NASA and IBM Launch Open-Source AI Model to Map Lunar Ice and Craters
The new Lunar Foundation Model outperforms existing vision systems in identifying critical resources for future Artemis missions.
NASA and IBM have released the NASA-IBM Lunar Foundation Model, an open-source AI system designed to help planetary scientists map the Moon's surface. Now available on Hugging Face, the tool aims to accelerate the discovery of lunar ice and the classification of craters to support future human exploration.
The model demonstrates significant performance gains over previous technology. It reduced errors in identifying lunar ice prospectivity by 23% when compared to Microsoft's SwinV2-B vision system. Additionally, the model outperformed SwinV2-B by 19% in the identification and classification of craters, achieving these results while utilizing only half the amount of training data.
Overcoming Lunar Lighting Challenges
Developing the model required a departure from standard AI training methods. Traditional masked autoencoder training failed due to the Moon's "knife-edged" shadows and the high degree of similarity between various craters. To solve this, researchers implemented a non-traditional approach that divided the lunar surface into "wedges," similar to the segments of an orange, to better handle the unique lighting and shadow challenges of the lunar environment.
This project is a component of the broader NASA-IBM AI4Science collaboration, which also includes the "Surya" model developed for heliophysics. To support the model, the partners released a first-of-its-kind open-source co-registered dataset. This library contains over two million data points sourced from JAXA's SELENE spacecraft as well as NASA's Gravity Recovery and Interior Laboratory (GRAIL) and Lunar Reconnaissance Orbiter (LRO).
Implications for the Artemis Era
By providing both a high-performing foundation model and a massive open dataset, NASA and IBM are lowering the technical barrier for the global scientific community. This accessibility is critical as NASA prepares for the Artemis missions, which seek to establish a sustainable human presence on the Moon.
The ability to precisely locate water ice is essential for long-term habitation, as it can provide life support and fuel for deep-space exploration. By automating the analysis of vast amounts of orbital data, the model allows scientists to identify these critical resources more efficiently than ever before.
Future Outlook
As the scientific community integrates the Lunar Foundation Model into their workflows, the focus will shift toward verifying these AI-identified sites through future lunar landings. While the model provides a powerful predictive tool for ice prospectivity and crater mapping, the next phase of the Artemis program will determine how these digital maps translate into physical resource extraction on the lunar surface.