On December 14, 2017, NASA announced that a deep neural network had discovered a previously missed planet in data from the Kepler space telescope. Christopher Shallue, a software engineer at Google, and Andrew Vanderburg, then a NASA Sagan Postdoctoral Fellow at the University of Texas at Austin, trained a convolutional neural network to tell genuine transiting planets from false positives in Kepler light curves.
The model learned from about 15,000 previously vetted Kepler signals and could rank a plausible planet signal above a false positive 98.8 percent of the time. The team then pointed it at 670 star systems already known to host multiple planets, where weak signals were most likely to have been overlooked. It surfaced Kepler-90i, a rocky planet about 30 percent larger than Earth that orbits its star every 14.4 days with an average surface temperature above 800 degrees Fahrenheit.
The find brought the Kepler-90 system to eight known planets, tying it with our own solar system as the star then known to host the most planets. The same network also identified Kepler-80g, an Earth-sized sixth planet in a five-planet resonant chain. The result is a landmark example of machine learning recovering signal that human-tuned pipelines had passed over.