Machine Learning for Local Detection of Separators in Three‐Dimensional Magnetic Fields

Abstract Magnetic reconnection is a major plasma phenomenon occurring in various key environments ranging from the Sun and near‐Earth space to astrophysical plasmas. While magnetic reconnection is relatively well‐understood under two‐dimensional (2D) settings, it remains challenging to characterize in three‐dimensional (3D) magnetic fields. In the 3D setting, such as in the Earth's magnetotail, magnetic reconnection tends to occur along separators, which are special field lines derived from the global topology of the magnetic field. However, current global approaches to locating separators can be computationally inefficient or provide inconsistent results on more complex magnetic fields. On the other hand, while local approaches do exist, they often liken the 3D magnetic field to a 2.5D field, where one of the three field components is assumed to be negligible, which is not always accurate. We employ two different types of machine learning approaches for local identification of the separators of 3D magnetic fields without the explicit assumption of a 2.5D setting. We compare the performance of these methods to those of a global approach and a local 2.5D approach. As a result, we show that, when trained on the separator output of an existing global approach, the machine learning methods can locate separators from large‐scale simulated magnetotail data. In addition, compared to the global approach, the machine learning methods provide a notable speed‐up in computation time and increased consistency of outputs over adjacent simulation time steps. Finally, we provide analysis on where each of the considered separator identification methods tends to succeed or fail.