In this research, we present two 1D Convolutional Neural Network (CNN) models that were trained, validated and tested using simulated light curves designed to mimic those expected from the Kepler Space Telescope during its extended mission (K2). We also tested them on real K2 data. Our light curve simulator considers different stellar variability phenomena, such as rotations, pulsations and flares, which along with the stellar noise expected for K2 data, hinders the transit signal detection, as …
Read moreIn this research, we present two 1D Convolutional Neural Network (CNN) models that were trained, validated and tested using simulated light curves designed to mimic those expected from the Kepler Space Telescope during its extended mission (K2). We also tested them on real K2 data. Our light curve simulator considers different stellar variability phenomena, such as rotations, pulsations and flares, which along with the stellar noise expected for K2 data, hinders the transit signal detection, as in real data. The first model effectively identifies transit-like signals in light curves, classifying them based on the presence or absence of such signals. Furthermore, the second model not only phase-folds the light curves but also eliminates stellar noise, a crucial step when fitting transits to the Mandel and Agol theoretical transit shape. The obtained results include an accuracy of $\sim 99\%$ when classifying the light curves based on the presence or absence of transit-like signals, and $MAPE\sim 6\%$ regarding to the transits’ depth and duration when phase folding the light curves, showing the great capabilities of 1D-CNN for automatizing the transit search in light curves, both on simulated and real data.
