Machine learning algorithms for detection and classification of plasma structures in multiple-X-line collisionless reconnection regions*

A crucial component of magnetic reconnection research is the analysis of in-situ data from spacecraft that study naturally occurring reconnection regions such as Earth's magnetotail. However, spacecraft can only sample a single point in space for each timestep, and trace a 1D path through the plasma. This limitation makes detection and identification of dynamic plasma structures difficult. Correctly identifying structures in multiple-X-line reconnection regions is crucial for understanding the physics of the coupling of the microscale to the macroscale, such as the potential role that the plasmoid instability plays in reconnection dynamics and energy transfer. Previous work investigating this physics used simple hand-tuned algorithms for detection and classification (Bergstedt et al. 2020). This work develops a more nuanced and robust classification algorithm which utilizes a set of simulated 'spacecraft' trajectories through 2D PIC simulations of reconnection to train a machine learning model to identify regions of data corresponding to plasmoids and current sheets. A range of models from Random Forest Classifiers to Convolutional and Recurrent Neural Networks are implemented, and their efficacies are compared.