NASA and IBM have presented the Lunar Foundation Model – an open artificial intelligence model for analyzing the lunar surface. It can be downloaded via Hugging Face and used, in particular, to search for areas where water ice may potentially be present. IBM reports this.
At first glance, this looks like an ordinary news story about yet another specialized neural network. But the practical angle here is different: NASA is essentially giving outside researchers not a ready-made answer of "where to search for ice," but rather a tool and raw data that can be questioned. This is fundamentally different from closed models, where the user only sees the result and is forced to trust it blindly.
Why old computer vision methods don't work directly here
During testing, the model showed better results than SwinV2-B – Microsoft's computer vision system. In the task of searching for potential ice deposits, it reduced the number of errors by 23%, and when detecting and classifying craters, it outperformed SwinV2-B by 19%, using half as much training data.
Model development was complicated by the specifics of lunar imagery. Due to the lack of an atmosphere, shadows on the Moon are very sharp, and the surface illumination changes depending on the Sun's position. Because of this, the same crater can look substantially different in different images. To avoid training problems, researchers divided the lunar surface into sectors and used different areas for training and testing.
This is not a cosmetic detail. Earth-based computer vision models are trained on light scattered by the atmosphere, where shadows are soft and predictable. The Moon breaks this logic: what looks like a "new texture" to a conventional algorithm may actually be the same crater at a different time of the "lunar day."
The model reduced the number of errors in ice searches by 23% compared to Microsoft's solution, using half as much training data.
The key is not the model, but the data behind it
Along with the model, NASA and IBM released an open dataset that can be used by other researchers to create their own AI models. It combines tens of thousands of images and instrumental data from NASA's Lunar Reconnaissance Orbiter (LRO) and Grail missions, as well as Japan's Selene. The data was synchronized across different types of measurements to correspond to the same areas of the lunar surface.
It is precisely the synchronization of three independent missions – American and Japanese – that transforms this release from another demo into a working tool. Any team planning a landing or drilling operation can now verify the model's prediction against primary data, rather than simply accept the conclusion "there is probably ice here" on faith.
The question is whether private lunar programs will take advantage of this: if an open model begins to influence the actual choice of landing sites for future resource extraction missions, it will become the first example of open-source AI directly determining the geography of the space race for the Moon.
- Recall that Artemis II astronauts filmed Earth on an iPhone 17 Pro Max.
- The secret US aircraft NT-43A RAT55 tracked NASA's Artemis II mission, which went beyond Earth's orbit.
- On April 1, NASA launched the crewed Artemis II mission around the Moon. The Artemis II space mission will also be featured in a live broadcast on Netflix.
