NASA's Apollo missions proved humans could reach the Moon. The Artemis program is testing whether we can stay.

Staying requires knowing the ground. And the Moon, it turns out, is badly mapped — not for lack of data, but because the data doesn't fit together.

On September 10, IBM Research and NASA did something about it, and then gave the result away.

This is the part that makes the achievement legible.

NASA's GRAIL mission mapped the Moon's gravitational field at 20 kilometres per pixel, to visualise its crust and subsurface. The Lunar Reconnaissance Orbiter images small boulders and crater rims at 1 metre per pixel.

That's a 20,000-fold difference in scale, measuring different physical properties, from different angles, on different missions. Add temperature readings, topography, and decades of imagery from both US and Japanese spacecraft.

Historically every scientific question got its own custom pipeline, because there was no shared representation to ask questions of.

To build one, IBM and NASA adapted TerraMind, the Earth-observation model IBM developed with the European Space Agency, chosen because it "excels at integrating different data types and resolutions and learning cross-modal correlations to fill in missing or noisy values."

It joins IBM's earlier models of the Earth (Prithvi EO) and the Sun (Surya). The Moon is the third world they've done this to.

A few details from the blog that explain why this isn't just a data-plumbing exercise:

The lighting is brutal. A lunar 'day' is two weeks of sunlight followed by two weeks of darkness, which dramatically changes how surface features look. With almost no atmosphere to scatter light, the Moon is "a land of sharp edges and stark contrasts, its peaks and valleys obscured by dark shadows and bright glare."

The poles are worse. The Moon is barely tilted relative to its orbit, so at the poles the Sun hangs permanently low on the horizon, casting long shadows that hide rocks, craters and other hazards.

The temperatures are absurd. As high as 250°F (121°C) in full sun; as low as -410°F (-246°C) in shadowed craters. Those shadowed craters are exactly where the ice would be — and exactly where you can see least.

The dust fights back. The Moon is covered in regolith, an abrasive sticky layer that distorts appearance depending on the Sun's angle, "in addition to tearing space suits, damaging lung tissue, and wrecking equipment."

So: the most scientifically valuable places are the least visible ones. That's the problem the model is aimed at.

Three priorities:

Craters are a clock — you can tell which parts of the Moon are oldest by counting them. More than 2 million large ones are catalogued; NASA wants the small ones, especially at the poles. Polar craters could provide water and shelter for astronauts and their crops, while high crater rims in near-constant sunlight could charge solar panels.

Volcanoes are a live scientific argument. Lunar volcanism was thought to have died out a billion years ago, but features called irregular mare patches may be far younger. "The ages of these features remain a matter of great debate," says NASA's Michael Barker, who co-led the project with IBM, "so the more we can understand their distribution and properties, the better chance we have of resolving this mystery."

Ice is the practical prize, and worth its own section.

For decades the Moon was thought to be bone dry, based on Apollo samples.

Then in 2008, a NASA radar instrument flying aboard India's Chandrayaan-1 spacecraft detected ice in 40 small craters near the north pole. "NASA put the amount of water at 600 million metric tons — enough to fill at least 240,000 Olympic-sized swimming pools."

Since then, NASA's retired SOFIA observatory found water molecules stuck to grains of lunar dust, and NASA estimates each cubic metre of lunar soil could hold as much as a 12-ounce bottle of water. The water was likely delivered over billions of years by comets and asteroids, or created by solar wind interacting with the soil.

Why this matters more than it sounds: water is drinking water, breathable oxygen, and rocket propellant. Every kilogram you don't have to launch out of Earth's gravity well changes the economics of a Moon base from implausible to arguable. It's the difference between visiting and staying.

But you can't use what you can't find — and it's hiding in craters that have not seen sunlight in billions of years.

On ice:

It makes that prediction by combining temperature and topographic data — slope, aspect, ice-stability depth, and maximum surface temperature.

On craters, and note the order here:

It matched a specialist Swin crater-detection model at 1-metre resolution. Then, at the coarser 100 metres per pixel, it "outperformed the same model by nearly 19% using half the training data."

On volcanic features, it beat a task-specific model by 3%, "bringing comparable accuracy at lower cost."

There's also a nice demonstration of crater detection in the wild: the model picks out a brand-new crater formed by the crash of a SpaceX rocket on August 5, near the Moon's Einstein Crater, highlighted against a field of already-documented ones.

Now the sentence that I think is the actual story:

Ninety percent untouched. And with that light a touch, a general model matched or beat models purpose-built for each individual task — while needing half the labelled training data.

That last part is the one to hold onto. In most scientific machine learning, labelled data is the binding constraint. You cannot crowdsource annotated lunar imagery; someone with real expertise has to sit down and label it. Halving that requirement doesn't just save money. It changes which questions are cheap enough to be worth asking.

This is the same shift that already happened in language and images, now arriving in planetary science: stop building a model per task; build a representation, and adapt it.

For context on why several countries are interested: four others besides the US have made soft landings and announced intentions to establish a presence. Lunar soil contains helium-3, a rare isotope and a potential fusion fuel, and some rocks are rich in rare earth elements.

And then there's astronomy. The Moon's far side has no earthquakes, no magnetic field, and little satellite interference.

Proposed projects lean on the Moon's bowl-shaped craters in wonderfully literal ways: three radio telescopes scheduled to land in the next few years to listen to the early universe; a NASA proposal to turn a nearly mile-wide crater into a giant antenna dish; a gravitational wave detector strung across a crater to listen for colliding black holes; an infrared telescope dropped into an ultracold polar crater to study exoplanet atmospheres.

A better map serves all of it.

Every performance number above comes from IBM and NASA, published on IBM's own research blog.

That isn't a reason to dismiss it. But it's a different thing from an independently replicated result, and it's worth saying plainly rather than letting percentages sound like verdicts. To their credit, the model is open-sourced, the technical report is published, and the benchmark datasets are available — so this is checkable in a way that most corporate AI announcements are not. It simply hasn't been checked yet by anyone outside.

"Open source" and "verified" are different claims. This is the first.

This is one of the better AI stories to bring to a kid, for a reason that isn't the Moon.

Ask them this: why is finding ice on the Moon hard?

The intuitive answer is "space is far away." The real answer is stranger and better — the ice is in craters at the poles where the Sun never reaches, so the most valuable places are the darkest ones. You are looking for something precisely where you can see least.

That's a genuinely lovely problem, and the solution isn't a better camera. It's combining clues: how cold is it, which way does the ground slope, how deep would ice stay stable, how hot does the surface get at its worst. No single measurement answers the question. Together they narrow it down.

That's what this model does, and it's a good mental picture of what AI is actually useful for — not knowing things, but combining many partial, noisy, mismatched clues into a decent guess about where to look next.

Two more that land well with kids:

And the ending is a good one: they gave the model away. Anyone can download it. A student with a laptop and some patience can now work with the same lunar model NASA is using.

Source: "Introducing IBM and NASA's new foundation model for the Moon," IBM Research, by Kim Martineau: https://research.ibm.com/blog/nasa-ibm-lunar-foundation-model

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter