NASA and IBM open-source lunar AI foundation model

NASA's Science office, in a 10 September 2026 announcement, released the NASA-IBM Lunar Foundation Model as an open-source multimodal model for lunar science. The weights sit on Hugging Face, the codebase is on GitHub, and the team plugged it into the open-source TerraTorch toolkit.

This is a NASA Science open-science model release with IBM Research, plus datasets and a companion paper. It is not a spacecraft launch, and it is not a commercial product sale.

The model was trained mainly on 17 years of Lunar Reconnaissance Orbiter data, NASA says, on roughly 2 million image tiles: more than 1 million high-resolution camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. GRAIL, Lunar Prospector, and JAXA SELENE data also went into the mix.

Researchers can fine-tune it on crater mapping, irregular mare patches (young volcanic features), and polar ice stability in permanently shadowed regions. Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer, said the point is turning petabytes of lunar data into something scientists can explore, not collecting another archive.

IBM Research, in its matching 10 September blog, says the team used LoRA adapters that left about 90% of the base weights frozen. On polar ice stability, IBM says the model cut error about 22% versus a SwinV2-B ImageNet baseline.

NASA is clear this is a research foundation model for scientists. Varying orbital lighting can hide smaller craters, and change detection between orbits is not perfect. It is not a flight-certified navigation product.

The release continues the NASA-IBM science-model line after Prithvi on Earth and Surya on the Sun. Other Earth-observation and launch notes on Techpresso include Pixxel's $100 million Series C, Stoke Space's $1 billion Series E for Nova, Isar Aerospace reaching orbit with Spectrum, and Google and Cathay's Asia-Pacific contrail work.

If you map lunar ice, irregular mare patches, or craters, download the Hugging Face checkpoint and run it through TerraTorch. Treat the 22% ice-error cut as a research benchmark on a frozen-weight adapter, not as a certified landing-site map.