NASA and IBM Release Open-Source Lunar Foundation Model

NASA and IBM have released the NASA-IBM Lunar Foundation Model, an open-source AI system now available on Hugging Face. The foundation model can perform several tasks, with a particular strength in identifying areas on the Moon's surface that may harbor ice.

In benchmark tests, the model's predictions were compared with reference maps built from terrain, thermal and environmental data. Against Microsoft's SwinV2-B vision system, the new model reduced errors by 23 percent and outperformed it by 19 percent in crater detection and classification while using half the training data.

The model was also validated using a real event: after a SpaceX Falcon 9 rocket crashed into the Moon on August 5, IBM fed an image of the impact into the model, which correctly identified the crash site as a new crater on its first attempt, despite its close overlap with an existing crater.

Training the model posed significant challenges. Unlike Earth observation, lunar shadows are pitch black and carry no pixel information. The team found that traditional training methods failed, so they divided the Moon into wedge-shaped regions, completely separating training and testing wedges to provide consistency.

The release also includes a first-of-its-kind open-source dataset, containing tens of thousands of images and instrument data from NASA's Lunar Reconnaissance Orbiter and GRAIL missions, as well as Japan's SELENE mission. The dataset is co-registered in a grid that aligns image pixels with other data modalities, totaling more than two million points, which IBM says will help the community build future lunar models.