NASA and IBM launch open-source AI to map the Moon’s hidden resources

WASHINGTON — The NASA-IBM Lunar Foundation Model, an open-source artificial intelligence system now available on the Hugging Face AI platform, is designed to improve how scientists study the Moon’s topography and search for critical resources like water ice. Initial testing by NASA and IBM researchers has demonstrated that the system is exceptionally skilled at pinpointing locations on the lunar surface that may harbor underground or surface ice.

Developing an AI capable of navigating the visual realities of the Moon required overcoming significant physical and environmental challenges that do not exist in Earth-based satellite mapping. Dr. Juan Bernabé-Moreno, the director of IBM Research Europe, UK and Ireland, noted that training computer vision models on lunar imagery is vastly different from training them on Earth observation data.

In direct performance comparisons, the NASA-IBM model showcased a substantial advantage over existing state-of-the-art vision systems. When tasked with identifying areas likely to contain ice, the model reduced mapping errors by 23 percent compared to SwinV2-B, a high-resolution image-processing vision system developed by Microsoft that is widely utilized as the benchmark for computer vision tasks.

On Earth, the atmosphere scatters sunlight through Rayleigh scattering, which fills geographical shadows with ambient, diffused light and softens their boundaries. Because the Moon lacks an atmosphere, there is no atmospheric scattering, causing lunar shadows to be incredibly sharp, high-contrast, and pitch black. When orbiters photograph these shadowed regions, the pixels inside the shadows contain zero visual information.

The model also demonstrated superior efficiency in identifying and classifying impact craters. It outperformed SwinV2-B by 19 percent in crater classification tasks, achieving this higher accuracy while requiring only half the training data of the Microsoft-developed baseline.

As the Sun’s position changes throughout the lunar day, these harsh shadows shift dynamically, meaning a single crater can appear completely unrecognizable from one orbit to the next. These extreme visual variations initially caused traditional AI training techniques to fail.

The AI’s real-time capabilities were put to the test following an event on August 5, when a SpaceX Falcon 9 rocket stage crashed into the lunar surface. When researchers fed post-impact orbital imagery into the model, it successfully identified the fresh impact site as a new crater on its first attempt, even though the new impact zone directly overlapped an existing crater.

Typically, computer vision models are trained using self-supervised learning, where portions of an image are masked out and the model is tasked with reconstructing the missing data based on learned patterns. Bernabé-Moreno noted that early attempts to train the model using this standard masking method—such as hiding 90 percent of a crater and asking the system to reconstruct it from the remaining 10 percent—proved to be a “complete disaster.”

IBM

Because many impact craters look highly similar when viewed from orbit, the model could not reliably reconstruct the masked features. To resolve this, the IBM and NASA research team designed an alternative training methodology, dividing the lunar sphere into longitudinal segments similar to the wedges of an orange. By completely separating the geographic wedges used for training from the wedges used for testing, the researchers provided the model with the structural consistency required to generalize features across the entire lunar surface.

Beyond the model itself, the collaboration has yielded an unprecedented, open-source dataset that is expected to serve as a foundational resource for the global scientific community. The dataset co-registers more than two million data points, aligning high-resolution imagery pixel-by-pixel with other scientific sensor readings.

“One of the reasons we’ve never seen a comprehensive, powerful lunar model before is that we didn’t have the data organized in the right way,” Bernabé-Moreno said.

To build this unified data grid, the team compiled and synthesized decades of orbital observations from multiple international space missions. NASA’s Lunar Reconnaissance Orbiter (LRO), launched in June 2009, has spent over a decade collecting detailed topographic, thermal, and imaging data using instruments like the Lunar Orbiter Laser Altimeter (LOLA) and the Diviner Lunar Radiometer Experiment.

NASA’s Gravity Recovery and Interior Laboratory (GRAIL), launched in September 2011, utilized twin spacecraft named Ebb and Flow to map the Moon’s gravitational field with high precision. Japan’s Selenological and Engineering Explorer (SELENE), also known as Kaguya, launched by the Japan Aerospace Exploration Agency (JAXA) in September 2007, acquired highly detailed global maps of lunar chemistry, mineralogy, and topography.

By combining these distinct modalities—including gravity, elevation, and multi-spectral imaging—into a single, co-registered grid, researchers have created a resource that will persist even as individual AI models are superseded by newer technologies.

The evaluation incorporated a rigorous workflow combining terrain models, thermal measurements, and other environmental datasets. This technological breakthrough comes at a critical juncture for international space exploration.

The urgency of precise lunar mapping was underscored on April 6, when NASA’s Artemis II mission completed its historic lunar flyby. Carrying astronauts Reid Wiseman, Christina Koch, Victor Glover, and Jeremy Hansen, the flight marked the first time a crewed spacecraft had visited the vicinity of the Moon since the conclusion of the Apollo program in 1972.

As NASA prepares for subsequent missions under the Artemis campaign, including crewed landings near the lunar South Pole, the ability to locate water ice has become a strategic priority. Water ice trapped in permanently shadowed regions of the lunar poles is highly valuable, as it can be harvested to produce breathing oxygen and hydrogen-based rocket fuel.

The Lunar Foundation Model builds on previous collaborations between IBM and NASA. In 2023, the two organizations released an open-source geospatial foundation model for Earth science, utilizing data from the Harmonized Landsat Sentinel-2 satellites to track deforestation, climate change, and urban sprawl. The Artemis II mission represented the farthest human journey into deep space to date.