A rough guide for going back to the Moon

NASA’s Apollo missions proved that humans could make it to the Moon. The Artemis program now underway will test whether we can live and work there, too. NASA’s long-term vision calls for building a base on the Moon where astronauts can carry out research, test new technologies, and prepare for a push onward to Mars.

A good map is handy when exploring any unfamiliar place, but when the terrain is as rugged and otherworldly as the Moon’s, it’s essential. Today, IBM and NASA are open-sourcing the most thorough model for mapping the Moon to date. It consolidates and harmonizes decades of data collected on US and Japanese missions flown over our closest celestial neighbor.

Called NASA-IBM Lunar Foundation Model, it’s the first AI model to integrate observations captured in a range of modalities, and at different viewing angles and spatial scales. By distilling these disparate measurements into one representation, researchers have created a model for the Moon that, like earlier models of the Earth and Sun, Prithvi EO and Surya, can be quickly adapted to different tasks and reused.

A lunar AI model has many potential uses, but NASA has initially prioritized three: mapping the smaller, uncatalogued craters that pockmark the Moon’s surface; investigating its volcanic history; and scouring craters at both poles for ice, which could provide future astronaut crews with a source of drinking water, oxygen, and fuel.

“Experienced travelers know to get the lay of the land before setting out for a foreign destination,” said Juan Bernabé-Moreno, director of IBM Research Europe for Ireland and UK. “We hope that our AI model can help the science community explore the lunar landscape and help the next generation of astronauts find their way around before heading into space.”

A Rosetta stone for multi-sensor data

The new model addresses a longstanding challenge for lunar scientists handling high-volume, multi-sensor data at varying scales. NASA’s Gravity Recovery and Interior Laboratory (GRAIL) mission, for example, mapped the Moon’s gravitational field at a scale of 20 kilometers-per-pixel to visualize its crust and subsurface while NASA’s Lunar Reconnaissance Orbiter (LRO) swooped in to hunt for ice nestled in the Moon’s dark polar craters and to image small boulders and crater rims at scales of 1 meter-per-pixel.

To align such a diverse collection of measurements, IBM and NASA looked for an AI architecture fluent in multiple scientific dialects. They went with a version of TerraMind, the Earth-observation model developed by IBM and the European Space Agency (ESA). TerraMind excels at integrating different data types and resolutions and learning cross-modal correlations to fill in missing or noisy values.

One area in which scientists will look to the new AI model for help is understanding the Moon’s difficult lighting conditions. A lunar ‘day’ consists of two weeks of sunlight followed by two weeks of darkness, which can dramatically change the appearance of the Moon’s surface features.

This disorienting effect is amplified by the Moon’s negligible atmosphere. Without clouds or a thick atmosphere like Earth’s 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 Moon is covered in an abrasive, sticky layer of dust and rock called regolith that can also create visual distortions depending on the Sun’s angle (in addition to tearing space suits, damaging lung tissue, and wrecking equipment).

At the poles, the landscape gets even blurrier. Unlike the Earth, the Moon is only slightly tilted relative to its path around the Sun. This means the Sun hangs low on the polar horizon, causing the terrain to cast long shadows which can obscure rocks, craters, and other dangers from astronauts.

Unusual lighting isn’t the only complication tied to lunar data. Scientists often want to bring coarse satellite data into sharper focus. Take lunar weather. The long lunar day combined with the lack of a climate-regulating atmosphere can push temperatures on the Moon as high as 250°F (121°C) in full sun and as low as -410°F (-246°C) in its shadowed craters.

Though scientists generally know where the extremes are, a detailed heat map could aid returning astronauts in finding ice or safer spots to build a base. Scientists have run simulations to try and bump up the resolution of NASA’s thermal data, but such simulations take time, money, and significant computational resources. A higher resolution heat map produced with a lunar AI model like IBM and NASA’s could help to quickly isolate compelling sites to study further with simulation.

Fire, falling asteroids, and ice

Scientists have long probed the Moon’s origins for clues about our home planet. The leading theory suggests the Moon was born in a spectacular collision with Earth at the dawn of our solar system. Rocks brought back by Apollo astronauts provided evidence that a Mars-sized object smashed into Earth about 4.5 billion years ago, flinging enough molten debris into space to create the Moon.

For hundreds of millions of years, a deep ocean of magma covered the Moon’s surface. The magma eventually cooled enough to form a crust, while continued volcanic activity below periodically brought fresh lava to the surface. The volcanic activity is thought to have died down about a billion years ago but recent evidence suggests that some volcanic features, called irregular mare patches, may be significantly younger.

