• machine learning
• magnetic reconnection
• plasma physics

Magnetic reconnection is a fundamental process in plasma physics being the only one able to rearrange the large-scale connections of magnetic field lines, allowing for important topological modifications of the field, despite it occurs in very small regions with respect to the system size. This change in the global topology allows the system to reach lower energy states otherwise forbidden and converts large amount of magnetic energy into kinetic energy, thermal energy and particle acceleration. Magnetic reconnection occurs in a large variety of space environments such as the solar corona, the Earth's magnetosphere, the turbulent solar wind.

Due to its importance and uniqueness, during the last decades magnetic reconnection has been extensively studied using theoretical models, numerical simulations and satellites' data. Still, some important questions remain to be elucidated. In particular, for what concerns simulations, detecting magnetic reconnection is a difficult task and finding reconnection signatures ask for a visual, not automatic, investigation. This is particularly true when simulations are not initialized with "ad-hoc" configurations, suitable for reconnection, but when current sheets were reconnection possibly develops are randomly generated by the turbulent dynamics, even in a simplified 2D geometry. The 3D case situation is even more complex since 3D reconnection dynamics is still a matter of debate even from a theoretical point of view.


The goal of this PhD thesis is to study magnetic reconnection in the context of space collisionless plasma, in particular in current sheets that are self-consistently generated by the plasma motion either by large-scale magnetohydrodynamic vortices (emerging after the development of the Kelvin-Helmoltz instability at the Earth's magnetospheric flanks) or by small-scale vortices in kinetic turbulence (as those developing in solar wind).

The main work of this Thesis addresses the possibility to use automatic techniques to individuate magnetic reconnection events in 2D Hybrid Kinetic simulations of plasma turbulence. These techniques are based on supervised (CNN) and unsupervised (KMeans and DBscan) machine learning methods. For what concerns the 3D case, the state-of-the-art of our work is presented. In particular, we statistically analyze magnetic reconnection events in a two-fluid 3D simulation of Kelvin-Helmholtz mediated magnetic reconnection at the Earth’s magnetospheric flanks. Finally, concerning 3D Hybrid Kinetic simulations of turbulence, we present a statistical analysis of current structures where potentially reconnection can occur, while the development of a machine learning technique to automatically individuate reconnection in 3D simulations is still ongoing.