A neural network called "ExoMiner" has identified a total of 301 new exoplanets. However, the discoveries are not based on new observations, but on archive data from the Kepler space telescope. Based on their light curves, the stars studied were considered candidates for planetary systems in which a planet passes in front of the central star and darkens it. As the development team reports in the "Astrophysical Journal", Exominers studied the light curves of the stars and was able to reliably distinguish the darkenings caused by these so-called transits from brightness fluctuations for other reasons.
A new network is a machine learning system with a special architecture from several layers that tries to reproduce nervous systems. "If Exominer says something is a planet, then you can be sure that it is a planet," says Hamed Valizadegan, project manager of Exominer at the AMES Research Center of NASA. However, the planets identified in this way are only considered validated. In order to confirm them safely, they must be detected with a different method regardless of this than just the light curve. One possibility, for example, would be to measure the slight "tumbling" of her mother star, which comes about due to the attraction of the companions.
The specialists around Valizadegan were able to independently verify the conclusions of the algorithm, because, according to the team, the latter decides transparently. "Unlike other machine learning programs that search for exoplanets, ExoMiner is not a black box – why it decides for or against a planet is not a mystery," says Jon Jenkins, also from the Ames Research Center. The neural network works with already raw processed data of possible transits. In doing so, one checks whether an episode of lower brightness fits the shape that a transient exoplanet would imprint on the light curve.
Exomine analyzes the mathematical properties of light curve and theoretical models and then judges the probability of an exoplanet. The machine learning system was previously trained by the team around Valizadegan with data known exoplanets. As the group reports, other algorithms are struggling. The best previous system of confirmed exoplanet has identified about three quarters correctly in a test set, but exominer, on the other hand, under the same conditions 93 percent, writes the working group in the publication. The 301 newly validated planets increase the total number of exoplanets discovered to 4880. The experts are confident that Exominer will also significantly simplify the evaluation of the data of future missions such as Tess or Plato.
