IBM and NASA open-source a Lunar Foundation Model

IBM Research and NASA Science announced an open Lunar Foundation Model — a moon-mapping AI trained mainly on Lunar Reconnaissance Orbiter data, using more than one kind of image and more than one resolution.

An open moon-mapping model now has dual IBM and NASA primaries — Artemis-era mapping tooling, not a lander announcement. The percentage gains stay the agencies’ claims.

IBM Research published “Introducing IBM and NASA’s new foundation model for the Moon,” and NASA Science published “NASA, IBM Launch AI Foundation Model for Lunar Science” (Rachel Wyatt). Both name the open-source NASA-IBM Lunar Foundation Model. Those agency pages are the primary.

A foundation model is a general model you later adapt to many tasks. IBM says this is the first AI model to combine lunar observations across sensor types, viewing angles, and spatial scales, and it sits in the line of earlier IBM/NASA Earth and Sun models Prithvi EO and Surya. NASA’s wording is more cautious: among the first open-source AI models built specifically for lunar science. Those are the agencies’ framings — not an independent “first” audit.

NASA says the model was trained mainly on Lunar Reconnaissance Orbiter (LRO) data — roughly 2 million image tiles, comprising more than 1 million high-resolution camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. It was also trained on imagery and terrain data from NASA’s GRAIL (Gravity Recovery and Interior Laboratory), NASA’s Lunar Prospector, and JAXA’s Selenological and Engineering Explorer (SELENE).

IBM says the architecture is a version of TerraMind, the Earth-observation model developed by IBM and ESA (European Space Agency). Fine-tuning used lightweight LoRA adapters — Low-Rank Adaptation, a small add-on that left 90% of the base model’s weights frozen. NASA says weights are hosted on Hugging Face, the complete codebase is on GitHub, and the model is integrated into the open-source TerraTorch toolkit. Wiring details are in Sources.

Priority tasks on both pages are mapping smaller uncatalogued craters, investigating volcanic history via irregular mare patches (IMPs), and prospecting polar craters for ice. IBM lists those as NASA’s initial priorities. NASA says the pre-trained model can be adapted to those tasks with only small amounts of labeled data.

IBM’s published comparisons versus task-specific Swin / SwinV2 models — a family of vision transformers: a 22% cut in error rate identifying whether a dark polar crater might contain ice, versus a SwinV2 transformer trained for ice prospecting; crater detection at 1 meter-per-pixel matched a Swin model trained for crater detection; at 100 meters-per-pixel the NASA-IBM model outperformed that same model by nearly 19% using half the training data; on mapping irregular mare patch extents it outperformed a task-specific Swin by 3%. NASA’s page is more cautious: the model matched or exceeded several strong baselines, with comparable results on crater mapping and IMP segmentation and a clear advantage on estimating polar ice stability. Those percentages and rankings are IBM’s / NASA’s benches —.

This is a mapping and analysis tool drop, not a lander or Artemis schedule announcement. IBM’s page puts the model in the context of Artemis-era exploration and polar-ice prospecting. It does not announce a landing date, lander, or certified landing site.