To read the Moon, NASA and IBM’s AI is told where the Sun is

A crater can remain unchanged while its photograph changes dramatically. Sunlight can pick out a rim, throw part of the interior into shadow or flatten the contrast that makes relief easy to recognise. For a machine learning to interpret the Moon, the conditions of observation are part of the problem.[2][3]

NASA and IBM’s Lunar Foundation Model, announced on 10 September, makes an unusually sensible accommodation: it receives information about the Sun, viewing geometry and image footprint alongside the observations. Researchers have already recorded those conditions. Giving them to the model lets it use that knowledge while learning what the surface looks like.[1][2]

The open release pairs a reusable model with SomBench, a collection that brings disparate lunar measurements into spatial alignment. Two September preprints describe the work. Together, they offer an illuminating answer to a practical question: how can years of orbital observations help a scientist who has only a small collection of carefully labelled examples?[2][3][4]

First, make the observations refer to the same ground

SomBench draws on four missions: Lunar Reconnaissance Orbiter, Japan’s Kaguya, GRAIL and Lunar Prospector. Their instruments describe different properties of the Moon, including reflected light, elevation, temperature, gravity and hydrogen abundance. Those measurements arrive at different resolutions and in different map projections. A useful learning example must establish which pieces describe the same patch of ground.[3]

The SomBench collection contains nearly two million bundles anchored to two kinds of camera image: approximately one metre per pixel for local detail and 100 metres per pixel for wider context. That is a hundredfold difference in linear scale. By simple area arithmetic, one 100-metre pixel covers the same ground area as 10,000 one-metre pixels. Keeping track of those scales matters when combining a detailed crater rim with a much coarser environmental measurement.[2][3]

The two families train a shared model at their native resolutions. Images and terrain information become numerical tokens; observing conditions become a separate contextual input. During pretraining, the system predicts hidden tokens from the information it receives. This gives it repeated opportunities to learn relationships between surface shape, appearance and illumination before anyone asks it to identify a crater.[2]

That distinction is useful beyond lunar science. An image records both an object and a particular encounter between the object, light and an instrument. Here, the researchers can tell the model something about the encounter. They do not have to discard the accompanying measurements and hope it reconstructs them from the picture.

What the crater experiment actually found

The clearest result concerns the wider-scale camera images. Researchers adapted the pretrained model to a crater-detection task with 800 training tiles, 100 validation tiles and 100 test tiles. They compared it with several established image models, including SwinV2-B, whose earlier training used ImageNet. The lunar model, adapted using half the available training data, exceeded the highest-scoring ImageNet baseline in this comparison using the full set.[2]

The practical attraction is the possibility of doing more with labels that take human effort to create. The experiment measures detection performance, rather than hours saved by a mapping team. It also evaluates the combined training recipe: separate experiments are still needed to isolate the contributions of the geometry inputs and mixed-resolution training.[2]

Performance depends on the task. On metre-scale crater detection and the outlining of irregular mare patches, the leading models were close enough that their differences fell within the reported variation between runs. That is useful context for the stronger wider-scale crater result: a lunar model can transfer well without winning every form of lunar mapping.[2]

The map a model is learning to make

A fourth task concerns polar ice prospectivity. Its target is an existing modelled map assembled from terrain, temperature and shadow-related inputs, expressing where ice might occur near the surface. The AI’s success here measures how well it reproduces that map. Establishing the presence of ice requires a different chain of evidence.[2][4]

The release is therefore most useful as a starting point that scientists can adapt to a defined question. The public model card points to the checkpoint and fine-tuning tools, while identifying limits in absolute positioning and generated terrain values. A plausible-looking surface is insufficient when a project needs a reliable elevation or location.[4]

For lunar researchers, the lasting contribution may be the combination: observations that line up, a model that can learn across them, and enough observing context to make shadows informative. The crater rim and the shadow beside it belong in the same scientific picture. They simply tell us different things.

Report an error or suggest a correction