NASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science
NASA and IBM released an open-source lunar foundation model trained on 17 years of orbiter data for ice and crater tasks.
NASA and IBM Research released the open-source NASA-IBM Lunar Foundation Model, pretrained from scratch on SomBench, described as the largest co-registered multimodal lunar corpus. The dataset holds nearly 2 million tile bundles from 17 years of Lunar Reconnaissance Orbiter observations, plus GRAIL, Lunar Prospector, and JAXA Kaguya data across more than 30 layers. Built on the TerraMind architecture but trained from scratch, it takes illumination geometry as explicit input and uses FlexiViT for variable patch sizes. On polar ice prediction it cut error by up to 22 percent versus SwinV2-B and improved coarse crater detection by nearly 19 percent with half the labeled data; it is not reliable for absolute geodetic positioning.
- Open-source NASA-IBM Lunar Foundation Model pretrained on nearly 2 million multimodal tiles.
- Training used 17 years of LRO imagery plus GRAIL, Lunar Prospector, and Kaguya data.
- Polar ice prediction error dropped up to 22 percent versus SwinV2-B.
- Coarse crater detection beat SwinV2-B by nearly 19 percent with half the labels.
- LoRA matched full fine-tuning, but the model is unfit for absolute positioning.
- Organizations
- NASAIBM ResearchJAXA
- Countries
- United StatesJapan
From the original article
5 min read at the-decoder.com
NASA and IBM have released the Lunar Foundation Model, one of the first open-source AI models for lunar science. Trained on nearly 2 million tile bundles, mostly from 17 years of Lunar Reconnaissance Orbiter data, it cuts the error in predicting polar ice deposits by up to 22 percent compared to the strongest model it was tested against.
By Jonathan Kemper · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.