Back

NASA and IBM Release Open-Source AI Model for Lunar Science

At a glance

  • NASA and IBM released the Lunar Foundation Model on September 10, 2026.
  • The model is available on Hugging Face and GitHub.
  • It was trained on data from four lunar missions and over 2 million images.

NASA and IBM have made a new open-source artificial intelligence model available to the public, designed to support lunar research and exploration by providing advanced analysis of lunar surface data.

The NASA-IBM Lunar Foundation Model, released on September 10, 2026, is accessible through Hugging Face and its full codebase is published on GitHub. The model was developed using a comprehensive lunar dataset that combines more than 30 spatially aligned data layers from nine instruments across four different missions, including NASA’s Lunar Reconnaissance Orbiter (LRO), GRAIL, and JAXA’s SELENE/Kaguya.

Researchers can use this AI model to estimate the likely stability of ice patches on and beneath the Moon’s surface, with a particular focus on permanently shadowed regions near the lunar poles. The model also supports scientific tasks such as crater detection and mapping volcanic features known as irregular mare patches.

The Universities Space Research Association (USRA) contributed to the project by providing expertise in planetary science, dataset development, and scientific evaluation. USRA supported the assessment of the model’s performance in areas such as crater detection, segmentation of irregular mare patches, and regression analysis for polar ice prospectivity.

What the numbers show

  • The model reduces error in predicting lunar ice areas by up to 22% compared to SwinV2-B.
  • It was trained on about 2 million image tiles, including over 1 million high-resolution images.
  • For mapping irregular mare patches, the model outperforms SwinV2-B by approximately 3%.

According to NASA Science, the training dataset for the model includes more than 1 million high-resolution (1 meter) images and nearly 964,000 multispectral images at 100 meter resolution. This diverse dataset allows the model to address a range of lunar science objectives.

IBM stated that the model achieves similar accuracy to the SwinV2-B baseline for crater detection at meter-scale resolution, while requiring only half as much training data and providing greater efficiency. For the task of mapping irregular mare patches, the model demonstrates a performance improvement of around 3% over the baseline.

The model’s open-source availability is intended to support the broader scientific community by enabling further research and development in lunar science. By making the code and datasets publicly accessible, NASA and IBM aim to facilitate collaboration and reproducibility in lunar data analysis.

With the integration of data from multiple lunar missions and instruments, the NASA-IBM Lunar Foundation Model provides a resource for researchers working on lunar geology, ice detection, and related fields. The model’s performance across several benchmarks has been evaluated with input from planetary science experts.

* This article is based on publicly available information at the time of writing.