Artificial intelligence is being shoved into seemingly everything these days, but NASA and IBM may have found a use for it that is genuinely difficult to dismiss. The two organizations have released an open source AI foundation model designed specifically to understand the Moon.
Called the NASA-IBM Lunar Foundation Model, the system was trained from scratch using a massive collection of lunar remote-sensing data. According to the model card on Hugging Face, its training data includes roughly 2 million lunar data bundles and more than 1 million high-resolution images, some with resolution around one meter per pixel.
This isn’t ChatGPT with a spacesuit. The model is designed to analyze scientific information about the lunar surface, combining different types of data that would otherwise require researchers to work across separate datasets and specialized tools.
The information fed into the model includes imagery, topography, slope, illumination geometry, radar observations, mineral information, gravity measurements, and hydrogen data. By learning relationships between these different sources, the AI can potentially help scientists identify patterns that humans might overlook.
NASA and IBM have already tested the model on several practical scientific tasks. Those include detecting lunar craters, analyzing volcanic features, and identifying locations that could have a higher likelihood of containing water ice.
That last application could become particularly important if humans are serious about spending extended periods on the Moon. Water ice isn’t merely useful for giving astronauts something to drink. Water can be separated into hydrogen and oxygen, potentially providing breathable oxygen and ingredients for rocket propellant.
In other words, finding accessible lunar ice could eventually help determine where humans establish infrastructure on the Moon. Rather than manually examining enormous quantities of lunar observations, researchers could use AI to help narrow the search.
There are measurable results behind the project too. In an ice prospectivity benchmark, the lunar model achieved a root mean squared error of 0.0293, compared with 0.0377 for the strongest baseline reported in the model documentation. On one crater detection benchmark, it reached a mean average precision of 0.2581 compared with 0.2420 for the best baseline.
The AI doesn’t win every contest, however. On another meter-scale crater detection benchmark, a baseline model scored 0.1552 while the NASA-IBM model reached 0.1543. Those numbers are extremely close, but they are a useful reminder that a foundation model doesn’t automatically beat specialized alternatives at every task.
IBM says the model can outperform commonly used approaches by as much as 23 percent on certain lunar mapping tasks. The “as much as” part is important. That doesn’t mean this AI suddenly understands the Moon 23 percent better than scientists or every competing system.
There is an even bigger limitation. NASA and IBM aren’t telling anyone to hand important lunar mission decisions over to AI. The model documentation explicitly warns that it has not been validated for operational tasks such as landing-site certification or hazard clearance.
Its predictions about ice require similar caution. The system estimates ice prospectivity. It doesn’t look at a patch of lunar dirt and definitively declare that water is sitting underneath it. Scientists would still need observations, analysis, and ultimately physical evidence to confirm what is actually there.
What makes the project more compelling is that NASA and IBM aren’t keeping the technology locked away. The model has been released through Hugging Face under the Apache 2.0 license, allowing researchers and developers to experiment with it, modify it, and build upon the work.
That openness could prove just as important as the model itself. One AI system isn’t going to build a lunar base, find every deposit of water, or tell astronauts where to land. Giving researchers around the world a common foundation for analyzing lunar data, however, could accelerate the scientific work needed to answer those questions.
There is plenty of justified skepticism surrounding AI right now, particularly when companies bolt chatbots onto products that worked perfectly well without them. This is different. NASA has accumulated an enormous amount of information about the Moon, and humans have limited time to examine it all.
If AI can help scientists make sense of that data faster, locate overlooked geological features, and identify promising places to search for water, it could become a useful tool in humanity’s return to the lunar surface.
AI probably isn’t going to put humans on the Moon by itself. But this open source model could help us figure out where to go once we get there.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz