NASA and IBM release open-source lunar AI foundation model
NASA and IBM have launched an open-source AI foundation model for lunar surface analysis, available on Hugging Face. The model outperforms a standard vision system by 23% in identifying potential ice deposits and by 19% in crater classification while using half the training data. It also correctly identified a new crater formed by a recent rocket impact, demonstrating its practical utility for Artemis-era research.
Lunar imagery poses unique hurdles because the Moon lacks an atmosphere to scatter sunlight, creating pitch-black shadows that obscure data. Standard training methods, which mask image sections for reconstruction, failed since many craters look nearly identical from orbit. Researchers solved this by dividing the lunar surface into wedge-shaped segments, strictly separating training and testing data to ensure consistent visual patterns.
The release follows the successful Artemis II crewed flyby in April 2026, supporting upcoming missions. Alongside the model, NASA and IBM provide a unique dataset for researchers to build their own tools. The model's utility was proven when it accurately flagged a new crater from a SpaceX Falcon 9 impact in August, despite overlapping an existing feature.
This open-source release could democratize lunar research, allowing smaller institutions and independent scientists to analyze orbital data without building expensive proprietary systems. By reducing the training data required and improving accuracy in ice detection, the model may accelerate planning for resource extraction and safe landing zones for Artemis-era missions. However, reliance on AI for identifying geological features could introduce subtle biases, so human verification will likely remain essential for critical mission decisions.