NASA and IBM have collaborated to develop an AI model aimed at improving lunar exploration. This model, known as the NASA-IBM Lunar Foundation Model, is an open-source tool available for download through Hugging Face. It is designed to analyze extensive lunar data, enabling scientists to create more accurate maps of the Moon's surface. The AI system was trained using a diverse dataset that includes tens of thousands of images and scientific measurements from several space missions, including NASA's Lunar Reconnaissance Orbiter (LRO) and GRAIL, as well as Japan's SELENE mission. This dataset combines data from over 30 different layers collected by nine instruments across four separate missions, offering a comprehensive view of the lunar surface. In testing, the model outperformed a well-known vision system called SwinV2-B, which was developed by Microsoft for high-resolution image processing. The AI model reduced errors in detecting potential ice deposits by 23 percent and improved crater detection by 19 percent, even when using only half the training data. Notably, it successfully identified a new crater created by a SpaceX Falcon 9 rocket crash in August 2023, even though the site overlapped with an existing crater. This highlights the model's ability to distinguish between overlapping features in lunar terrain. One of the biggest challenges in lunar data analysis is the Moon's lack of an atmosphere. This results in extreme lighting conditions, with deep shadows and intense sunlight that can confuse traditional image analysis techniques. To help address this challenge, the research team created a benchmark dataset called SomBench, which includes nearly 2 million overlapping map patches, or "tiles." These tiles combine data from various instruments and images taken at different times and angles, creating a more complete and accurate representation of the lunar surface. The model uses a technique known as masked-token learning, where parts of the dataset are hidden, and the AI must predict the missing information. This process is repeated across millions of examples, allowing the AI to understand complex relationships between elements like lighting, terrain, and geography. Researchers can adapt the model using efficient methods like low-rank adaptation, which allows them to fine-tune the AI without the need for expensive and time-consuming retraining. This model is expected to serve as a foundation for future lunar research, helping improve crater detection, mapping, and the search for water ice in the Moon's polar regions. The release of this dataset marks the first time such a unified and publicly accessible collection of lunar data has been available for use in machine learning.