NASA, IBM Release Open-Source AI Model For Analyzing Moon Data, With USRA Support

NASA and IBM Research have released an open-source artificial intelligence model designed to help scientists analyze the large and diverse datasets returned by lunar missions, with planetary science support from the Universities Space Research Association (USRA).

The NASA-IBM Lunar Foundation Model was pretrained using SomBench, a multimodal lunar dataset containing nearly two million co-registered data bundles spanning 11 modalities and two spatial scales. The model combines imagery, topography, illumination geometry, thermophysical properties, mineralogy, radar, gravity and other geologic and environmental data describing the lunar surface.

USRA's contribution was led by Dr. Rachel Slank, an associate scientist with the organization's Science and Technology Institute on assignment at NASA's Marshall Space Flight Center. Slank served as the team's planetary science subject-matter expert, working across both the science and modeling groups to connect lunar science priorities and the physical characteristics of planetary datasets with decisions on model development, application and evaluation.

The model was evaluated on three downstream benchmarks: crater detection at regional and meter scales, segmentation of irregular mare patches, and regression of lunar polar ice prospectivity. The tasks span identifying impact features, mapping unusual volcanic landforms and integrating environmental data tied to the stability and distribution of polar volatiles.

Across all three benchmarks, the pretrained model matched or outperformed comparison models based on ImageNet pretraining, as well as an architecturally identical model initialized without lunar pretraining, according to the study. Researchers reported particularly strong label efficiency in crater detection, indicating lunar pretraining can reduce the amount of task-specific labeled data required for some applications.

The model's multimodal design allows it to learn relationships among different types of lunar observations instead of treating each dataset separately. It operates across both regional-scale Wide Angle Camera data and meter-scale Narrow Angle Camera data, while incorporating terrain, illumination geometry and other surface properties.

NASA and IBM released the pretrained model, fine-tuning code and benchmark datasets openly, intending to give the planetary science and AI communities a reusable foundation for developing lunar research applications.

A central part of the effort was development of SomBench, which served as both the pretraining corpus for the model and a benchmark suite for evaluating machine learning methods on lunar science problems spanning three themes: impact processes, volcanic history and polar volatiles.

Slank led development of the high-resolution Lunar Reconnaissance Orbiter Camera (LROC) Narrow Angle Camera crater benchmark, manually identifying more than 49,000 lunar craters. The dataset combines high-resolution lunar imagery with co-registered digital terrain models to evaluate crater detection at meter-scale resolution. She also contributed to the broader SomBench datasets and applications and reviewed both the SomBench and Lunar Foundation Model studies.

"One of the biggest challenges was bringing together lunar datasets that span very different instruments and spatial resolutions, from 1 meter to 20 kilometers per pixel, while still preserving the scientific value of each dataset," said Dr. Slank. "Because I worked across both the science and modeling teams, I could help make sure those differences were considered as the data was brought together and used to train and evaluate the foundation model."

CONTACT: https://huggingface.co/collections/nasa-ibm-ai4science/nasa-ibm-lunar-fm-and-downstream-models