Neural Network Based Flow Recovery and Magnetic Structure Identification in the Solar Atmosphere

Energy transport is a fundamental aspect of solar atmospheric dynamics. Plasma motions
can generate concentrated magnetic flux, driving large-scale transients such as solar flares
and coronal mass ejections, which pose an increasing threat to space- and ground-based
infrastructure. However, direct measurement of plasma flows on the Sun remains unfeasible, raising a critical question: how can plasma flows be accurately recovered from observations?
In a step toward answering this question this thesis employs the neural network Deep-
Vel for estimating flows in new scenarios including active regions and the chromosphere, thus enabling the estimation of flows by assimilating data from state-of-the-art simulations. DeepVel was further extended to higher regions of the solar atmosphere, where a reduced optical thickness results in less coherent apparent motions. By training on a simulation containing
swirling motions in the chromosphere, the networks ability to identify coherent swirls against an incoherent background was tested using a vortex detection method. Network performance was evaluated by undergoing cross-validation with simulations, performing error analysis and comparing with the widely-used Fourier-based local correlation tracking technique. Its ability to imitate the physics present was also tested for the first time by identifying through
Lagrangian coherent structures in the recovered flows.
Results highlight that DeepVel is highly capable for identifying coherent flow structures,
that determine the evolving flow dynamics, across simulations. These flow structures were
shown to correspond well to the presence of emerging active regions thus presenting a promising performance for use with real-world observations for forecasting and tracking the presence of active regions. Despite currently available data, both simulated and observed, tests indicate success in applying DeepVel to chromospheric plasmas.