Automatic 3D Reconstruction of Coronal Mass Ejections Based on Dual-viewpoint Observations and Machine Learning

Abstract Coronal mass ejections (CMEs) are the major driver of severe space weather events, causing substantial economic losses to space-based and ground-based human assets. It is essential to advance our understanding of CME propagation dynamics for better predictions. We have developed an algorithm that automatically reconstructs CME structure, integrating dual-viewpoint observations with machine learning techniques. It consists of three stages: (1) region acquisition, (2) model construction, and (3) function optimization. First, we use two independent convolutional neural networks to identify and detect CMEs in the coronagraph images from two spacecraft. Next, we construct the projections of the graduated cylindrical shell (GCS) model in the fields of view of the coronagraphs. In the final step, optimal parameters are retrieved by minimizing the function that quantifies the morphological discrepancies between the image of the GCS model and the CME detection. Four CME events are reconstructed and analyzed to demonstrate the accuracy of our algorithm. A statistical analysis of 97 CME events from 2007 to 2018 is conducted to investigate both the two-dimensional (2D) and three-dimensional (3D) parameters. According to our statistics, the actual velocities are underestimated by 8%, and the widths are overestimated by 47% because of the projection effect from 2D observations. Furthermore, the widths and velocities are investigated to have a positive correlation coefficient of 0.67 (2D) and 0.52 (3D). The proposed method can further be used to provide CME initial parameters for magnetohydrodynamics simulations, enabling a deeper understanding of CME kinematics.