On December 14, 2017, NASA and Google announced that a neural network had discovered an eighth planet orbiting the Sun-like star Kepler-90, about 2,545 light-years away. The find, named Kepler-90i, tied that system with our own solar system for the most planets known around a single star. The work was led by Christopher Shallue, a Google AI software engineer, and Andrew Vanderburg, then a NASA Sagan Fellow at the University of Texas at Austin, and was accepted by The Astronomical Journal.
NASA’s Kepler space telescope hunts planets by watching for the tiny dimming when a planet crosses, or transits, its star. Sorting real transits from noise and false signals is hard to do by hand across hundreds of thousands of stars. Shallue and Vanderburg trained a neural network on a set of 15,000 previously vetted Kepler signals to learn the difference between true planets and false positives, reaching 96 percent accuracy on a held-out test set. They then pointed it at 670 stars already known to host multiple planets, looking for weaker signals humans had missed.
In those faint signals the network turned up two new worlds: Kepler-90i and Kepler-80g. Kepler-90i is a rocky planet about 30 percent larger than Earth, orbiting its star every 14.4 days with a surface estimated to exceed 800 degrees Fahrenheit. The discovery was an early, well-publicized example of deep learning extracting findings from an existing scientific archive rather than from new observations.