NASA’s recent discovery of another exoplanet, called Kepler-90i, about 2,545 light years away, is significant for a few reasons. It furthers the work of the space agency’s Kepler spacecraft, launched in 2009 to search for other planets that could be suitable for life. It also marks the first discovery of a solar system with a number of planets that matches our own solar system’s eight (something fans of Pluto are still trying to change).
And while it continues a process of space discovery based not so much on looking at the stars, but looking at the data, it marked the first discovery resulting from a deep learning neural network — the same kind used to distinguish cats and dogs in Google photos — that caught what human scientists or other automated methods had missed.
“Just as we expected, there are exciting discoveries lurking in our archived Kepler data, waiting for the right tool or technology to unearth them,” Paul Hertz, director of NASA’s Astrophysics Division in Washington, said in a NASA announcement. “This finding shows that our data will be a treasure trove available to innovative researchers for years to come.”
Signs of (Possible) Life
NASA launched Kepler in 2009 with the intention of a 3.5-year mission, which was extended and later adjusted into a second mission called K2 after two of the spacecraft’s four reaction wheels used to direct the spacecraft failed. The goal is to search regions of the Milky Way galaxy for Earth-sized exoplanets in habitable zones — not too close or too far from a star — that could support life.
To date, Kepler has discovered more than 1,000 exoplanets in about 440 star systems (four discovered under K2), though not all lie in a habitable zone. (Kepler-90i, by the way, doesn’t qualify as it’s so close to its star it orbits every 14.5 days and has an average surface temperature of about 800°F. It’s a high-speed pizza oven without any atmosphere. The system's outermost planet, Kepler-90h, is about the same distance from its star as the Earth is to the sun, though NASA says other star systems are likely more promising.)
Kepler uses a photometer to identify planets by the “dimming effect,” detecting a reduction of light from a star when a planet passes in front of it. While size can make some planets relatively easy to spot this way, the process has some vagaries, such as interactions with other planets, that can result in false positives. Sifting through the noise in all the collected data to find valid evidence of planets is where machine learning comes in.
NASA describes K2 as a community-driven mission, inviting outside researchers to peruse its data through its Guest Observer program. The mass of available data collected by Kepler attracted Christopher Shallue, a software engineer with Google AI, who worked with Andrew Vanderburg, a NASA Sagan postdoctoral fellow and astronomer at the University of Texas at Austin, to apply a neural network to the search.
What Neural Networks Can See
“Machine learning really shines in situations where there is so much data that humans can't search it for themselves,” Shallue said. And although machine learning had been used before to rummage through Kepler’s data, Shallue and Vandenberg brought a deep learning neural network to the problem, as they explain in a research paper by Shallue and Vendenburg accepted for publication in The Astronomical Journal.
After training the network on 15,000 previously vetted signals from Kepler’s catalogue, and managing to eliminate false positives 96 percent of the time, they started looking for weaker signals in 670 star systems that already had multiple known planets.
“It’s like sifting through rocks to find jewels,” Vandenberg said. “If you have a finer sieve, then you will catch more rocks, but you might catch more jewels, as well.”
Between Kepler and other methods, more than 3,000 exoplanets have been confirmed, and Kepler’s data alone has identified more than 3,000 other possibilities. But the search for alien life is just in its infancy.
Astronomers have estimated the Milky Way could hold 40 billion Earth-size planets in habitable zones around either sun-like stars or red dwarfs. The closest of them are about 12 light years away. Any close encounters with intelligent life are likely still far off. Meanwhile, the Kepler researchers plan to refine their neural network in the hope of boldly going where no data extraction has gone before.