For the first time, astronomers have used a machine learning algorithm – a form of self-teaching AI – to confirm the existence of exoplanets in data collected by the now-retired Kepler space telescope. The algorithm, which was designed to differentiate real worlds from false positives in telescope data, confirmed a grand total of 50 exoplanets ranging from gas giants the size of Neptune to alien planets smaller than the Earth.
In 1995, a team of European astronomers made the monumental announcement that they had discovered the first world confirmed to orbit an alien star – a hot gas giant roughly half the size of Jupiter, called 51 Pegasi b. Ever since the discovery of that first fateful planet, humanity has been feverishly scouring the stars for evidence of other worlds that might be lurking throughout our galaxy.
Over time, planet hunting techniques evolved, and dedicated telescopes such as the Kepler and later TESS were launched to search vast swathes of sky for evidence of hidden worlds. This evidence came in the form of a periodic, telltale dip in the light of a distant star that occurs as a planet passes between its disk and the wide eye of a watching telescope. This type of exoplanet discovery is known as the transit method.
These efforts have not been in vain. Astronomers have discovered over 4,200 worlds orbiting beyond our solar system, and over 5,000 candidate exoplanets.
These candidate worlds require further observation to make sure that dips in light recorded by telescopes are not the result of other phenomenon, such as the presence of another co-orbiting star, interference from background objects, or minute errors in the observatory’s harvesting the data.
Astronomers are now turning to machine learning to help them sift through the veritable ocean of telescope data and weed the false positives out from the honest-to-science exoplanets.
Machine learning algorithms are able to, as the name would suggest, learn from past experiences in order to progressively improve their accuracy and performance over time.
Scientists from the University of Warwick Department of Physics and the Alan Turing Institute built an exoplanet-finding algorithm, and trained it by feeding it two large data samples captured by the now-defunct Kepler Space Telescope. One of the datasets was populated by already-confirmed planets, and another by known false positives.
The team then unleashed the algorithm on a sample of unconfirmed exoplanet candidates also from the Kepler archives. Any exoplanet that had less than a one percent probability of being a false positive was classified as confirmed.
"In terms of planet validation, no-one has used a machine learning technique before. Machine learning has been used for ranking planetary candidates but never in a probabilistic framework, which is what you need to truly validate a planet," comments Dr David Armstrong, of the University of Warwick, one of the authors of the new paper. "Rather than saying which candidates are more likely to be planets, we can now say what the precise statistical likelihood is."
The algorithm was able to statistically confirm the existence of 50 exoplanets from the candidate data, ranging from diminutive alien worlds smaller than the Earth all the way up to enormous gas giants the size of Neptune.
These worlds can now be targeted for follow up observations by more specialized and powerful telescopes that will be able to determine the key characteristics of the worlds.
According to the authors of the study, the machine leaning algorithm was completely automated, and was able to separate out the false positives from the real exoplanets faster than would otherwise be possible. The scientists believe that multiple exoplanet confirmation methods should be combined when confirming discoveries in the future, including their own automated algorithm.
"Almost 30 percent of the known planets to date have been validated using just one method, and that’s not ideal," Dr. Armstrong explains. "Developing new methods for validation is desirable for that reason alone. But machine learning also lets us do it very quickly and prioritize candidates much faster."
Moving forward, the researchers intend to continue training the algorithm, and are hoping to apply their algorithm to larger candidate exoplanet samples collected by TESS, and future missions such as ESA’s planned PLAnetary Transits and Oscillations of stars (PLATO) mission.
The paper has been published in the Monthly Notices of the Royal Astronomical Society.
Source: University of Warwick
