Recovering coherent flow structures in active regions using machine learning

Analysing high-resolution solar atmospheric observations requires robust techniques to recover plasma flow features across different scales, especially in active regions. Current methodologies often fall short in capturing subgranular-scale flows, and there is limited research on the errors introduced by velocity estimation techniques and analysing the properties of recovered flows in the presence of kG magnetic flux density. This study concentrates on validating the effectiveness of the DeepVel neural network in recovering subgranular to mesogranular-scale topological plasma flow features throughout the total evolution of a simulated active region by tracking tracers, and reproducing coherent patterns. The neural network was trained on the r2d2 radiative MHD simulation depicting the emergence and decay of a magnetic flux tube. DeepVel achieved strong correlations (exceeding 0.7) with flows from an unseen muram simulation, despite being trained on a model with a simpler radiative transfer and lacking thermal resistivity. DeepVel was able to capture the detailed topology well, e.g. the structure of vortical and diverging structures across all scales present in the flows. DeepVel performed slightly less well in the umbra, this is likely explained by magnetic field suppression and reduced contrast. Differences in velocities introduced by DeepVel did not affect Lagrangian analysis; consequently, we demonstrate for the first time that the DeepVel-recovered velocities accurately reflected the flow’s transport barriers. These findings highlight the precision and reliability of the DeepVel and its ability to emulate plasma flows surrounding and within active regions.