NASA and IBM Research have launched the NASA-IBM Lunar Foundation Model, an open-source AI tool designed to transform the analysis of lunar surface data. While NASA has previously integrated AI into orbital operations, this release marks a shift toward providing the global scientific community with a pre-trained foundation model capable of generalizing across multiple lunar research domains through minimal fine-tuning.
The model was trained on a massive dataset comprising roughly 2 million image tiles, primarily sourced from the Lunar Reconnaissance Orbiter (LRO). This includes over 1 million high-resolution images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. To ensure comprehensive coverage, the training set also incorporated data from NASA's GRAIL and Lunar Prospector missions, as well as JAXA's Selenological and Engineering Explorer.
From Manual Mapping to Automated Discovery
Traditionally, identifying lunar features required scientists to manually sift through petabytes of data or use narrow machine-learning tools. The new foundation model automates these processes, allowing researchers to rapidly map craters and identify "irregular mare patches"—young volcanic features that challenge current theories on the Moon's thermal evolution.
Scraped da https://science.nasa.gov/science-research/artificial-intelligence-lunar-foundati
A critical application is the estimation of polar ice stability. The model excels at identifying stable ice patches in permanently shadowed regions, a key requirement for planning future human habitats. In benchmark tests, the NASA-IBM model outperformed other baselines specifically in ice prospectivity mapping.
Detecting Surface Changes
The model's versatility is demonstrated by its ability to recognize novel surface changes. In a test involving the impact of a SpaceX rocket body—an event previously reported by AlexTech.ai—the model successfully detected the newly formed crater despite the post-impact image being excluded from its initial pre-training. This capability allows scientists to automatically monitor natural impacts and anthropogenic changes across the lunar surface.
The project is part of a broader AI-for-science strategy, joining the Prithvi models for Earth observation and the Surya model for heliophysics. The complete codebase is available on GitHub to encourage global experimentation.
No comments yet. Be the first!