At a press conference, the annoucement came of the first exoplanets discovered thanks to the TensorFlow machine learning engine created by Google. Researchers Christopher Shallue and Andrew Vanderburg trained this system to make it recognize exoplanets in the data collected by NASA’s Kepler space telescope. The two exoplanets announced are Kepler-90i and Kepler-80g but it’s only the beginning for a new way to look for exoplanets, especially the smaller ones that leave very weak traces.
Google released its TensorFlow machine learning engine as a free / open source library a little over 2 years ago, allowing anyone to freely use it adapting it to very different taks. Christopher Shallue is a Google engineer who works for the division of the artificial intelligence company and together with NASA Sagan Postdoctoral Fellow Andrew Vanderburg adapted TensorFlow to search for exoplanets. An article that describes this research was accepted for publication in “The Astronomical Journal”.
The Kepler space telescope allows the discovery of exoplanets using the transit method. Basically, when an exoplanet passes in front of its parent star it causes a tiny eclipse that can be detected with very sophisticated instruments such as Kepler’s, which can measure the very small changes in a star’s brightness due to a transit. The confirmation of the existence of an exoplanet and even more of various exoplanets in a star system requires an analysis of the collected data.
The training of the new system was conducted using a set of data collected from the observations of NASA’s Kepler space telescope. 15,000 signals previously vetted from the huge Kepler’s catalog and the TensorFlow-based neural network successfully identified exoplanets but also false positives in 96% of cases.
At that point, the researchers tried to provide the system with weaker data that belong to 670 star systems in which a number of exoplanets had already been discovered. They believe that this type of system is an ideal candidate to find more of them.
The largest exoplanets are the easiest to find while the analysis of the weakest traces, the ones from small exoplanets, can be really difficult. Andrew Vanderburg compared it to the search for jewels in rocks with a sieve: having a finer sieve you find more rocks but you can also find more jewels. In the use of a neural network applied to weaker signals, this means that the researchers have found more false positives but also new exoplanet candidates.
Among the two confirmed exoplanets, Kepler-90i was given great importance because it’s part of a star system a little over 2,500 light years from Earth, where 8 planets have now been confirmed. It’s the first system with a number of planets equal to those of the solar system and even in their positions there are similarities because the rocky planets are the closest to their star while the larger planets are outside.
The big difference between the two star systems is in size because the star Kepler-90 has a much smaller system. The orbit of its outermost planet, Kepler-90h, is similar to that of the Earth while the one just discovered, Kepler-90i, is very close to its star so its year lasts about 14.4 Earth’s days. Their star is a little smaller than the Sun so Kepler-90i is a little bit bigger than the Earth but the conditions on it are probably similar to those on Mercury.
The other exoplanet discovered thanks to this research is Kepler-80g, the sixth of its star system. Its size is similar to the Earth’s and four of its neighbors are in a situation that sees an orbital resonance like in the TRAPPIST-1 system with the consequence that they’re very stable.
This is just the beginning for this type of research. Other exoplanets may be discovered that are small or are further from their parent star so they don’t transit all the time in front of it. This type of analysis can also be adapted to the data collected by other planet hunters, allowing a leap forward in the search for them.