The computer used telescope data to identify the planet while looking for exoplanets in the system. In October 2016, Shallue, who is interested in astronomy, made a decision to use about a fifth of his time at work (Google calls it 20% time, and allows its employees to pursue a pet project during those hours) to find a way to use machine learning to more quickly cipher through Kepler's data.
The discovery was based on observations gathered by Nasa's Kepler Space Telescope. But the similarities with our home planet probably end there: Kepler-90i completes one orbit every 14.4 Earth days and is therefore probably much too hot to host life. The planet is much closer to its star than the Earth is to the Sun, which means it is a sizzling hot planet with temperatures exceeding 800 degrees Fahrenheit.
The machine learning technique was also used to find a new Earth-sized planet, called Kepler 80g, around a different star.
That's one more than the previous extrasolar record, which had been held jointly by Kepler-90 and the TRAPPIST-1 system.
Lead astronomer on the project Andrew Vanderburg said he believes Kepler-90 may have more planets that haven't yet been detected.
Google helps NASA discover an eighth planet
"Maybe there are systems out there with so many planets that they make our eight-planet solar system seem ordinary". "Today Kepler confirms that stars can have large families of planets just like our solar system".
While machine learning has previously been used in searches of the Kepler database, this research demonstrates that neural networks are a promising tool in finding some of the weakest signals of distant worlds.
The Kepler Space Telescope launched in 2009, and has scanned some 150,000 stars. Kepler-90i is the eighth planet found around its parent star.
"This is a really exciting discovery, and we consider it to be a successful proof of concept to be using neural networks to identify planets, even in challenging situations where the signals are very weak", said Christopher Shallue, senior software engineer at Google in Mountain View, California. Doing so could spark interest among machine learning practitioners to become citizen scientists, "or encourage citizen scientists to use machine learning in their efforts as well", he said.
After the neural network was trained manually to identify passing exoplanets by analyzing 15,000 signals from the Kepler catalogue, it started giving correct results for true planets and false positives in 96 percent of cases.
As of December 14, Kepler has confirmed 2,341 exoplanets. Machine learning is an approach to artificial intelligence in which computers "learn". Light curves show how the brightness of a star drops off when an orbiting planet passes in front of it. "You have little planets inside and enormous planets outside, however everything is scrunched in substantially nearer", Vanderburg said.