An Iterative, Deep Learning Approach for Switchback Classification in Parker Solar Probe Data

Abstract A neural network model known as ParkerNet has been implemented for classifying switchbacks in Parker Solar Probe (PSP) data. ParkerNet is a binary classification neural network model that combines convolutional neural network layers with bidirectional long short-term memory layers. We employ a targeted, human-in-the-loop approach, where a small set of labels is initially provided to the network for training, and select predictions are iteratively corrected and fed back for retraining. The predictions from the network are compared to two switchback catalogs by J. Huang et al. and F. Pecora et al. ParkerNet only needed approximately 12% of data labeled as switchbacks to demonstrate strong performance, showing high agreement with the core/spike region in the J. Huang et al. catalog. The application of ParkerNet to PSP data highlights the potential of a data-forward approach to unify the identification and characterization of switchbacks and provides a framework for future switchback detection.