NASA-IBM Lunar Foundation Model: the open-source Moon AI, explained

What NASA and IBM actually shipped

The NASA-IBM Lunar Foundation Model is, as far as anyone can tell, among the first open-source AI models built specifically for lunar science. It was developed by IBM Research together with NASA's IMPACT team and a group of academic partners: the Universities Space Research Association, the SETI Institute, the University of Maryland, Baltimore County, and Howard University.

The idea is the same one behind every foundation model you have heard of. Instead of training a narrow model to do one task, you pretrain a large one on a mountain of raw data so it learns the general "shape" of a domain, then fine-tune it cheaply for specific jobs. ChatGPT did this for text. This does it for the surface of the Moon.

"That's a real opportunity we see with AI: turning large-scale data into discoveries."

Kevin Murphy, NASA Chief Science Data Officer

Worth noting up front: NASA and IBM have not published the model's parameter count, so I won't invent one. For scale reference, its sibling solar model Surya is 366M parameters. What they did detail thoroughly is the data, which is the interesting part anyway.

What it was trained on: 17 years of orbital data

This is where the work went. The team assembled what they call the first open-source unified lunar dataset: more than 30 spatially-aligned layers from 9 instruments across 4 missions, adding up to roughly 2 million image tiles.

The bulk of it comes from NASA's Lunar Reconnaissance Orbiter, which has been photographing the Moon for 17 years. That gave the team over a million high-resolution camera images at 1-metre detail, plus around 964,000 multispectral images at 100-metre resolution. Data from NASA's GRAIL and Lunar Prospector missions and Japan's Kaguya orbiter filled in the gaps.

Aligning all of that so every layer lines up over the same patch of ground is unglamorous, expensive work. It is also the reason the model is any good. The lesson repeats everywhere in AI: the moat is rarely the architecture, it is the clean, domain-specific training data nobody else has bothered to assemble.

What the model can do

Once trained, the model can be pointed at a handful of lunar science tasks that used to each need their own custom pipeline:

• Estimate polar ice. Water ice sitting in permanently shadowed craters near the poles is the single most valuable resource on the Moon, since it is potential fuel and life support for a sustained Artemis presence. The model estimates where ice is stable and likely to be.
• Map craters. It detects and classifies craters at both metre and roughly 100-metre scales, which feeds directly into landing-site selection, hazard avoidance, and dating the surface.
• Find volcanic features. It flags irregular mare patches, young volcanic formations that tell scientists about the Moon's thermal history.
• Spot changes over time. Run it across images from different dates and it surfaces surface changes. In testing, it detected novel impact craters that had appeared in images taken after its training cut-off.

NASA is careful to frame this as a research assistant, not an oracle. It surfaces candidates for scientists to verify. That distinction, a capable model that proposes while a human disposes, is exactly the guardrail I wish more AI products shipped with by default.

The number that matters: specialised beat generic

Here is the result that made me want to write this up. IBM benchmarked the model against SwinV2-B, a strong, widely used vision model pretrained on ImageNet, a giant general-purpose image dataset. In other words, big-and-generic versus smaller-and-specialised.

The specialised model won across the board:

• Up to 22% lower error on polar-ice estimation.
• Nearly 19% better at context-scale crater detection, using half the training data.
• 3% better on volcanic feature detection, even with imperfect labels.
• Comparable accuracy on metre-scale craters, but noticeably more efficient.

NASA's own summary rounds this up to a model that "exceeds widely used methods by up to 23%". Read past the specific figure and the point is simple: a model that learned the Moon beat a model that learned everything, using less data to do it.

"The model gives scientists a foundation to explore the Moon at scale."

Juan Bernabe-Moreno, Director of IBM Research Europe

Why a "foundation model for one domain" is the real story

You could be forgiven for tuning out a Moon-mapping model as niche space news. I would argue the opposite. This is one of the cleanest public demonstrations of a principle that decides whether almost any AI project succeeds: relevance beats raw scale.

A general model has seen a lot of everything and none of your thing. A specialised model has seen your thing exhaustively. For a bounded, high-stakes domain, the second one wins, which is the same reason an AI agent trained on your own knowledge base answers better than a raw language model wired to a chat box.

This is the part of my day job I think about constantly. A support AI is not useful because it runs on a big model. It is useful because it has read every one of your past tickets, your help center, your internal docs, and your macros, and it answers the way your best agent would. Bolt a generic model onto a help widget with no company context and you get confident, wrong answers, the exact failure mode that gives rule-based chatbots and thin AI wrappers a bad name.

At eesel, this is the whole design. eesel is an AI teammate platform, and you hire ready-to-work teammates for specific jobs, the AI helpdesk teammate being the one most relevant here. That teammate is not a blank model. It learns from your historical tickets and knowledge sources the moment it connects, so it arrives already knowing your product the way NASA's model arrives already knowing the Moon.

And crucially, like NASA's "propose, don't decide" framing, you get to check the work before it goes live. eesel simulates the AI on thousands of your real past tickets so you can see exactly how it would have replied and what it would have resolved, before a single customer talks to it. No flip-the-switch-and-hope.

The pricing follows the same "pay for what's relevant" logic: eesel bills per ticket it actually handles, not per seat, so a specialised teammate that resolves your repetitive questions costs you when it works. If you run support and you have been waiting for the AI hype to turn into something that actually knows your business, the Moon just gave you the proof of concept. Try eesel free and point it at your own tickets.

Where to get it, and what's next

Everything is open. The weights live on Hugging Face under the nasa-ibm-ai4science collection, the code is on GitHub as NASA-IMPACT/NASA-IBM-Lunar-Foundation-Model, and it has been folded into the open-source TerraTorch toolkit so researchers can fine-tune it for their own tasks.

It is also part of a bigger arc. This model joins Prithvi, the NASA-IBM geospatial model for Earth, and Surya, their model of the Sun. The stated plan is to keep going into planetary science, astrophysics, and the biological and physical sciences. NASA is quietly building a family of domain foundation models, one hard scientific domain at a time, and open-sourcing each one.

That is the strategy I would bet on. Not one model that knows everything, but a fleet of specialists that each deeply know one thing. It works for the Moon. It works for the Sun. It works, in a far more mundane way, for your support inbox.

Frequently asked questions