In a first for astronomy, scientists trained a neural network to sift through scads of data from a planet-hunting telescope, and it found a whole new world.
Dubbed Kepler-90i, the newfound planet had been hiding in the buckets of data gathered by NASA’s Kepler spacecraft. It joins seven other planets circling a star roughly 2,500 light-years away, which means the Kepler-90 system ties our own planetary family for hosting the most known worlds.
“Kepler has already shown us that most stars have planets,” NASA’s Paul Hertz said during a press conference revealing the discovery. ”Today, Kepler confirms that stars can have large families of planets, just like our solar system.”
In the days leading up to the press conference, speculation ran wild about a possible detection of alien life. While that is unsurprisingly not the case, the announcement does showcase how machine learning can help us learn more about potentially exciting worlds across the galaxy.
A Hundred Thousand Things To See
Launched in 2009, the Kepler spacecraft spent four years staring at 150,000 stars in a single patch of sky. Its mission was to watch for brief dips in starlight caused by planets crossing the faces of their stars. When they find such a signal in the data, scientists can also figure out how big a planet is and how far it orbits from its star.
Kepler has identified 2,525 planets so far, and more are likely waiting to be found in the data. But it’s not an easy task to confirm a planet as real. Sorting through the massive amount of Kepler data by hand is an impossible task for humans—those data contain several quadrillion possible planetary orbits. And just because a star dims doesn’t mean a planet is the culprit: Starspots, stellar partners, and other objects can mimic a planet’s fingerprint.
That’s why Google AI’s Chris Shallue decided to use neural networks to tackle the problem. Machine-learning approaches have been used to filter and classify Kepler data before, but Shallue’s neural network offered a far more powerful algorithm.
“I became interested in applying neural networks to astronomy when I learned that the Kepler mission had collected so much data that it was impossible for scientists to examine it all manually,” he says. “Our idea was to turn this technique to the skies and teach a machine learning system how to identify planets around faraway stars.”
A New Fantastic Point of View
As the name implies, neural networks are based on the way a human brain works. They can be trained to identify and classify things, such as the difference between photos of dogs and photos of cats. Eventually, after looking at enough examples, the computer is able to classify cats and dogs on its own.
Shallue trained a network to recognize the distinct fingerprints of planets. He pulled 15,000 actual planet signatures from the Kepler data set and started training the system to tell the difference between real worlds and signals that can masquerade as planets. When he tested how well his machine learned, he found that it correctly identified planets 96 percent of the time.
Then it was time to put the algorithm to work for real. Shallue and Andrew Vanderburg, of the University of Texas at Austin, asked it to scrutinize some 670 stars in the Kepler field that already had known planets, since planets are more likely to exist in multiples.
Then they fed it signals from those systems that were not considered strong enough for humans to manually examine. In those signals, the machine identified two new planets, which are described in a study that will be published in The Astronomical Journal.
“These two planets have weak signals that were missed in all previous searches of these stars,” Shallue says.
With New Horizons to Pursue
One of the planets, Kepler-80g, is the sixth known world in its system. Roughly the size of Earth, Kepler-80g takes 14.6 days to orbit its star, which is smaller and redder than the sun.
The neural network also recognized Kepler-90i. Slightly bigger than Earth and with a year that lasts just two Earth-weeks, this planet is also the third rock from its sun, a star that’s a bit bigger and hotter than the sun. Two similarly small planets circle the star closer than Kepler-90i, while the planets orbiting farther from the star get progressively bigger.
Though this cast of otherworldly characters is large, they’re all bunched together—all eight planets are squeezed within the same distance from their star as Earth is from the sun.
“Kepler-90i is not a place I’d like to go visit,” Vanderburg says. “The surface is likely scorching hot, we calculated that it probably has an average temperature of about 800 degrees Fahrenheit.”
He adds that Kepler-90 may host even more planets yet to be discovered. He and Shallue are planning to run the entire Kepler data set through their neural network and see what pops out. But don’t worry too much about computers replacing actual human astronomers just yet.
“This will absolutely work alongside astronomers,” says NASA’s Jessie Dotson. “You’re never going to take that piece out. You’ve gotta have those initial classifications in order to train your machine learning—and then it can go through a lot more signals than humans can.”