A PhD Thesis in collaboration between UNIPI and Aix-Marseille University exploiting the competencies and research interests addressed by the AIDA project.
Magnetic reconnection is a fundamental process in plasma physics. It is responsible for changes in the magnetic field topology involving the conversion of large amounts of magnetic energy into heating and particle acceleration. It occurs almost everywhere in space environment as, for instance, the solar corona, the turbulent solar wind or at the boundary of the Earth's magnetosphere. Despite decades of theoretical and experimental studies of this process, numerical simulations and analysis of direct satellites' data represent today a major research line in space physics because of the key role of reconnection in space plasmas and because important issues still remain unsolved. In particular, there is no verified technique to automatically detect reconnection events in 2D and 3D simulations data or satellites' 1D time-series. Since very huge data are routinely produced to study this topic, an automatic algorithm able to speed-up the analysis is highly demanded. This problem has been topic of study in the PhD Thesis of M. Sisti (to be defended in September 2021), an expert in the field of magnetic reconnection. The issue has been approached by M. Sisti through artificial intelligence tecniques in collaboration with members of the European project AIDA. Two different approaches to automatically detect reconnection in 2D-3V Vlasov simulations have been developed: supervised and unsupervised technique. In collaboration with Andong Hu and Jannis Teunissen, members of the AIDA project from Centrum Wiskunde & Informatica of Amsterdam, they created the first labeled dataset of current sheets from hybrid kinetic simulations of plasma turbulence to train a CNN in identifing magnetic reconnection. The trained CNN turned out to be high performance in individuating reconnection sites. Moreover, the model trained turns out to be even able to correctly catch reconnection events in cases in which the human labeling was wrong. At the same time, M. Sisti in collaboration with Giorgio Pedrazzi from CINECA, Bologna, has developed an unsupervised machine learning algorithm exploiting both KMeans and DBscan clustering algorithms and based on a threshold on the aspect ratio of the current sheets, which turned out to be very accurate in identifying reconnecting current sheets, even if the precision decreases a bit during fully developed turbulence regimes (as expected). This work, published into two distinct papers on the Astrophysical Journal, has demonstrated the possibility of using machine learning techniques to analyze big data from numerical simulations and open the path for future applications to 1D time series from satellite data and to fully 3D numerical simulations.
Workflow used to create the labeled dataset of image of current sheets from 2D HVM simulations of turbulence used for train the CNN in Hu et al. (2020). A group of experts was asked to label over 3000 images identifying magnetic reconnection events.