Scientists discover two exoplanets through machine learning and Kepler data
A team from Google and scientists from several American universities have discovered two new exoplanets. This was done on the basis of data from the Kepler space telescope’s mission, which ended last year, to which machine learning was applied.
The scientists describe in a paper that they trained a neural network called AstroNet-K2 before the discovery. According to co-researcher Anne Dattilo, this discovery is significant because it is the first time a neural network has been applied to the K2 data. K2 refers to the extended mission of the Kepler space telescope.
After one of the space telescope’s four gyroscopic flywheels failed in 2012, a second flywheel gave up the ghost in 2013. As a result, the telescope could no longer be kept stable and could no longer aim precisely at the stars to be studied. In the end, a technical solution gave the telescope a second life, although those observations were less precise due to the slight instability and the data was accompanied by the necessary noise. That makes it relatively difficult to research this K2 data, says Dattilo.
The modified algorithm was trained on 27,634 stars of K2. Dattilo told The Register that it only took 40 minutes to train the algorithm on a laptop, but it took months for the research team to figure out how it could work based on the K2 data. AstroNet-K2 is quite successful, according to the researchers, but the researchers write that the model is not yet ready to fully automatically detect and identify planetary candidates. When applied to a number of stars, too many false positives come up, incorrectly indicating that a signal may be a planet candidate. The number of false reports still exceeds the number of detected real planets. In combination with human input, this is a big step forward, because it significantly reduces the data set to be viewed.
The neural network is trained for short dips in a star’s brightness, which can be caused by a passing planet. This means that the algorithm is only trained to detect exoplanets that are in a certain vicinity of their star. Exoplanets that are far from their star, or exoplanets with anomalous shapes are not recognized. This means that discovering these types of anomalous planets still requires a human eye.
AstroNet-K2 builds on research by scientists from Harvard and Google, who announced in December last year that they had discovered an exoplanet based on Kepler data using Google machine learning. This open source model later appeared on Github and the same should happen with AstroNet-K2.
The discovered exoplanets are called K2-293b and K2-294b. They are in fact identical planets located 1230 and 1300 light-years from Earth in the constellation Aquarius. Both objects are slightly larger than Earth, are very close to their stars and have short orbits. Because they are close to their stars, it is quite hot. The discoveries made through the algorithm have also been confirmed by two terrestrial telescopes.
With AstroNet-K2, even more exoplanets will be discovered, as there is still plenty of K2 data left. Moreover, Dattilo thinks that the algorithm can also be used with some changes for the data from the TESS space telescope, the ‘successor’ to Kepler that became operational last year.