A new machine learning model allows the global research community to map craters and volcanic features across the Moon's surface. Developed by NASA and IBM, the NASA-IBM Lunar Foundation Model was trained on 17 years of observations from the Lunar Reconnaissance Orbiter and other missions using a dataset of 2 million co-registered tile bundles called SomBench. This Vision Transformer encoder-decoder outperforms SwinV2-B baselines in crater detection by nearly 19% at 100 meter scale.
The collaboration follows previous efforts to create the Prithvi Earth-observation and Surya heliophysics models to allow researchers to process petabytes of data without new hardware. This lunar system reduced root-mean-square error for polar ice prospectivity by up to 22% compared to previous benchmarks and is distributed via Hugging Face under an Apache 2.0 license. Fine-tuning code is available on GitHub for use with the TerraTorch toolkit.
Key sources
- SOURCE@nasascience_“trained on 17 years of lunar data to map craters, identify geologic features, and even help predict where ice may hide”x.com
- SOURCE@ibmnews“one of the first publicly available AI models built for scientific lunar exploration”x.com
- SUPPORT@brianroemmele“outperforms them by nearly 19 percent at 100-meter scale while using only half the training labels”x.com