Scientists at the University of Warwick have devised a new machine learning technique that has proven the presence of fifty possible planets.
For the first time, astronomers have analyzed a sample of putative planets to identify which are genuine and which are “fakes,” or false positives, and to calculate the likelihood that each candidate is a real planet. This has been done using a machine learning-based artificial intelligence technique.
In addition to reporting their findings, they also conducted the first extensive comparison of planet validation methods in the Monthly Notices of the Royal Astronomical Society. In order to statistically corroborate upcoming exoplanet findings, they argue that a variety of validation techniques, such as their machine learning algorithm, should be used.
Numerous exoplanet surveys comb through massive amounts of telescope data looking for indications that a planet is transiting—that is, moving between the telescope and its star. This causes the telescope to detect a distinctive dip in the star’s brightness, although it might also be caused by a binary star system, interference from a background object, or even small camera faults. A planetary validation procedure can be used to sort through these erroneous positives.
A machine learning-based algorithm, developed by researchers from The Alan Turing Institute and the Departments of Physics and Computer Science at Warwick, can distinguish between real and false planets in large samples of thousands of candidates discovered by telescope missions like NASA’s Kepler and TESS.
Using two sizable datasets of verified planets and false positives from the now-retired Kepler spacecraft, it was taught to identify actual planets. After applying the technique to a dataset of Kepler’s unconfirmed planetary candidates, the researchers discovered fifty new confirmed planets—the first of which was verified by machine learning. Prior machine learning methods have ranked candidates, but they have never independently ascertained the likelihood that a candidate was a genuine planet—a necessary step in the planet confirmation process.
These fifty planets have orbits that range from as long as 200 days to as short as one day, and they range in size from worlds as big as Neptune to smaller than Earth. Now that these fifty planets have been verified as real, astronomers can use specialized telescopes to prioritize these for more observations.
The technique we have created allows us to take fifty candidates and upgrade them to genuine planets, says Dr. David Armstrong of the Department of Physics at the University of Warwick. Our goal is to utilize this method on sizable groups of potential candidates from present and upcoming missions such as PLATO and TESS.
“Nobody has ever employed a machine learning technique for planet validation. Although machine learning has been applied to rate potential planets, it has never been done so within a probabilistic framework—which is what is required to genuinely assess a planet. We can now determine the exact statistical likelihood rather than only listing the candidates that have a higher probability of being planets. A planet is deemed validated when the probability of a candidate being a false positive is less than 1%.
“Probabilistic approaches to statistical machine learning are especially suited for an exciting problem like this in astrophysics that requires incorporation of prior knowledge — from experts like Dr. Armstrong — and quantification of uncertainty in predictions,” stated Dr. Theo Damoulas, Deputy Director, Data Centric Engineering, and Turing Fellow at The Alan Turing Institute. An excellent illustration of how the added processing complexity of probabilistic approaches pays off handsomely.
The system is faster than current methods once it is constructed and taught, and it can be fully automated. This makes it perfect for examining the thousands of planetary candidates that may be found in current surveys like TESS. The scientists contend that it ought to be among the instruments used in concert to validate planets in the future.
“It’s not ideal that nearly 30% of the planets that have been discovered to date have been validated using a single method,” continues Dr. Armstrong. For that reason alone, it is desirable to develop new ways of validation. However, machine learning also makes it possible for us to work quickly and rank prospects much more quickly.
“The algorithm still needs to be trained, but after that, it will be lot simpler to apply to new candidates. Additionally, you can add fresh insights to gradually enhance it.
It is expected that a survey such as TESS will yield tens of thousands of planetary candidates, and it would be excellent to be able to examine them all uniformly. We can accomplish it more effectively with the help of quick, automated methods like this one, which can validate planets in fewer steps.
