NASA and IBM Open-Source a Lunar Foundation Model Built on Orbiter Imagery

On 10 September 2026, NASA and IBM Research released the NASA-IBM Lunar Foundation Model — among the first open-source AI models built specifically for lunar science. Trained primarily on imagery from NASA’s Lunar Reconnaissance Orbiter (LRO), the model is available on Hugging Face, with code on GitHub for anyone who wants to experiment.

Why a foundation model for the Moon

Lunar science has always been data-rich and labour-intensive. Orbiter cameras produce vast mosaics; researchers still spend enormous effort labelling craters, tracing volcanic features, and estimating where ice might remain stable near the poles. A foundation model — trained once on a large, shared corpus, then adapted to narrower tasks — is meant to compress that work.

According to NASA, the training set leans on roughly two million lunar image tiles. Evaluations covered crater mapping, segmentation of irregular mare patches (unusual-looking volcanic features that may challenge timelines of lunar cooling), and polar ice stability estimates. NASA reported that the model matched or exceeded several strong baselines across those tasks, with a clearer edge on polar ice stability.

Open by design

The release sits inside a broader NASA–IBM “AI for science” collaboration that already produced Earth-observation models. Making weights, code, and benchmarks public is deliberate: it lets university labs, mission planners, and industry teams compare approaches on the same footing instead of reinventing lunar feature extractors from scratch.

That matters for products as much as papers. Faster, more consistent maps of the lunar surface feed landing-site selection, resource prospecting, and hazard analysis for upcoming crewed and robotic missions. The science is still the point — but the engineering path from pixels to decisions is getting shorter.

What to watch next

Foundation models are only as good as the questions asked of them. Expect follow-on fine-tunes for specific instruments, latitudes, and lighting conditions, plus debates about how much human review still belongs in the loop before a model-generated map influences a mission plan.

For JSIPE readers, this is a clean example of the journey we care about: orbital science → machine learning engineering → tools that put lunar insight into more hands.