NASA-IBM Lunar AI Maps Ice, Volcanoes, Craters Using 2M LRO Images

NASA-IBM Lunar Foundation Model Overview

NASA, in collaboration with IBM Research and academic institutions, has developed the NASA-IBM Lunar Foundation Model , one of the first open-source AI models purpose-built for lunar science. The model and its training data are publicly available on Hugging Face, allowing researchers to fine-tune it for diverse downstream tasks.

Training Data and Architecture

The model employs a multimodal, multi-resolution design incorporating approximately 11 data modalities at two spatial scales: 1-meter and 100-meter resolution. Training data originates from the Lunar Reconnaissance Orbiter (LRO) collected over 17 years , forming a dataset larger than all other NASA planetary missions combined. The dataset comprises roughly 2 million image tiles , including over 1 million 1-meter resolution Narrow Angle Camera (NAC) images and nearly 964,000 100-meter resolution multispectral images . Additional data sources include NASA's GRAIL (gravity), Lunar Prospector , and JAXA's SELENE (topography).

The model uses a ViT-B encoder–decoder architecture pre-trained via masked reconstruction, learning to infer missing information from diverse inputs. The training incorporates observational geometry parameters such as illumination angle, solar reference coordinates, and geographic extent.

Downstream Lunar Science Tasks

After pre-training, the foundation model is adapted for three challenging lunar science tasks:

Polar ice mapping : Identifying potential surface and subsurface ice in permanently shadowed regions (PSRs) near the lunar poles, where ice can persist for billions of years. The model reproduces ice stability distribution maps at four polar locations.

Young volcanic structure detection : Locating irregular mare patches (IMPs) , rare volcanic features that appear geologically young and challenge existing models of lunar thermal evolution.

Crater detection and mapping : Cataloging impact craters whose size-frequency distributions enable surface age dating and reconstruction of early solar system history.

Performance and Open-Source Release

Benchmarking against several strong baselines shows the NASA-IBM model achieves comparable or superior performance . It matches baselines on crater mapping and IMP segmentation, while demonstrating a clear advantage in estimating polar ice stability .

To support the global research community, the team released:

Comprehensive machine-learning-ready pre-training datasets and benchmark suites.

Integration into the open-source toolkit TerraTorch .

A companion paper on Hugging Face ensuring reproducibility and enabling scientists to build, compare, and optimize future lunar exploration AI models.

Broader NASA-IBM AI4Science Collaboration

The lunar foundation model is part of NASA's AI4Science (AI4S) strategy led by the Office of the Chief Science Data Officer. Other models in this ongoing partnership include:

Prithvi : Pre-trained on Earth observation data for disaster detection, flood mapping, and crop yield prediction.

Surya : Trained on high-resolution solar observation data to forecast space weather events such as solar flares.

Original article:

https://science.nasa.gov/science-research/artificial-intelligence-lunar-foundation-model

Model collection:

https://huggingface.co/collections/nasa-ibm-ai4science/nasa-ibm-lunar-fm-and-downstream-models