolutions and payloads.

This model is of great significance for lunar resource exploration. Permanently shadowed craters on the Moon have extremely low temperatures that can preserve water ice for millions of years, yet optical methods struggle to observe it directly. The model combines multi-resolution, multi-modal observations to estimate the occurrence and stability zones of polar ice, helping researchers narrow down the scope for field verification. Water ice is not only related to drinking water and breathing oxygen for lunar bases, but can also be electrolyzed to produce oxygen and hydrogen, providing propellant raw materials for deep space missions. This is precisely the starting point for long-term human lunar habitation missions and subsequent missions to Mars.

At the same time, for future spacecraft safe landing and geological understanding of landing sites, the new model can assist in rapidly extracting and classifying craters, reducing the workload of manual interpretation.

The model is incorporated into NASA's Chief Science Data Office "Science AI" strategy and was completed through collaboration among multiple centers, universities, and research institutes. For the academic community, a unified, reproducible lunar dataset is far more important than single algorithm metrics. As future lunar exploration missions continue to generate new imagery, researchers can also fine-tune on the same base model, connecting multi-mission, multi-temporal observations into a consistent understanding rather than starting from scratch each time.

The significance of the "NASA-IBM model" lies in transforming fragmented lunar data into a globally shared intelligent platform, so that scientific discovery no longer depends on data possession but on the ability to make good use of models. In the future, teams from various countries can access cutting-edge research at extremely low cost, and manual interpretation will thus be transformed into a reusable tool. For the scientific community, this will greatly accelerate understanding of craters, volcanoes, and polar resources, accumulating essential knowledge for future lunar bases and even Mars landing missions. This model will likely also be frequently seen in future deep space exploration research, extending from the Moon to Mars and even farther asteroids, driving another leap in humanity's understanding of the universe.