NASA and IBM's Lunar Foundation Model Advances Moon Science with AI
NASA and IBM Research have collaborated to create the NASA-IBM Lunar Foundation Model, an open-source tool designed to transform decades of lunar observation data into a powerful resource for machine learning. The model, trained on 17 years of data from the Lunar Reconnaissance Orbiter (LRO) and other missions, excels at predicting ice deposits at the lunar poles and detecting craters. Unlike task-specific algorithms, this foundation model is pretrained on large volumes of unlabeled data, making it adaptable to various tasks with minimal labeled examples, ideal for lunar research where labeled data is scarce.
The dataset used to train the model, called SomBench, is the largest co-registered multimodal lunar corpus to date. It includes nearly 2 million tile bundles from 11 modalities and two spatial scales, with high-resolution images from the Narrow Angle Camera (NAC) and multispectral images from the Wide Angle Camera. The model also incorporates data from the GRAIL mission, Lunar Prospector, and JAXA's Kaguya/SELENE probe, totaling over 30 spatially aligned data layers from nine instruments and four missions.
One of the model's key advantages is its ability to account for lunar lighting conditions explicitly, rather than forcing the model to infer them from raw pixels. This approach, combined with a technique called FlexiViT, allows the model to adapt to different image patch sizes without retraining. The model was tested on tasks like crater detection and predicting polar ice deposits, where it outperformed existing baselines, particularly in ice prediction, reducing error by up to 22 percent.
While the model is a valuable tool for analysis, it is not a substitute for physical measurements. The researchers note that it is not suited for absolute geodetic positioning, as latitude, longitude, and elevation values can sometimes be inaccurate. The model is publicly available on Hugging Face, with code on GitHub and integration into the open-source toolkit TerraTorch. This collaboration is part of the broader NASA-IBM 'AI for Science' initiative, which aims to develop foundation models for scientific research.