Does this star have a planet? A new algorithm could help astronomers predict, on the basis of a star’s chemical fingerprint, whether that star will host a giant gaseous exoplanet.
“It’s like Netflix,” Natalie Hinkel, a planetary astrophysicist at the Southwest Research Institute in San Antonio, Texas, told Eos. Netflix “sees that you like goofy comedy, science fiction, and kung fu movies—a variety of different patterns” to predict whether you’ll like a new movie.
Likewise, her team’s machine learning algorithm “will learn which elements are influential in deciding whether or not a star has a planet.” Hinkel is lead author on a paper, accepted for publication in the Astrophysical Journal, that presents this new technique.
Picking Out Patterns
Past studies show that iron-rich stars are more likely to host a giant exoplanet. This new algorithm draws on an extensive database of stellar compositions, the Hypatia Catalog, to test whether that trend holds for groups of elements together.
The team looked at combinations of some of the most common planetary ingredients: light and gaseous volatiles, oxygen-loving lithophiles, iron-loving siderophiles, and iron. The algorithm randomly selects a subset of Hypatia stars known to have giant planets, identifies patterns in their chemical compositions, and decides how important the patterns are when hosting a planet.
By training the program on about 300 planet-hosting stars, the team found that a few elements were consistently good planet predictors. In addition to iron, “the elements that ended up being the biggest indicators were carbon, oxygen, and sodium,” Hinkel said. That makes sense, she said: Gas giant planets need gaseous elements, and planetary cores need iron.
Nickel, to the researchers’ surprise, turned out to be not very important despite forming in a way similar to iron. Sodium, however, was much more important than expected, something they are still trying to explain.
New Stars to Search
The team also applied the predictive patterns to Hypatia stars not yet known to host a planet. By repeating the prediction hundreds of thousands of times for each star, the algorithm calculated the likelihood that a star is a planet host. The team confirmed the algorithm’s accuracy by “hiding” planet hosts among the test stars. The program picked them out about 75% of the time.
All told, the team tested more than 4,200 stars not currently known to have a planet. About 350 of those stars had a more than 90% probability of hosting a giant exoplanet.
“I think that it’d be super useful to be able to know where to look ahead of time,” Hinkel said.
An Important Tool for the Future
“This work is a wonderful example of the power of large data sets such as the Hypatia Catalog, especially when combined with machine learning approaches,” said space scientist Shawn Domagal-Goldman at NASA Goddard Space Flight Center in Greenbelt, Md., who was not involved with this research. A similar approach might also help select stars around which to search for signs of life, he said.
Increasingly, “we are swamped with more data than is feasible to make sense of,” said Steven Desch, an astrophysicist at Arizona State University in Tempe who also did not participate in this study. Techniques like this “draw our attention to patterns we might have missed that are probably clues to how planets are formed.”
“As exoplanet data continues to build, algorithms like this one are going to be increasingly important tools,” Desch said.
—Kimberly M. S. Cartier (@AstroKimCartier), Staff Writer