One of the biggest challenges of astronomy is also the most obvious: space is large and it takes a long time to analyze everything. This is why artificial intelligence has been a great help for this science. It turns out that the same artificial vision tools developed for tasks like driving cars without a driver are also perfect for classifying large amounts of astronomical data. So much so, astronomers announced this month that they had used AI to find 6,000 new craters on the Moon.
Now, this is not so significant in itself. It is estimated that the Moon has hundreds of thousands of craters, mainly caused by impacts with asteroids and meteors. This is due to some factors. First, because the Moon has no atmosphere, these objects have a free path to the surface (unlike Earth, where the friction of the air slows them down and reduces them in size); and second, because there is no climate on the Moon, the marks they leave are not softened by erosion. The final result is the crater face satellite we all know.
But using AI to find these craters is important, as it demonstrates another way in which machine learning can automate a labor-intensive task. The less time astronomers spend photographing the moon, labeling the craters by hand, the more they can focus on other, more challenging investigations. Also, the more we know about the craters of the Moon, the better we can theorize about the history and formation of our Solar System.
A sample image showing the performance of the neural network. The blue circles are craters identified by humans that the network detected successfully; the red circles are new craters that the network found; and the purple ones are those that were lost.
The tool used for this particular research is what is known as a convolutional neuronal network, or CNN. This is a common technique that is particularly good for classifying visual data. As explained by the researchers who did the work in an unpublished article, they trained their network using a set of data from craters previously identified by humans. Once the program learned what the craters looked like, it was released into a new section of the surface of the Moon (approximately one third of its total surface). There he found 6,000 new craters.
As the scientists who did the work, from the universities of Toronto, Penn State and Arizona State, write in their newspaper, the system was consistent and, most importantly, fast. "Once trained, our CNN greatly increases the identification speed of the crater, taking minutes to generate predictions for tens of thousands of lunar DEM," they write. (A DEM is a digital elevation map and is the type of standard image used to find and classify craters.) "This, of course, everything is done passively, freeing the scientist to perform other tasks."
This system was not perfect, and when it was tested against a human crater observer, it only found 92 of the same characteristic. But this work shows that AI is an extremely capable tool that can accelerate basic astronomical research. To date, similar methods have been used to detect gravitational lenses, discover new exoplanets, identify pulsating stars and classify galaxies. Space may be large, but humans now have computers to help them sift through the cosmos.