NASA and IBM fold four lunar instrument archives into one reusable AI backbone

Shadow on an airless body is both the nuisance and the signal. Relief is read from shading, and the same low sun that picks out a ridge can flood a crater floor with black that looks like missing data; a frame shot at another hour washes the same geology flat ^[10]. A general vision model has to learn to separate illumination from topography from pixels alone. The Lunar Foundation Model is handed the answer instead: solar angles and spacecraft positions enter the architecture as explicit metadata, so capacity that would go into undoing the lighting can go into the terrain ^[14]. That is a physics prior, and it is the part of the design most likely to earn its keep.

The second engineering problem is scale. NASA's Lunar Reconnaissance Orbiter and its companions deliver regional maps at 100 metres per pixel and terrain scans at 1 metre per pixel ^[9], a factor of 100 in linear resolution and 10,000 in ground area sitting behind each pixel ^[16]. The coarse layers carry ultraviolet reflectance and elevation; the fine ones carry detailed camera frames with meter-scale slope and orientation, and thermal behaviour, surface mineralogy and local gravity anomalies are layered in alongside ^[12].

Training hides parts of a tile and asks the model to predict what was concealed, over millions of examples, until it has learned how lighting, terrain structure and physical geography move together ^[13]. That objective rewards a model for knowing how elevation and visible brightness co-vary across the archive. What it does not tell you is whether the model finds a crater it has never seen, or ice in a floor that is never lit. Reconstruction loss and crater recall are different numbers, and the account of the paper given to Live Science reports neither a task-level score nor a margin over the SwinV2-B backbone it is meant to displace ^[17][3].

The relationships the model absorbs are statistical associations between co-registered instruments ^[13]. That is solid ground for filling in a channel you did not observe. It is thinner ground for the inferential work further down the list, such as collating volcanic features into a story about the moon's thermal evolution ^[8], where the quantity of interest was never in the training tiles at all.

If the lighting prior is doing the work the architecture implies, it should show up as sample efficiency: fewer hand-labelled craters needed to reach a map of given quality than a general-purpose backbone requires. That is a measurable claim about a scarce resource, labelled lunar imagery, and it is the one worth holding the model to.