A Novel Convolutional Neural Network–Long Short-term Memory Model for Interplanetary Coronal Mass Ejection Detection

Abstract Current models for detecting interplanetary coronal mass ejections (ICMEs) using spacecraft in situ measurements rely heavily on manual processes, resulting in time-consuming procedures, inconsistencies in identification criteria, and gaps in existing catalogs. Although several machine learning–based detection models have been proposed, there is still a need for improved accuracy. This study presents a convolutional neural network–long short-term memory (CNN-LSTM) model with dynamic loss function, designed to efficiently process complex multidimensional spatiotemporal data. The model leverages LSTM to capture evolving patterns in time series data, while employing a CNN to extract spatial features, making the model highly adaptable. Using in situ data collected by the Wind spacecraft from 1997 October 1 to 2016 January 1, the proposed model demonstrates strong competitiveness. Trained on data from 1998 January 1 to 2009 December 31, 189 out of 230 ICMEs are detected in the test set, achieving an F1 score of 81.29%.