NASA and IBM Unveil AI Model to Map Moon’s Surface in Unprecedented Detail

In a significant advancement for lunar exploration, researchers from NASA and IBM have collaborated to develop the Lunar Foundation Model (LFM), an artificial intelligence system designed to synthesize decades of fragmented lunar data. The partnership aims to transform massive, disjointed troves of robotic mission information into a cohesive, high-resolution map of the moon’s surface.

Historically, scientists have relied on general-purpose transformer models like SwinV2-B to analyze lunar imagery. However, data collected from orbiting spacecraft comes from a mismatched array of sensors—ranging from high-resolution optical cameras and laser altimeters to radar reflectance tools and spectrometers—making it difficult to create a unified picture. The resulting datasets are vast but labor-intensive to process, often hindering accurate surface analysis.

To overcome this bottleneck, the team released their findings on September 10 in a technical paper. The LFM is an open-source model available via the Hugging Face repository, allowing researchers worldwide to fine-tune the AI for specific scientific inquiries. Key applications include generating reliable crater maps for safe landing zones, analyzing crater compositions to understand the moon’s interior history, and identifying heavily shadowed regions that may harbor subsurface ice critical for long-term bases.

Lunar observation presents unique computational challenges distinct from weather or geospatial modeling. The moon lacks an atmosphere, leading to extreme sunlight geometry that creates deep, deceptive shadows or washes out geological details. Additionally, instruments capture data at vastly different scales, from broad regional maps at 100 meters per pixel to detailed terrain scans at 1 meter per pixel.

The researchers addressed these issues by compiling SomBench, a layered benchmark dataset comprising nearly 2 million overlapping map patches. This system aligns data from various instruments and resolutions, layering lower-resolution overviews with high-resolution camera shots, slope maps, thermal behavior readings, mineralogy data, and gravitational anomalies.

LFM utilizes a technique called masked-token learning, where parts of the dataset are hidden, forcing the AI to predict and fill in the blanks. By learning relationships between lighting, terrain structure, and physical geography across millions of examples, the model becomes adept at interpreting complex lunar environments. Furthermore, explicit metadata regarding solar angles and spacecraft positions is integrated directly into the architecture, helping the AI distinguish actual terrain features from shadow-based illusions.

In evaluations, the LFM demonstrated superior performance compared to previous models. It outperformed SwinV2-B in crater detection by nearly 19% while requiring half the training labels. For polar ice prospectivity within the top meter of regolith, the model reduced errors in identifying high-potential ice areas by up to 22%. Its ability to map meter-scale craters and rare volcanic features, known as irregular mare patches, was also competitive with top custom-built models.

Scientists can customize the LFM using lightweight fine-tuning methods such as low-rank adaptation, which adjusts a small set of weights without the computational expense of training from scratch. While the team cautions that the AI serves as a pattern-recognition assistant rather than a replacement for physical measurements, it represents a major step toward creating an intelligent toolkit for both robotic and crewed lunar missions. Over time, the model is expected to improve further through the integration of terrain, illumination, and thermal layers.