NASA and IBM have released the NASA-IBM Lunar Foundation Model, a publicly available artificial intelligence system designed to help scientists map the moon’s surface and process decades of lunar data. Launched on a Thursday, the open-source model is available to download from Hugging Face, according to CNET. The project builds on more than 50 years of collaboration between IBM and NASA, dating back to the Apollo missions.
NASA and IBM Release New Lunar Foundation Model
The model aims to support future exploration under NASA’s Artemis program by helping researchers analyze petabytes of observation data gathered over the last five decades. Traditionally, scientists relied on manual examination of maps and images or used limited transformer models such as SwinV2-B, a 2022 vision system designed for general image recognition tasks.
Performance and Capabilities in Identifying Lunar Features
In benchmark tests conducted by NASA and IBM, the new foundation model outperformed the baseline SwinV2-B system across several key geographic and environmental analysis categories. When identifying potential ice deposits on the lunar surface, the NASA-IBM model reduced errors by 23 percent compared to SwinV2-B. In crater classification tasks, the model outperformed the baseline system by 19 percent while utilizing only half the training data.
The system is also equipped to detect volcanic formations known as irregular mare patches (IMPs) and evaluate lighting, terrain structure, and physical geography. According to TechRadar, identifying ice deposits is vital because water and oxygen will be crucial for establishing a human base on the moon and creating rocket fuel for future Mars missions. Crater and slope analysis can additionally help scientists identify safe landing sites free of hazards such as boulders and steep slopes.
Overcoming Training Challenges With SomBench
Training an AI model on lunar data presented unique hurdles compared to Earth observation. Dr. Juan Bernabé-Moreno, director of IBM Research Europe, UK and Ireland, noted that traditional computer vision techniques—which typically involve masking parts of an image and asking a model to reconstruct the missing 90 percent—proved to be a complete disaster
because many lunar craters look nearly identical when photographed from orbit.
To establish consistency, the research team divided the moon into wedges, completely separating the training wedges from the testing wedges. Additionally, the team compiled a layered benchmark dataset named SomBench, which consists of nearly 2 million overlapping map tiles. SomBench organizes data into aligned tracks so that information from different instruments, resolutions, and viewing angles is properly coordinated. This dataset brings together tens of thousands of maps and images collected by nine instruments across four moon missions, utilizing data from NASA’s Lunar Reconnaissance Orbiter and the GRAIL mission.
Real-World Verification and Open-Science Goals
The model’s capabilities were tested when a SpaceX Falcon 9 rocket crashed into the moon. When an image of the impact was fed into the system, it correctly identified the crash site as a new crater on its first attempt, even though it closely overlapped with an existing crater. Bernabé-Moreno described the performance as working “fantastically.”
Campbell Watson, senior research manager at IBM Research, stated that open-sourcing both the model and the underlying dataset continues a tradition of broad scientific access and collaboration, providing researchers globally with a shared foundation that can be adapted for new inquiries without requiring them to build an AI model from scratch.
