A Neural Network based Search for Earth Analogs in Kepler Data

Abstract

Kepler was designed to find Earth-sized planets in Earth-like orbits, but its catalogue of Earth-analogue planet candidates is contaminated by instrumental artifacts. To try to solve this problem, we have developed a neural network architecture that allows to classify individual transits of possible Earth-like signals into three different categories (planet candidates, false positives, and noise) using pixel-level data. We use a branched convolutional neural network that receives as inputs the normalized time series of pixels surrounding the target star. The training data consists of short segments of pixel level data (created using a sliding window in time) for likely planets, false positives, and stars without known signals in Kepler data. The initial results show that the architecture has an 82% accuracy classifying individual transits; future work combining multiple transits could improve this accuracy significantly. The main novel contributions of this research are: the use of pixel level data as opposed to the aggregated cumulative flux for the detection of exoplanets, the proposal of a branched convolutional neural network architecture to address the problem, the filtering algorithms of Kepler data for the creation of the datasets, and the flow correction pre-processing algorithms. Ultimately, we plan to develop this algorithm and deploy it to search the full Kepler dataset for previously hidden Earth analogue exoplanets.