NASA and IBM Release Open-Source AI Model for Lunar Exploration

The NASA-IBM Lunar Foundation Model represents a shift in how researchers process planetary data. Rather than analyzing imagery, topography, and mineralogy in isolation, the system uses a multimodal approach to find relationships across 11 different data types. At the heart of this project is SomBench, a massive training corpus that spans spatial scales from regional Wide Angle Camera observations to meter-scale Narrow Angle Camera imagery.

Dr. Rachel Slank of the Universities Space Research Association served as a key bridge between the science and modeling teams at NASA’s Marshall Space Flight Center. Her work included the manual identification of over 49,000 lunar craters to create a high-resolution benchmark for the model. This manual labor proved effective; tests show the model requires less labeled data than traditional architectures to achieve high accuracy in crater detection, irregular mare patch segmentation, and polar ice prospectivity analysis.

By releasing the pretrained model, fine-tuning code, and the SomBench dataset on Hugging Face, the team intends to lower the barrier for planetary scientists looking to apply machine learning to their research. The model consistently outperformed traditional ImageNet-based benchmarks, proving that pretraining on lunar-specific data is essential for mapping the complex surface environments of the Moon.