Light curve imaging for exoplanet detection with deep learning: a conceptual trial

Abstract

The ever-growing number of space missions has made manual searching for exoplanet candidates infeasible due to the increasing volume of data. Consequently, the astrophysics community has extensively employed machine learning methods not only to handle the sheer amount of available data but also to enhance the sensitivity of detections concerning the signal noise inherent in relevant observational cases. This work builds upon and refines a previous review study, also presenting a conceptual trial for an alternative method to those previously discussed in the literature for classifying potential exoplanet signals using deep learning. We developed, trained, and evaluated Convolutional Neural Network (CNN) models to analyze light curves from the Kepler space mission, allowing inference on whether a given signal refers to an exoplanet or not. The distinction of this work lies in the imaging of these light curves before they are passed to the CNNs, which practically increases the number of dimensions available for analysis and enables the use of powerful and successful computer vision techniques for classification problems. Our best model ranks plausible planet signals higher than false-positive signals 97.22% of the time in our test dataset and demonstrates promising performance on entirely new data from other datasets. Our best model also shows a moderate capacity to generalize what it learned with data from other space missions, such as K2 and TESS. A good performance on entirely new data is a critical characteristic for upcoming space missions such as PLATO, and is work in progress at time of writing. Additionally, we provide new perspectives on how this imaging method can be further explored and tested in future works.