NASA and IBM Release Open-Source AI Model to Advance Lunar Exploration and Mapping

NASA and IBM Research have jointly launched the NASA-IBM Lunar Foundation Model, marking a significant expansion of open-source artificial intelligence tools designed specifically for planetary science and lunar exploration. Released publicly on Hugging Face alongside a complete codebase hosted on GitHub, the new model processes petabytes of historical lunar observation data to help researchers evaluate rugged terrain, identify volcanic features, and estimate subsurface water ice. The development represents a shift away from traditional manual mapping methods and task-specific algorithms toward multi-modal foundation models capable of generalizing across large spatial datasets.

Trained on data collected over the past 17 years by NASA's Lunar Reconnaissance Orbiter, the model incorporates approximately two million image tiles, including over one million high-resolution camera images captured at one-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. Additional terrain and imagery data from NASA's Gravity Recovery and Interior Laboratory, the Lunar Prospector, and the Japan Aerospace Exploration Agency's Selenological and Engineering Explorer were integrated into the training pipeline. This multi-instrument foundation is intended to support the establishment of a sustained human presence on the Moon under ongoing agency planning.

The scale of information gathered by the Lunar Reconnaissance Orbiter exceeds that of all other NASA planetary missions combined, creating both a scientific resource and a data management challenge. To overcome these limitations, IBM and NASA scientists assembled what is described as the first unified, machine-learning-ready lunar dataset aggregating over 30 spatially aligned layers from nine instruments across four distinct missions. This pre-training dataset allows researchers to avoid building specialized algorithms from scratch for individual tasks, instead utilizing fine-tuning techniques on top of broad pre-acquired knowledge.

Within NASA, the Impact AI team based at the Marshall Space Flight Center in Huntsville, Alabama, collaborated directly with researchers from the Science Mission Directorate Planetary Science Division, the Goddard Space Flight Center in Greenbelt, Maryland, and the Ames Research Center in California's Silicon Valley. The resulting software infrastructure is fully integrated into the open-source TerraTorch toolkit, supported by a companion technical paper designed to ensure reproducibility across the global academic and industrial research community.

Identifying subsurface resources is a primary objective for future lunar missions, particularly for water and oxygen extraction required for a prospective lunar base and rocket fuel production for crewed missions to Mars. The newly released foundation model assists scientists in predicting where ice patches are likely to remain stable both on and beneath the lunar surface, particularly within permanently shadowed regions where extremely low temperatures can preserve ice deposits for billions of years.

In benchmark evaluations highlighted in technical documentation authored by IBM and NASA, the foundation model reduced error rates when identifying areas with high potential for lunar ice by up to 22 percent compared to baseline SwinV2-B ImageNet models. Researchers noted that the system maintains fine-scale prospectivity patterns present in reference data, offering higher accuracy and efficiency when handling multi-resolution inputs from polar regions.

Although the Moon is generally considered to be volcanically inactive today, it experienced complex geological processes in its past that continue to puzzle planetary scientists. The foundation model accelerates the identification and segmentation of unusual volcanic structures known as irregular mare patches, which appear relatively young and challenge established scientific timelines regarding the rate at which the lunar interior cooled.

By leveraging the model's pattern recognition capabilities, researchers can map these volcanic formations with lower fine-tuning costs. According to project evaluations, the foundation model captures the geographic extent of irregular mare patches with notable precision compared to traditional architectures, achieving comparable accuracy while improving computational efficiency for scientists investigating the Moon's thermal history.

Crater counting and measurement serve as essential methods for dating lunar terrains and reconstructing the geological history of the inner solar system, while also guiding safe landing site selection for future missions. The foundation model enables scientists to classify craters at meter-scale resolution while outperforming baseline models by nearly 19 percent at context-scale resolution using only half of the standard training data volume.

Demonstrating the model's utility for tracking dynamic changes, test runs involving Lunar Reconnaissance Orbiter imagery captured before and after a SpaceX rocket body impact near Einstein crater showed that the model successfully detected existing formations and highlighted the newly created impact site. Because the post-impact imagery was deliberately excluded from the initial pre-training phase, the test proved the model can be fine-tuned to detect novel surface alterations between orbital observations.

The lunar model joins a growing portfolio of scientific artificial intelligence tools developed through the ongoing partnership between NASA and IBM under the agency's strategy for open science. Previous releases include the Prithvi family of models focused on Earth observation data to assist with disaster monitoring, flood mapping, and crop yield forecasting, as well as the Surya heliophysics model trained to predict space weather phenomena such as solar flares.

"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. 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."

The complete codebase, pre-training datasets, and benchmark collections are now publicly accessible via GitHub and Hugging Face, allowing international researchers to deploy, test, and adapt the software for independent lunar studies. The open-access release is intended to streamline collaborative research efforts across academic institutions including Howard University, the University of Maryland, Baltimore County, the SETI Institute, and the Universities Space Research Association.

Further evaluation milestones and community-driven fine-tuning tasks will be managed through the open-source repository ecosystem as research teams begin applying the foundation model to broader planetary science objectives ahead of upcoming exploration milestones.