IBM Releases Open Source Lunar Foundation AI Model with NASA

IBM and NASA have launched an open-source artificial intelligence system called the NASA-IBM Lunar Foundation model. Available through the Hugging Face platform, the system aims to support researchers in analyzing decades of geological data gathered from the Moon's surface. Alongside the release, the organizations are providing public access to a unified, machine-learning-ready dataset containing thousands of lunar maps and imagery layers.

Previously, planetary scientists relied heavily on manual data evaluations or smaller, task-specific models. These traditional tools were computationally intensive and often lacked the precision required for complex geographical studies. By contrast, the newly released foundation architecture allows scientists to fine-tune a single core system across a variety of research tasks, eliminating the necessity of engineering specialized models from scratch for every distinct surface mapping challenge.

The newly accessible dataset combines over 30 spatially aligned layers derived from nine separate instruments across four historical space missions. Key inputs include tens of thousands of geophysical maps and high-resolution images gathered by instruments on NASA’s Lunar Reconnaissance Orbiter and the GRAIL mission. Combining these disparate sources into a standardized, machine-ready repository permits advanced multimodal training across disparate planetary science datasets.

Initial testing reveals the model improves performance when detecting geological structures such as craters and Irregular Mare Patches, which are key volcanic formations. It also decreases error rates when identifying subsurface ice deposits. Pinpointing ice locations is critical for long-term space exploration, as water and oxygen reserves are necessary to support personnel on the Moon and manufacture fuel for potential future Mars missions.

Juan Bernabe-Moreno, Director of IBM Research Europe, UK and Ireland, stated that the foundation model provides a scaleable base for connecting observations across scientific instruments and identifying subtle patterns. IBM, which has collaborated with NASA for more than 50 years, expects the technology will assist in identifying safe landing sites free from hazardous boulders or steep slopes during future astronaut operations.

By making both the foundation model and the underlying multimodal dataset publicly accessible, the partnership enables open scientific exploration across global research institutions. Academic researchers and students can now leverage pre-trained geospatial AI models directly without needing to construct custom model architectures or consolidate disparate space agency archives independently. This open-source distribution significantly lowers technical barriers for studying planetary geology and surface features.

Why it matters for AI for Students, Study & Research

For student researchers and academics in data science and astronomy, this release eliminates the friction of processing raw, unaligned space agency datasets. Instead of spending months building custom models or cleaning sensor data, researchers can fine-tune a pre-trained foundation model on Hugging Face to study lunar surface features, accelerate thesis research, and contribute to open-source space exploration tools.

What to do about it

• Access the NASA-IBM Lunar Foundation model weights and repository directly on Hugging Face.
• Download the unified multimodal dataset containing spatially aligned imagery and sensor maps from NASA missions.
• Adapt the foundation model for specific research tasks like identifying craters, ice deposits, or volcanic features.

Tools mentioned

• Hugging Face

Source: IBM is launching a new open source AI model to get NASA back to the Moon — and making petabytes of lunar data available to study – TechRadar (techradar.com)