The Pulse
NASA and IBM Open-Source a Lunar Model for Ice and Crater Mapping
NASA and IBM released the Lunar Foundation Model, an open-source system trained on decades of lunar observations. The model is designed to help researchers identify potential ice deposits, craters and volcanic formations across data from mu
AI.info Team ·
NASA and IBM Put a Shared Model Behind Lunar Mapping
NASA and IBM released an open-source foundation model on September 10 that is designed to analyze the Moon across instruments, image resolutions and scientific tasks. The NASA-IBM Lunar Foundation Model is intended to help researchers locate potential ice deposits, map craters and study volcanic formations without building a separate machine-learning system for every question.
The release addresses a disagreement built into lunar science: the Moon has been observed in extraordinary detail, but the resulting data remains difficult to combine. Scientists have often had to work through maps and images by hand or train narrow models for individual applications. IBM and NASA argue that a shared model can expose relationships hidden across those sources, while the software release makes the work available for researchers outside the two organizations.
“NASA has spent decades building an extraordinary scientific record of the Moon, but collecting data is only part of the job,” said Kevin Murphy, chief science data officer and acting chief data and AI officer at NASA Headquarters in Washington. “We also have to make data easier for scientists to explore and use.”
More Than Two Million Lunar Image Tiles
The model draws primarily on observations from NASA’s Lunar Reconnaissance Orbiter, which has surveyed most of the Moon’s surface during a mission that began in 2009. NASA says training used roughly 2 million image tiles, including more than 1 million high-resolution camera images at about 1 meter per pixel and nearly 964,000 multispectral images at 100 meters per pixel.
Additional terrain and gravity information came from missions including NASA’s GRAIL and Lunar Prospector, along with data from Japan’s SELENE/Kaguya mission. IBM and NASA describe the accompanying release as an open-source lunar dataset that brings together more than 30 spatially aligned layers from nine instruments across four missions. The model and data are available through Hugging Face, while the fine-tuning and inference code is published in a NASA-IMPACT GitHub repository.
Ice, Craters and Volcanic Patches Are the First Tests
Potential ice deposits in permanently shadowed regions are one of the model’s main targets. Water could provide drinking supplies and oxygen, while hydrogen and oxygen derived from water could support propellant production for later missions. The IBM-NASA technical results say the model reduces error in identifying areas with high ice potential by up to 22% compared with a widely used SwinV2-B model trained on ImageNet.
Crater mapping offers a different test. Craters help scientists estimate the age and composition of lunar terrain, while detailed maps can inform landing-site selection and surface-route planning. At roughly 100-meter context resolution, the model outperformed the same comparison system by nearly 19% while using half the training data, according to the release.
The system also segments irregular mare patches, subtle volcanic formations that can help scientists study the Moon’s thermal history. IBM and NASA report a 3% improvement over the comparison method on that task, with similar accuracy achieved at lower fine-tuning cost.
The Open Release Has a Defined Boundary
Researchers receive model weights, configuration files, datasets and downstream examples, but not the full pretraining code. The GitHub repository describes itself as a fine-tuning and inference release, with examples for crater detection, volcanic-feature segmentation and ice prospectivity. That distinction matters: downloading the model is straightforward, while reproducing the original training run requires infrastructure and details beyond the public repository.
The model is also a starting point rather than a replacement for geological review. Its benchmark results measure performance on selected research tasks, not autonomous mission decisions. Scientists will still need to account for sensor limitations, uncertain labels and the differences between orbital observations and conditions encountered by a lander or crew.
NASA Adds the Moon to Its Foundation-Model Program
The lunar system joins the Prithvi family of open models developed by NASA and IBM for scientific data. Earlier releases focused on Earth observation and weather, while the Surya model applies a similar approach to heliophysics. The lunar release extends that program from monitoring Earth and the Sun to interpreting a planetary surface.
Juan Bernabe-Moreno, director of IBM Research Europe, UK and Ireland, said the model gives scientists “a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation, and providing an open platform the global research community can build on.”
For now, the concrete result is a public model trained on lunar observations that previously sat across separate instruments, resolutions and missions. Researchers can download the weights, inspect the benchmarks and adapt the system to new lunar mapping problems through the Hugging Face release.