AI has found an 8-planet system like ours in Kepler data

Our solar system is no longer the sole record-holder for most known planets circling a star.

An artificial intelligence algorithm sifted through data from the planet-hunting Kepler space telescope and discovered a previously overlooked planet orbiting Kepler 90 — making it the first star besides the sun known to host eight planets. This finding, announced in a NASA teleconference December 14, shows that the kinds of clever computer codes used to translate text and recognize voices can also help discover strange new worlds.

The discovery, also reported in a paper accepted to the Astronomical Journal, can also help astronomers better understand the planetary population of our galaxy. “Finding systems like this that have lots of planets is a really neat way to test theories of planet formation and evolution,” says Jeff Coughlin, an astronomer at the SETI Institute in Mountain View, Calif., and NASA’s Ames Research Center in Moffett Field, Calif.

Kepler 90 is a sunlike star about 2,500 light-years from Earth in the constellation Draco. The latest addition to Kepler 90’s planetary family is a rocky planet about 30 percent larger than Earth called Kepler 90i. It, too, is the third planet from its sun — but with an estimated surface temperature higher than 400° Celsius, it’s probably not habitable.  

Story continues below graphic

The seven previously known planets in this system range from small, rocky worlds like Kepler 90i to gas giants, which are all packed closer to their star than Earth is to the sun. “It’s very possible that Kepler 90 has even more planets,” study coauthor Andrew Vanderburg, an astronomer at the University of Texas at Austin, said in the teleconference. “There’s a lot of unexplored real estate in the Kepler 90 system.”

Astronomers have identified over 2,300 new planets in Kepler data by searching for tiny dips in a star’s brightness when a planet passes in front of it. Kepler has collected too much data for anyone to go through it all by hand, so humans or computer programs typically only verify the most promising signals of the bunch. That means that worlds that produce weaker light dips — like Kepler 90i — can get passed over. Vanderburg and Christopher Shallue, a software engineer at Google in Mountain View, Calif., designed a computer code called a neural network, which mimics the way the human brain processes information, to seek out such overlooked exoplanets.

Researchers previously automated Kepler data analysis by hard-coding programs with rules about how to detect bona fide exoplanet signals, Coughlin explains. Here, Vanderburg and Shallue provided their code with more than 10,000 Kepler signals that had been labeled by human scientists as either exoplanet or non-exoplanet signals. By studying these examples, the neural network learned on its own what the light signal of an exoplanet looked like, and could then pick out the signatures of exoplanets in previously unseen signals.

The fully trained neural network examined 670 star systems known to host multiple planets to see whether previous searches had missed anything. It spotted Kepler 90i, as well as a sixth, Earth-sized planet around the star Kepler 80. This feat marks the first time a neural network program has successfully identified new exoplanets in Kepler data, Jessie Dotson, an astrophysicist at NASA’s Ames Research Center said at the teleconference.

Vanderburg and Shallue now plan to apply their neural network to Kepler’s full cache of data on more than 150,000 stars, to see what other unrecognized exoplanets it might turn up.

Coughlin is also excited about the prospect of using artificial intelligence to assess data from future exoplanet search missions, like NASA’s TESS satellite set to launch next year. “The hits are going to keep on coming,” regarding potential exoplanet signals, he says. Having self-taught computer programs help humans slog through the data could significantly speed up the rate of scientific discovery.