UT Astronomer Discovers New Planet in Our Rival Solar System

Kepler-90i might not have piqued anyone’s attention had it been discovered under different circumstances. The rocky, super-earth-sized exoplanet is about 2,545 light-years away from Earth. It circles Kepler-90, its sun, at a rapid clip, making a full orbit every 14.5 days. With a balmy average surface temperature of 817 degrees Fahrenheit, Kepler-90i doesn’t hold much promise for life.

Yet this inhospitable planet with a humdrum name and faraway star was a consequential find. It was the eighth planet found orbiting Kepler-90 in the constellation Draco, meaning that our solar system is no longer the only one with eight planets that we are aware of. More importantly, Kepler-90i wasn’t discovered by humans. It was discovered by an artificial intelligence (AI) algorithm.

Andrew Vanderburg, a UT astronomer, and Christopher Shallue, a software engineer at Google, facilitated the discovery. The two worked with data from the Kepler Space Telescope, which has taken pictures of 200,000 stars every half hour for more than four years.

Read: Life Under the Stars at UT’s McDonald Observatory

With the Kepler dataset, researchers can find planets by looking at light changes caused by the dimming of any given star. This occurs when a planet transits—or passes in front of—the star. Kepler’s dataset contains about two quadrillion (10 to the 15th power) possible orbits of planets. To sift through this vast trove of potential planet signals, astronomers have relied on automated tests and human eyes. Yet weak signals that could full well be planets are often 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,” said Paul Hertz, director of NASA’s Astrophysics Division in Washington. “This finding shows that our data will be a treasure trove available to innovative researchers for years to come.”

It all started when Shallue cold-emailed Vanderburg. He sensed that astronomy was a field being overrun by data to the point that humans didn’t know what to do with it and he wanted to help. This exchange was the beginning of a long and fruitful partnership, in which the two pooled complementary expertise.

Using open-source software from Google called TensorFlow, they created a neural network, a type of A.I. system, and taught it what planets look like in the Kepler dataset. The model trained on a database of 15,000 previously vetted signals from the Kepler exoplanet catalog.

When searching for exoplanets, astronomers struggle with false positives: signals that look like a planet but are in fact instrumental glitches or something non-planetary transiting a star. But Vanderburg and Shallue were able to train the neural network to decipher whether a signal was a planet or a false positive. The model correctly identified true planets and false positives 96 percent of the time.

Asked about the age-old concern of robots taking humans’ jobs, Vanderburg shrugs. He isn’t worried about astronomers. A.I. won’t replace their jobs, it will just make them more efficient. The process of deciding if a signal is a planet or not a planet is redundant and time-demanding, Vanderburg said. “I’ve probably gone through 35,000 signals by eye over the past few years. If I could replace the time I spent doing that myself with a computer that does it just as well or better … then that’s a benefit for me and all astronomers.”  

Read: How One UT Alumnus Helped Continue Some of Einstein’s Most Ambitious Work

One downside of neural networks and artificial intelligence is that they behave like black boxes. It’s difficult to decipher how and why they make decisions. “That makes it hard to recreate what’s going on, especially when something goes wrong,” Vanderburg said. “The system is able to tweak itself until it can almost perfectly tell if a signal is a planet or not a planet. It does this by changing lots of different numbers and weights—thousands and thousands of them.”

Given the model’s accuracy and its far-reaching capabilities, though, Vanderburg is excited. “We previously didn’t have the capacity to go through all [the signals] and identify the ones that looked real,” he said. Astronomers will now be able to look at weaker, more speculative signals. And inevitably, they will discover new planets.

This may bring us closer to answering the oft-pondered question: are we alone? Since there are ten billion Earth-like planets in our galaxy, and a hundred billion galaxies in the universe, a new era of A.I.-enabled astronomy might pay big dividends in the search for extraterrestrial life or planets that can support it. When new planets are discovered, astronomers can search them for biosignatures, which NASA’s James Webb Space Telescope will be equipped to do. The Webb telescope, to be launched on an Ariane 5 rocket from French Guiana in spring 2019, will serve thousands of astronomers.

“There are missions being planned right now—LUVOIR, X-Hab—to find life on other planets,” Vanderburg said. “When we use this technique to try to measure more precisely the number of earth-like planets and earth-like orbits around sun-like stars, that’s going to have a direct impact on efforts to detect life outside of our solar system.”

A.I.-driven discoveries will allow astronomers and space agencies to decide how to allocate resources, where to send space-faring vessels and powerful space telescopes, and what missions to plan for the future.

When Vanderburg earned his PhD in astronomy and astrophysics from Harvard last year, A.I.-turbocharged planet-hunting wasn’t in the cards. But this project changed everything. He’s excited by the potential now, and doesn’t see how he wouldn’t combine the two down the road.

“It’s a really up-and-coming field,” he said. “I’m going to keep going for it.”

Photo via Nasa/Ames Research Center/ Wendy Stenzel


No comments

Be the first one to leave a comment.

Post a Comment