Trained on more than thirty layers of data from nine instruments and four missions, the model reduces the error in detecting lunar ice deposits by up to 22%.
On September 10, 2026, IBM and NASA made available the NASA-IBM Lunar Foundation Model, one of the first open foundation models dedicated to the scientific exploration of the Moon. The stated goal is to transform decades of multi-instrument observations, which are difficult to use together, into results that can be used to prepare for a sustainable human presence on the satellite.
The training set aggregates more than thirty spatially aligned layers from nine instruments spread over four missions, including the Lunar Reconnaissance Orbiter and GRAIL for NASA, as well as the Japanese SELENE/Kaguya mission, i.e. tens of thousands of images and maps. The authors announce an error reduction of up to 22% on ice detection compared to a SwinV2-B reference trained on ImageNet, a gain of nearly 19% for the detection of craters at the contextual scale of a hundred meters, and a more modest improvement of 3% on the identification of volcanic formations.
Three use cases were used as a test bed: potential ice deposits in the polar regions in permanent shadow, re-enactment of volcanic history via irregular pond spots, and crater mapping for landing site selection and obstacle avoidance. The model is published on Hugging Face and joins the Prithvi family of foundation models jointly developed by IBM and the space agency.
Read the full article on the original source
Go to the source →