Scientists have applied a brand of artificial intelligence to data from the exoplanet-hunting Kepler satellite, resulting in the discovery of the first eight-planet system outside our own.
Neural networks, an application of artificial intelligence, can do a little bit of just about anything, from diagnosing breast cancer to driving cars. Now, they can help detect planets, too. A unique collaboration between an astronomer and a Google engineer has resulted in a neural network that combed through a preliminary sample of Kepler-detected exoplanets, making two Earth-size discoveries along the way.
Kepler-90i is a likely rocky planet belonging to the first eight-planet system outside our own. It’s a scrunched system: All eight planets would fit inside Earth’s orbit. Kepler-80g is the sixth planet in its system, orbiting its star in resonance with four of its companions. The resonance keeps the orbits stable, even though they’re even more tightly packed than Kepler-90.
Both planets are Earth-size, but given their roughly 14-day orbits, neither one offers a pleasant place to live. Surface temperatures are estimated at 710K (820°F) and 420K, respectively.
These are only two of the 2,525 confirmed planets discovered by Kepler to date. But it’s not just about the final tally: The newest finds are important for the possibilities they promise.
In Search of Earth Twins
Kepler was launched in 2009 with the goal of detecting an Earth-size planet at an Earth-like distance from a Sun-like star — Earth’s twin, in other words. But the main Kepler mission was cut short after four years, and Earth-twins remain at the very edge of what Kepler can detect using traditional analysis.
Here’s what the traditional method looks like: Kepler monitors a star, recording its brightness over time as a light curve. Then an algorithm looks for repeating dips in the light curve that could indicate a transiting planet. Humans then sort through the results — some 30,000 signals in all — and cull those coming from non-planetary sources such as binary stars.
But what if we could examine all the signals that Kepler finds, even ones that the pipeline doesn’t deem terribly likely to be a planet? Some real planets are probably hiding in those signals, right on the edge of detection, but the vast amount of data is simply too overwhelming for humans to contemplate.
A Neural Network to Find Them All
Big Data is something Google engineer Christopher Shallue is familiar with — he used to help the company target display ads on Gmail and Google Maps. Now he works for Google Brain, the company’s division for artificial intelligence. When he read about the massive scale of data Kepler was collecting, he wondered if AI could help identify new planets. He approached astronomer Andrew Vanderburg (University of Texas, Austin) to find out.
Neural networks are computer algorithms loosely based on the myriad connections between neurons in the brain. Each “neuron” is a simple mathematical formula that acts like a switch, turning on when it recognizes a certain pattern, and each set of neurons passes its results on to the next layer. Combine enough layers, along with training data to teach the machine what to recognize, and you can classify an image as a cat or a dog — or a light curve as a planet or not.
Shallue built a neural network and trained it to recognize planets using 15,000 Kepler light curves previously studied by humans. Then Vanderburg ran light curves from a preliminary sample of 670 stars through the standard pipeline to identify signals, even weak ones, to feed to the algorithm.
The neural net caught four likely planets (“likely” in this case means that the algorithm was correct 96% of the time in tests). Two of these, Kepler-90i and Kepler-80g, appear to be real planets — there’s only 1 chance in 10,000 that they’re not. The results will appear in the Astronomical Journal (full text here).
Granted, Kepler-90i and Kepler-80g are both too scorched to be Earth twins, but they come from a small selection of Kepler data. Shallue and Vanderburg want to expand their data set to include all 150,000 stars in the Kepler field.
Still, they hesitate to speculate on how many more planets they might unearth. “It’s hard to guess how many we’d find because they’re in a new regime,” Vanderburg says. “These are the smallest planets at longer periods.”
The duo also cautions against extrapolating these results to other missions, and even into deeper Kepler data. Neural networks may be called artificially intelligent, but the algorithm is only as good as it was taught to be, and it cannot generalize what it’s learned in the same way that humans can. If astronomers want to apply similar methods to data collected by the upcoming Transiting Exoplanet Survey Satellite (TESS) mission, set to launch in 2018, they’ll need to build another algorithm and train it anew.
The Eight-Planet System
Kepler-90i orbits a star that’s a little more massive and brighter than the Sun. As in the solar system, and unlike many systems that Kepler has discovered, the planets around this star are segregated — the smaller ones are closer in, the bigger ones farther out.
At first glance, that seems to agree with standard models of how planets form — larger gas giants form farther out where molecules can condense into ice. But all eight planets in Kepler-90 orbit well within the snow line.
“The planets must have migrated inward,” Vanderburg says. Early in the system’s history, the planets may have interacted with the protoplanetary disk from which they formed, slowly spiraling inward and maintaining their order. But, he cautions, “we don’t entirely understand the process of planet formation and planet migration.”
Not to mention, there’s still a lot we don’t know. While Kepler has detected eight planets in the Kepler-90 system, there could very well be more beyond Kepler’s measurements. Vanderburg offers a tantalizing possibility: “It would almost be surprising to me if there weren’t more planets around this star. . . .We’re just scratching the surface.”