To much fanfare, NASA just announced the discovery of a new rocky planet, Kepler-90i. Orbiting a star some 2,545 light years away, the new planet is roughly 1.3 times the size of Earth and blistering hot—around 800 degrees Fahrenheit. The find is unique for a few reasons: For one, Kepler-90i is the eighth planet in the previously discovered Kepler-90 solar system, making this far-off system the only family of exoplanets with as many planets as our own.
But perhaps more impressive is how researchers found the planet: by using a Google-developed artificial neural network. The basic idea behind neural networks is that, instead of programming specific rules into a computer, researchers feed that computer a large set of data and the system develops its own way to accomplish the specific task. Layers of computer "neurons" each do simple computations, passing the output on to another layer, says Chris Shallue, a Google software engineer specializing in neural networks, in a press conference.
Such systems have previously been used to translate between languages, identify breast cancer tumors, or even identify hotdogs vs. not hotdogs. "Our idea was to turn this technique to the skies and teach a machine learning system how to identify planets around far away stars," Shallue says. His team’s results have been accepted for publication in the The Astronomical Journal.
To use this kind of system for exoplanet hunting, researchers turned to the vast database of planetary candidates that the space telescope Kepler has amassed since its launch in 2009. Kepler has monitored the brightness of roughly 200,000 stars, watching for faint dips in the light—the telltale sign of an orbiting planet. Researchers or citizen scientists on the ground then sort through this data by hand (sometimes with the help of statistical techniques) to identify the most likely planetary candidates.
It's a grueling and time-intensive process: From the roughly 35,000 signals of possible exoplanets orbiting stars outside our solar system, researchers have so far confirmed 2,525 exoplanets."This process is like looking for needles in a haystack," says Shallue.
To help narrow down possible candidates scientists mostly focus on the biggest dips in light from the stars, which signify the most probable planetary candidates. So Shallue and the team at NASA turned a neural network on the fainter dips, hunting for planets in the data.
The researchers first fed the system around 15,000 Kepler stars with already labeled orbiting planets. Then it gave the neural network data from 670 stars that were candidates in the search for multi-planet systems. The analysis turned up two new exoplanets: Kepler-90i and Kepler-80g, which is the 6th planet in its system. Statistically, Vanderburg says, there is only a one in 10,000 probability that these are false positives.
The neural network was able to find these needles in the haystack far more quickly and efficiently than humans would have, says Shallue. "This is a really powerful technique," emphasizes Courtney Dressing, an astronomer at University of California at Berkeley who was not involved in the study. "And they could take this and apply it to more stars and perhaps find more planets as well."
The new technique also raises the possibility that systems with eight—or even more—planets aren't so odd after all. "In our solar system we know that we have eight planets because we're in the system, we can look at all the planets," says Dressing. But if you took a system exactly like ours and placed it 30 light years away, what would we see from Earth? We'd probably we'd see our massive gas giant Jupiter and possibly Earth, says Dressing. "But we probably wouldn't know about any of the other planets," she says.
Even Kepler-90 may be harboring more orbiting bodies, says Andrew Vanderburg, a postdoctoral fellow at the University of Texas, Austin who is an author on the new study. "It would almost be surprising to me if there weren't any more planets around this star," he says, noting that the planet has a large area surrounding it that researchers have yet to examine.
The researchers hope to further hone their AI system to improve its ability to identify false positives and tease through the complexities of the Kepler data. They also plan to take advantage of more of the supplementary information that Kepler collects about the stars, Dressing explains.
The idea of this planetary abundance is exciting for a couple of reasons, Dressing says. For one, it means that there's just more places in our galaxy where life could have evolved. But it could also "change our picture of how planets form in the first place," she adds.
The new study could also be the start of a timely collaboration. The Transiting Exoplanet Survey Satellite is set to launch in March of 2018 to peer at the nearby stars on the hunt for orbiting bodies. "That data set will be so big, that if we're able to use the sophisticated computational tools and neural networks to classify planets, we'll be much further along in characterizing planets than we would be if we relied on human eyes [alone],” says Dressing. To her, this newest finding "highlights the advantage of bringing together people from different skillsets to look at a new problem.”