NASA-IBM AI Model Spots Moon Ice with Unprecedented Accuracy

NASA and IBM have released an open-source AI model that predicts where ice may be buried in the Moon's permanently shadowed polar craters. The NASA-IBM Lunar Foundation Model is available for download on Hugging Face and has been shown to pinpoint potential ice areas with 22% less error than a widely used baseline algorithm.

The model was built to serve three initial scientific priorities, each of which has a direct connection to the Artemis mission's plans. The first priority is finding ice in permanently shadowed regions, where surface temperatures can drop to -410°F (-246°C). A NASA-IBM authored technical paper shows that the model reduced error by up to 22% compared to a SwinV2-B transformer pre-trained on ImageNet.

The team also used LoRA (Low-Rank Adaptation) to adapt the model for specific tasks. This means researchers can take the public Lunar Foundation Model and fine-tune it for new applications using only labeled examples and standard computing hardware.