NASA and IBM Release Open-Source Lunar Foundation Model, Improving Polar Ice Prediction by 22%

NASA and IBM Research, in collaboration with academic institutions, have released the open-source NASA-IBM Lunar Foundation Model, an AI foundation model for lunar science. Trained on 17 years of observations from the Lunar Reconnaissance Orbiter (LRO) and other missions, the model and its machine-learning-ready dataset are now openly published on Hugging Face. For broader context, explore our Top 100 AI Tools.

A Foundation Model Built on Decades of Lunar Data

The NASA-IBM Lunar Foundation Model was developed from scratch, utilizing the TerraMind multimodal Earth-observation architecture. Its training leveraged SomBench, a comprehensive corpus of nearly 2 million tile bundles across 11 modalities. This extensive dataset primarily incorporates 17 years of observations from NASA's Lunar Reconnaissance Orbiter (LRO), supplemented by data from GRAIL, Lunar Prospector, and JAXA's Kaguya/SELENE missions.

The model's ability to process imaging geometry, such as illumination angles and sun position, as explicit input enhances its analytical capabilities for lunar surface features.

Enhanced Accuracy in Lunar Analysis

Initial evaluations demonstrate the model's improved performance in key areas of lunar analysis. Compared to existing baselines, the NASA-IBM Lunar Foundation Model reduces polar ice-deposit prediction error by up to 22 percent. Furthermore, it improves coarse crater detection by nearly 19 percent, offering more precise insights into the Moon's geological history and potential resources.

The SomBench Dataset: A Rich Resource for Researchers

The accompanying SomBench dataset is a critical component of this release, providing a robust foundation for further research and development. It comprises nearly 2 million tile bundles across 11 modalities and two distinct spatial scales. Specifically, the dataset includes approximately 1 million Narrow Angle Camera images, offering a resolution of about 1 meter per pixel, and around 964,000 multispectral Wide Angle Camera images, with a resolution of 100 meters per pixel.

This comprehensive dataset spans over 30 spatially aligned data layers, collected from nine instruments across four different lunar missions. Its open availability on Hugging Face aims to foster collaborative research and accelerate discoveries in lunar science.

Implications for Lunar Exploration and Science

The release of an open-source lunar foundation model has broad implications for the scientific community. By providing a powerful, pre-trained AI model and a vast, curated dataset, NASA and IBM are enabling researchers worldwide to conduct more efficient and accurate analyses of lunar data. This could accelerate discoveries related to lunar geology, resource identification, and the planning of future lunar missions.

The open-source nature of the model encourages transparency and collaboration, allowing scientists to build upon the existing framework and contribute to its further development. This approach aligns with the principles of open science, making advanced AI news and tools more accessible.

Conclusion

The NASA-IBM Lunar Foundation Model represents a significant advancement in the application of artificial intelligence to planetary science. By combining extensive lunar observational data with a sophisticated AI architecture and making it openly available, NASA and IBM Research are providing a valuable resource for the global scientific community. This initiative is poised to enhance our understanding of the Moon and support future endeavors in lunar exploration.

Sources

• nasa-ibm-ai4science/NASA-IBM-Lunar-Foundation-Model · Hugging Face
• NASA-IBM-Lunar-Foundation-Model/NOTICE at main · NASA-IMPACT/NASA-IBM-Lunar-Foundation-Model · GitHub
• 📰 Hacker News AI Digest 2026-09-19 · Issue #3371 · duanyytop/agents-radar · GitHub
• Hacker News AI Digest 2026-09-20 · Issue #3384