The most valuable map of the Moon may not show roads or borders. It may show shadows.
Near the lunar poles, sunlight never reaches the floors of some deep craters. Temperatures there are cold enough for water ice to survive for extraordinary lengths of time. To future explorers, that ice could be drinking water, oxygen—and eventually hydrogen and oxygen for rocket fuel.
Finding it is difficult. The Moon has been photographed and measured for decades, but the evidence is spread across different instruments, wavelengths and missions. A bright patch on one map may need to be compared with temperature, radar and terrain data from several others.
NASA and IBM now want artificial intelligence to do more of that comparison. On 10 September they released the NASA-IBM Lunar Foundation Model, an open-source system trained on more than 30 layers of data from nine instruments aboard four NASA missions, including the Lunar Reconnaissance Orbiter. Reuters reported the launch.
An AI that reads maps instead of sentences
The phrase “foundation model” usually brings to mind a chatbot trained on enormous quantities of text. This model is different. It learns patterns across layers of lunar observations.
Imagine laying transparent maps on top of one another: one for elevation, another for temperature, another for reflected light and another for the chemical clues in the soil. A human researcher can compare them, but doing so across the entire Moon is slow. A model trained to recognise relationships between those layers can rapidly flag places that deserve closer inspection.
NASA and IBM say the system can help identify possible ice deposits in permanently shadowed regions, map craters for landing-site analysis and study volcanic features. In benchmarks supplied by the organisations, it identified selected surface features up to 23% more accurately than widely used methods.
That “up to” is important. A benchmark improvement is not the same as a discovery on the Moon, and performance on one mapping task does not guarantee the same gain on every task.
Why lunar ice matters so much
Launching anything from Earth is expensive because a rocket must lift its own fuel as well as its cargo. If future crews can use lunar resources, they may not need to carry every litre of water and every kilogram of propellant from home.
Water can be split into hydrogen and oxygen. Oxygen supports breathing and can act as an oxidiser for rockets; hydrogen can be a fuel. Ice could therefore make a sustained lunar outpost more practical and potentially turn the Moon into a staging point for deeper missions.
But a map of “possible ice” is only the beginning. Researchers must determine how much ice is present, how deeply it is buried, what it is mixed with and whether it can be extracted without using more energy than it is worth.
The same caution applies to landing sites. AI can help rank terrain, but mission planners must consider slopes, boulders, lighting, communications, dust and emergency options. No responsible agency will select a crewed landing site because an algorithm produced a colourful heat map.
Open tools can widen the lunar research team
The model joins IBM and NASA’s Prithvi family of open foundation models. Making it publicly available matters because lunar science is no longer conducted by one agency or one country. Universities, startups and national space programmes can test the model, adapt it to new tasks and—just as importantly—find where it fails.
Open access does not guarantee trustworthy results. It does allow outside researchers to reproduce tests and compare the tool with other approaches. That is healthier than treating a model’s output as a secret score that cannot be examined.
FutureTechDose recently covered another new view of the Solar System in Hubble’s discovery of a ten-sided wave around Saturn’s south pole. The common thread is not simply better cameras. Modern astronomy increasingly depends on software capable of finding weak patterns inside immense archives.
What the model has not done
The lunar AI has not found a confirmed reservoir ready to supply astronauts. It has not chosen an Artemis landing site, and it cannot replace direct measurements from landers, rovers or drilled samples.
Its value is narrower and more practical: it can help scientists decide where to look next. That may sound less dramatic than an autonomous AI explorer, but space missions are shaped by limited time, limited power and limited chances. Reducing a million possible locations to a few hundred promising ones can be enormously valuable.
The Moon is not short of data. It is short of easy answers. NASA and IBM are betting that an AI trained to see connections across decades of observation can turn that archive into a better guidebook for the explorers who follow.
Evidence status: Publicly released research tool with organisation-reported benchmark results. It assists analysis; it has not independently confirmed new ice deposits or selected mission sites.
Primary/reliable sources:
Leave a Reply