èƵ

AI has spotted nearly 7000 undiscovered craters on the moon

Artificial intelligence can automatically identify craters on the moon, and may eventually tell us more about the formation of the solar system
A mosaic of the moon's made from hundreds of images
A mosaic of the moon’s made from hundreds of images
NASA/GSFC/Arizona State University

AI is looking skyward. An artificial intelligence algorithm has learned to identify craters on the moon, discovering thousands of new ones in the process.

Craters are geological fingerprints left behind by impacting space debris, such as comets or asteroids. Tectonic activity and weathering have wiped most of the Earth’s surface craters clean, but the moon is still covered in them—many of which date back billions of years. Even from afar, the size distribution and number of craters can be used to reveal more about how the solar system formed.

Yet the process of identifying and counting craters is laborious, inefficient and prone to inconsistency. So a team led by Ari Silburt at Penn State University and Mohamad Ali-Dib at the University of Toronto, fed 90,000 images of elevation maps spanning the surface of the moon into a convolutional neural network – a type of algorithm adept at categorising images.

The algorithm was trained to identify craters larger than 5 kilometres in diameter. First it would attempt to identify the edges, and then check the shape against a set of known craters. If a similar shape was already on the system, then it could be confident it was looking at a crater rather than an edge from a mountain or other planetary feature.

New finds

When the machine was tested on images covering a third of the moon, it found 6883 new craters in just a few hours, almost doubling the total number of known craters of this size ().

The AI wasn’t perfect however. It falsely identified a few craters, missed some of the larger ones, and in about a quarter of the findings the positions deviated noticeably from the truth. “A human makes a judgement call based on a variety of indicators with topography only one factor,” says Peter Schultz, at Brown University.

Nevertheless, with further refinement, the system could be used to accelerate crater counting significantly. “Once our model has improved a bit more, we can use it to discover the hundreds of thousands of currently unidentified craters below 5 kilometres,” says Silburt.

It’s these smaller craters that reveal the most about the material near the Earth and Moon in the early Solar System, as well as when craters formed, he says.