Scientists using advanced "neural network" machine-learning software to sift through thousands of weak, previously unstudied signals fromhave identified two new worlds orbiting distant suns, researches announced Thursday.
One of them is the eighth planet now known to be orbiting the star Kepler-90, the first solar system other than Earth's to host at least that many worlds.
"What's different about this discovery is that we used machine learning to help identify planets that were missed by previous searches of the Kepler data," said Christopher Shallue, a senior software engineer at Google AI in Mountain View, Calif.
"The key contribution of machine learning here is it was able to search a much larger number of signals than humans would have been able to do in a reasonable amount of time."
The Kepler satellite, launched in 2009, has discovered 2,525 confirmed, using a 95-megapixel camera to measure the very slight dimming of a target star's light as a planet moves across its face as viewed from the spacecraft.
Up to this point, planet detections required astronomers to focus on the stronger signals to find planet candidates that then were confirmed or rejected based on additional observations and analysis to root out "false positives" and make sure the data reflected the transit of an actual planet.
Shallue and Andrew Vanderburg, an astronomer and NASA Sagan Postdoctoral Fellow at The University of Texas in Austin, decided to use neural network software that mimics how the brain identifies patterns to recognizein signals that were too weak for normal human analysis.
The researchers "trained" their neural network to recognize planets using 15,000 Kepler observations that had already been analyzed by astronomers. When all was said and done, the software learned to distinguish between real planets and false positives 96 percent of the time.
And in so doing, the technique found two new planets, one orbiting a star known as Kepler-80 and another orbiting Kepler-90, that star's eighth known world. It is the only solar system other than Earth's to feature at least eight planets.
All eight of Kepler-90's planets, however, circle closer to their sun than Earth does. Kepler-90i, the third planet from its star, is believed to be a rocky world with a surface temperature of around 800 degrees Fahrenheit, taking just 14.4 days to complete one orbit.
"The Kepler-90 star system is like a mini version of our solar system," Vanderburg said in a statement. "You have small planets inside and big planets outside, but everything is scrunched in much closer."
Shallue said he became interested in applying neural networks to astronomy "when I learned the Kepler mission had collected so much data that it was impossible for scientists to examine it all manually. Instead, scientists selected the strongest signals, which were the most likely to be actual, to receive the most attention.
"This process is like looking for needles in a haystack," he said. "Over 30,000 signals from the Kepler space telescope were examined, and about two-and-a-half thousand were confirmed to be actual planets."
To search for planets in the weaker signals "would be like looking for needles in a much, much larger haystack. There were simply too many weak signals to be examined using the existing methods."
The new technique may help astronomers find many more planets in the Kepler data, which is available online to researchers around the world. Shallue said the machine learning routines developed to study the Kepler data will be posted later for anyone to use.
"We built this model and trained this model using ... an open-source machine learning software library that was produced by Google," Shallue said. "We will actually be releasing all of the code (and) anybody can come in, and without any specialized hardware, and be able to train the exact same model we trained and potentially run their own searches of the Kepler data if they so desire."