NASA and IBM Release Open-Source Moon AI Model, Cuts Crater Detection Error by 23%

NASA and IBM have released the NASA-IBM Moon Foundation Model, an open-source AI system available for download on Hugging Face. The model is especially good at spotting areas on the Moon that may contain ice.

In tests, it cut detection errors by 23% compared with Microsoft's SwinV2-B model. For crater detection and classification, it performed 19% better using only half the training data. The team also fed images of a SpaceX Falcon 9 rocket impact into the model, and it correctly identified the new crater.

Training was challenging due to sharp lunar shadows and changing light. The team used a new pie-slice training method. A companion open-source dataset, the first of its kind, includes tens of thousands of images and over 2 million data points.