NASA and IBM released the NASA-IBM Lunar Foundation Model, an open-source tool trained on 17 years of lunar observation data, enabling scientists to better predict ice deposits and detect craters on the Moon. It is the first open-source foundation model for lunar science, according to the two organizations.

The model is trained on nearly 2 million tile bundles from 11 modalities and two spatial scales, including high-resolution images from the Lunar Reconnaissance Orbiter (LRO) and multispectral images from the Wide Angle Camera.

The dataset combines more than 30 spatially aligned data layers from nine instruments and four missions, including data from the GRAIL, Lunar Prospector, and JAXA's Kaguya/SELENE probe.

"NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job," said Kevin Murphy, NASA's chief science data officer. "The data also has to be easier for scientists to use." The model uses explicit lighting context, including illumination angles and sun position, instead of guessing these factors from raw images.

The model outperformed existing baselines, particularly in predicting polar ice deposits, where it cut prediction error by up to 22% compared to the best baseline, SwinV2-B. It also showed strong performance in crater detection and segmenting Irregular Mare Patches, which are young volcanic features challenging established cooling timelines.

"The model connects observations across instruments and reveals patterns that are difficult to spot in isolation," said Juan Bernabé-Moreno, director of IBM Research Europe.

"The model isn't suited for absolute geodetic positioning, as it can misrepresent latitude, longitude, and elevation values in some cases." The authors see it as a reusable foundation for downstream tasks, not a replacement for physical measurements.

NASA and IBM did not specify how the model will be integrated into future lunar missions, and some test datasets remain small. The model is publicly available on Hugging Face, with code on GitHub and integration into the TerraTorch toolkit. It is part of the NASA-IBM 'AI for Science' collaboration, which has been developing foundation models since 2022.

Source: thedecoder