GoKawiilOpen-source artificial intelligence model combines diverse lunar datasets to support scientific analysis of the Moon
The NASA-IBM model reproduces patterns of lunar ice prospectivity (scaled from blue to yellow), shown at four locations (left) near the Moon’s pole. Top row: reference ice prospectivity map of Mons Mouton near the lunar south pole; middle row: predictions from the ConvNeXt model; bottom row: predictions from the NASA-IBM model. The NASA-IBM model preserves many fine-scale prospectivity patterns in the reference data. Image credit: NASA/IBM Research
WASHINGTON, D.C., — September 18, 2026. Universities Space Research Association (USRA) contributed planetary science expertise, lunar dataset development, and scientific evaluation to the newly released NASA-IBM Lunar Foundation Model, an open-source artificial intelligence (AI) model designed to help researchers analyze the large, diverse datasets collected by lunar missions.
Developed through a collaboration led by NASA and IBM Research, the NASA-IBM Lunar Foundation Model was pretrained from scratch using SomBench, a multimodal lunar dataset containing nearly two million co-registered data bundles spanning 11 modalities and two spatial scales. The model brings together complementary information about the lunar surface, including imagery, topography, illumination geometry, thermophysical properties, mineralogy, radar, gravity, and other geologic and environmental data.
USRA's contribution to the project was provided by Dr. Rachel Slank, an associate scientist with USRA's Science and Technology Institute, on assignment at NASA’s Marshall Space Flight Center. She served as a planetary science subject-matter expert on the NASA-IBM Lunar Foundation Model team. Slank worked across both the science and modeling teams, helping connect lunar science priorities and the physical characteristics of planetary datasets with decisions about model development, applications, and evaluation.
The NASA-IBM Lunar Foundation Model was evaluated across three downstream benchmarks: crater detection at both regional and meter scales, segmentation of irregular mare patches (IMPs), and regression of lunar polar ice prospectivity. Together, these applications assess the model’s performance across a diverse range of lunar science challenges, from identifying impact features and mapping unusual volcanic landforms to integrating environmental datasets associated with the stability and potential distribution of polar volatiles.
Across all three benchmarks, the pretrained NASA-IBM Lunar Foundation Model matched or outperformed comparison models based on ImageNet pretraining, as well as an architecturally identical model initialized without lunar pretraining. The study also demonstrated particularly strong label efficiency in crater detection, suggesting that the representations learned through lunar pretraining can reduce the amount of task-specific labeled data required for certain applications.
The multimodal design of the NASA-IBM Lunar Foundation Model allows it to learn relationships among different types of lunar observations rather than treating each dataset independently. The model was designed to operate across both regional-scale Wide Angle Camera (WAC) observations and meter-scale Narrow Angle Camera (NAC) data while incorporating information such as terrain, illumination geometry, and other lunar surface properties.
By releasing the pretrained model, fine-tuning code, and benchmark datasets openly, the NASA-IBM Lunar Foundation Model team aims to provide the planetary science and AI communities with a reusable foundation for developing new lunar research applications.
A major component of this work was the collaborative development of SomBench, the dataset used both to pretrain the NASA-IBM Lunar Foundation Model and to support standardized evaluation of lunar machine learning (ML) applications.
... continue reading