Identification of Magnetic Reconnection in Plasma Turbulence Using a Dual-branch Lightweight Hybrid Convolutional Neural Network

Abstract Magnetic reconnection is a fundamental plasma process that plays a crucial role in energy dissipation in plasma turbulence, in which magnetic energy is rapidly converted into particle energy. However, the identification of reconnection within turbulence remains a significant challenge. To automatically detect magnetic reconnection in plasma turbulence, we propose a novel and efficient classifier called a dual-branch lightweight hybrid convolutional neural network (DHCNN), which, for the first time, achieves automatic localization of reconnection sites by effectively leveraging negative-valued information in physical variables. We adopt a temporal partitioning scheme together with diverse data augmentation strategies to enhance the model’s generalization ability. Our results show that the DHCNN accurately identifies reconnection sites, achieving a Matthews correlation coefficient ( M ) of 0.67, a true skill statistic ( T ) of 0.67, a recall ( R ) of 0.83, a specificity ( S ) of 0.84, and an accuracy ( A ) of 0.84. Finally, by analyzing the impact of different physical variables, we find that the total current density (| J |), in-plane components of the magnetic field ( B x , B z ), and energy conversion term ( D e ) are the primary drivers of the model’s performance.