Additive Attention for Vetting Transiting Exoplanet Candidates

Abstract In this study we propose a deep learning model architecture using long short-term memory (LSTM) networks and additive attention mechanisms for vetting transiting exoplanet candidates. Our method is applied to two different data sets (Kepler and TESS) comprising light curves, representing periodic fluctuations in star brightness that are potentially indicative of planetary transits. Making use of preprocessed data from the Kepler and TESS missions, we evaluate the effectiveness of LSTM-based approaches in contrast to conventional CNN methods. Through extensive experimentation, we demonstrate the efficacy of our final model, which integrates CNN, LSTM, and additive attention layers, coupled with feed-forward neural networks, achieving competitive performance with limited model complexity. Notably, the incorporation of additive attention enhances prediction accuracy across various model scales without significantly increasing model complexity. Our findings highlight the utility of additive attention in vetting transiting exoplanet candidates, offering understanding of the model decision-making processes. Overall, our proposed architecture presents a general-purpose approach for addressing transit detection challenges across different data sets, contributing to advancements in astronomical data analysis.