A New Methodology for Classifying Eclipsing Binaries with Kepler Data and Deep Learning

Abstract We present a new method for the automated classification of eclipsing binaries, into contact, detached, and semi-detached types using Kepler data. Phase-folded light curves are generated and chi-square vs. box size plots are constructed by comparing flux values to the median flux, revealing distinct class patterns. These patterns were first modelled using a polynomial damped sinusoidal function, whose period served as classification feature, achieving an overall accuracy of 86.5 percent. To capture more features and enhance accuracy, we trained a convolutional neural network, which improved the total accuracy to 90 percent, including 47 percent for the challenging semi-detached systems. However, several binaries displayed irregular chi-square signatures. To mitigate this, we incorporated simulated light curves generated with the PHOEBE modelling code, achieving 99 per cent accuracy in distinguishing contact and detached binaries. The resulting chi-square morphologies show a strong correlation with orbital period, and a subset of systems exhibit quarterly variability in their light curves and chi-square trends. We designate these as Temporally Varying systems. By measuring the normalized spread of the chi-square period across quarters, we define a statistical threshold that separates these systems from stable binaries. We reported four Temporally Varying systems not previously noted in the literature with magnetic activity that requires further investigation. Furthermore, cooler stars, namely late-F, G, K, and M types, display systematically higher variability than hotter stars. Cross-matching with catalogues of magnetically active stars indicates that stellar flares and starspots are the most likely causes of this enhanced variability.