NASA and IBM built a free AI to make sense of decades of moon data

NASA and IBM researchers released an open-source "Lunar Foundation Model" on Wednesday, a single AI backbone designed to fuse decades of mismatched lunar sensor data into one reusable map of the moon's surface.

The model, published as a technical paper and posted to GitHub along with its weights and a new benchmark dataset called SomBench, combines high-resolution optical imagery, laser altimetry, radar reflectance, and spectroscopy from past orbital missions. Earlier lunar AI work, including SwinV2-B in 2022, was trained on a single dataset at a time; LFM ingests all four modalities at once.

Named use cases in the paper and NASA's release include crater and landing-zone mapping for crewed and robotic missions, mineralogy inference from crater walls, subsurface ice detection in permanently shadowed regions, and reconstructing volcanic and thermal history. Because the model is open-weight and the training data and code are public, outside teams can fine-tune it for downstream mission tasks without waiting on a NASA-internal pipeline.

The IBM-NASA announcement positions LFM within a broader effort to apply foundation models, a single reusable AI backbone, to scientific data, following Prithvi for Earth observation and similar efforts in biology. Independent validation against ground truth, calibration drift across older sensors, and performance beyond the four named tasks remain open questions.