NASA and IBM released an open-source artificial intelligence model designed to map lunar craters, volcanic formations, and polar ice. Trained on seventeen years of satellite observations, the tool aims to support upcoming Artemis missions and help researchers analyze petabytes of planetary data.
Decades of planetary exploration have left space agencies with an enormous accumulation of satellite imagery, yet turning those archives into usable discoveries remains a formidable logistical hurdle. To bridge that gap, IBM and NASA released the NASA-IBM Lunar Foundation Model on September 10, 2026. Available publicly on Hugging Face under an open-source license, the system provides researchers with a pre-trained backbone that eliminates the need to build specialized mapping algorithms from scratch for every distinct research project.
Training on Decades of Lunar Reconnaissance Orbiter Data
The foundation model draws its core training from imagery gathered by NASA’s Lunar Reconnaissance Orbiter, which has mapped the lunar surface in high detail for roughly 17 years. According to mission documentation released alongside the tool, the training dataset covers approximately two million image tiles. This collection includes more than one million high-resolution Narrow Angle Camera images at roughly one-meter resolution and nearly 964,000 multispectral Wide Angle Camera frames at approximately 100-meter resolution.
Engineers supplemented the orbital imagery with co-registered terrain and gravity data from additional planetary missions. These sources include NASA’s Gravity Recovery and Interior Laboratory (GRAIL), the Lunar Prospector, and the Japan Aerospace Exploration Agency’s SELENE, also known as Kaguya. By unifying these diverse inputs into a machine-learning-ready framework, the model can identify complex spatial relationships across different instruments and resolutions.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job. We also have to make data easier for scientists to explore and use. The NASA-IBM Lunar Foundation Model shows what’s possible when we bring AI to NASA’s petabytes of scientific data. That’s a real opportunity we see with AI: turning large-scale data into new discoveries.”
Kevin Murphy, NASA’s chief science data officer and acting chief data and AI officer
Overcoming Lighting Challenges in Deep Space Observation
Analyzing lunar images presents unique obstacles that standard artificial intelligence models trained on terrestrial data cannot easily resolve. Dr. Juan Bernabé-Moreno, director of IBM Research Europe for the U.K. and Ireland, explained that lunar scenes lack an atmosphere to scatter light into shadows, leaving craters rendered in stark, pitch-black contrast. Because illumination angles shift dramatically depending on the orbital pass, identical geographic features can look entirely different from one image to the next.
To prevent the algorithm from memorizing specific tiles during training, the development team divided the Moon into distinct geographic wedges for testing and validation. The underlying architecture relies on a Vision Transformer encoder-decoder that explicitly incorporates illumination geometry as contextual data, ensuring the system evaluates surface features rather than momentary lighting anomalies.
Targeting Ice Deposits, Volcanic Features, and Impact Craters
Planetary scientists can fine-tune the backbone model to address three primary areas of lunar research: subsurface volatile distribution, thermal history, and terrain classification.
Identifying trapped ice reserves is considered critical for future exploration, as water and oxygen resources can sustain long-term bases and generate rocket fuel for missions traveling toward Mars.
The system also assists researchers studying irregular mare patches, which are unusual volcanic structures that challenge established models of lunar cooling. Operating on imperfect training labels, the model improved the identification of these formations’ boundaries by 3 percent.
Stress-Testing the System Against Recent Impact Scars
Independent research teams can access the complete fine-tuning code on GitHub, integrate the tools through the TerraTorch toolkit, and utilize bundled tile datasets on Hugging Face as NASA prepares for crewed Artemis flights.
