Classification of astronomical objects from PLATO observations using Deep Learning

Since the first discovery of a transiting exoplanet, there has been a large interest in the scientific community for exoplanetary detection. This has produced a surge in ground-based observational campaigns and the launch of specialized space instruments. In the near future, ESA will launch PLATO, a mission designed to identify Earth-sized exoplanets across wide fields. All these efforts produce vast data sets, demanding advanced automation to aid analysis. Algorithms like BLS and TLS have been developed, but they still require human oversight. Machine learning and deep learning techniques offer a distinct approach. Astronet, a convolutional neural network, has shown promise in exoplanet identification through deep learning, assessing signals as potential planets or false positives in Kepler data and identifying previously undetected exoplanets. This study adapts and tests Astronet on PLATO simulated stellar light curves to classify planet candidates, eclipsing binaries, and non-transiting planets. Achieving 98.33% accuracy, the model excels at distinguishing actual planet candidates from noise and false positives. This underscores the potential that would have a custom designed algorithm for PLATO in autonomous planet detection, facilitating large-scale data analysis and focusing scientists’ attention on potential discoveries.