NASA and IBM launched an AI model to study the Moon

NASA, in collaboration with IBM Research and academic institutions, has launched the NASA-IBM Lunar Foundation Model — an open AI model designed to analyze scientific data about the Moon. The model is available for use on Hugging Face, and its source code is published on GitHub.

The model was trained primarily on data collected by NASA's Lunar Reconnaissance Orbiter over 17 years. Approximately 2 million images were used for training, including more than 1 million high-resolution images with a resolution of up to 1 meter and nearly 964,000 multispectral images with a resolution of 100 meters. Data from the GRAIL, Lunar Prospector missions, and the Japanese SELENE spacecraft were also used in training.

The NASA-IBM Lunar Foundation Model enables researchers to analyze the lunar surface more quickly and adapt the model to various scientific tasks. It can be used to map craters, identify young volcanic formations, and assess the location and stability of ice deposits in polar regions.

A separate area of focus is the permanently shadowed regions near the Moon's poles, where temperatures may remain low enough for ice to accumulate over billions of years. The model helps evaluate where such deposits may be stable on and beneath the surface.

AI is also used to search for and measure craters. In one test, the model successfully detected existing craters and a new impact crater formed after the fall of a space object. Notably, data about the new crater was not used during the model's pre-training.

Testing results showed that the NASA-IBM Lunar Foundation Model performed comparably to or better than several other foundational models. A particularly notable advantage was the accurate determination of polar ice stability.

NASA notes that the model is part of a joint NASA-IBM program applying AI to scientific research. As part of the partnership, models are also being developed to analyze Earth data and predict space weather.

The NASA-IBM Lunar Foundation Model, training datasets, and test sets are publicly available. The model is integrated into the open-source TerraTorch toolkit, allowing researchers to use, compare, and refine it for further lunar research.

Source: science.nasa.gov