“The ages of these features remain a matter of great debate, so the more we can understand their distribution and properties, the better chance we have of resolving this mystery,” says Michael Barker, a NASA expert on lunar topography who co-led the project with IBM.

The Moon has neither tectonic plates to heave new rock to the surface nor wind and water erosion to break down mountains and erase the past. Nevertheless, asteroids and micrometeoroids are constantly bombarding its surface and reshaping the terrain. You can tell which parts of the Moon are oldest by counting its craters.

Scientists have globally catalogued more than 2 million large craters so far, and NASA wants to document the smaller ones remaining, especially at the poles. Polar craters could provide water and shelter for astronauts and their crops, while high crater rims immersed in near-constant light could charge solar panels to power equipment and scientific instruments.

For decades, the Moon was thought to be an arid wasteland from the rocks brought home by Apollo astronauts. Then, in 2008, a NASA radar instrument flown 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.

Several missions since then have found water in other forms and places. More recently, NASA’s retired SOFIA observatory discovered water molecules stuck to grains of lunar dust. NASA estimates that each cubic meter of soil across the Moon’s surface could hold as much as a 12-ounce bottle of water. Water was most likely delivered to the Moon over billions of years by falling comets and asteroids, or by solar winds interacting with the lunar soil.

The next wave of lunar exploration

IBM and NASA’s open-source foundation model is a first pass at translating petabytes of satellite data into a fast and flexible way of modeling the Moon’s surface. To fine-tune the model, the researchers used lightweight low-rank adapters (LoRAs) that left 90% of the base model’s weights frozen. Despite the light touch, the model matched or outperformed traditional machine-learning models at a range of tasks.

Faced with the challenge of identifying from an unfamiliar image whether a dark polar crater might contain ice, the NASA-IBM model reduced its error rate by 22% compared to a state-of the-art SwinV2 transformer trained for ice prospecting.

The NASA-IBM model also excelled at crater detection. It could pick out craters at one-meter resolution (where each pixel in an image represents a square meter) as accurately as a Swin model trained for crater detection. At a coarser resolution of 100 meters-per-pixel, the NASA-IBM model outperformed the same model by nearly 19% using half the training data.

And when it came to mapping the extent of those mysterious volcanic features, the NASA-IBM model outperformed a task-specific Swin model by 3%, bringing comparable accuracy at lower cost.

The US isn’t the only country with ambitious plans for the Moon. Four others have made soft landings there and announced intentions to establish a presence. Minerals are part of the draw. Lunar soil contains helium-3, a rare isotope created by the solar wind and a critical ingredient for making clean nuclear fusion energy. Certain rocks are also rich in rare earth elements widely used in cellphones and other electronics.

Then there’s science. Astronomers have long dreamed of observing deep space from the Moon’s vantage point. There’s neither earthquakes nor a magnetic field to rattle instruments. And its far side is largely free of satellite interference.

“You’d have to go out past Neptune to find a place this radio-quiet,” said Martin Elvis, an astronomer at the Center for Astrophysics I Harvard and Smithsonian.

On a recent call, Elvis and a colleague, Malgorzata Sobolewska, ticked off a list of proposed projects that ingeniously capitalize on the Moon’s bowl-shaped craters.

Three radio telescopes are scheduled to land in the next few years to listen in to the early universe, a time before planets or stars. One NASA proposal even calls for turning a nearly mile-wide crater into a giant antenna dish. Other projects range from stringing a supersensitive gravitational wave detector across a crater to listen for colliding black holes and dropping an infrared telescope in an ultracold polar crater to study exoplanet atmospheres in the Milky Way.

The US-led Artemis Accords in 2020 established a set of principles to guide space exploration and cooperation. IBM and NASA’s open-source model can be seen as extending that spirit of discovery and good will.

New ways of modeling the Moon

The lunar foundation model is part of a wider shift in creating new scientific knowledge. New computing technologies can not only accelerate familiar calculations, but they can prompt entirely new investigations. Here, AI brings fragmented observations into a common representation, helping scientists uncover patterns, identify uncertainty, and decide where to focus next.

Over time, this model could become part of a broader scientific-computing workflow. Classical supercomputers could continue to handle most workloads, with AI helping to organize observations, build faster approximations of complex systems, and identify promising hypotheses. As quantum computing matures, quantum processors could contribute to simulation and optimization problems that are too difficult or time-consuming for classical machines to handle.

“The next lunar breakthrough could come not from a single new instrument, but from algorithms that allow many instruments — and eventually, computing paradigms — to work together,” said Bernabé-Moreno.