Inferring Plasma Flows in the Solar Photosphere & Chromosphere using Deep Learning and Surface Observations

Direct measurements of plasma motions are limited to the line-of-sight component at the Sun's surface. Multiple tracking and inversion methods were developed to infer the transverse motions from observational data. Recently, the fully convolutional DeepVel & DeepVelU neural networks were trained in conjunction with detailed magnetohydrodynamics (MHD) simulations of the Quiet Sun and sunspots to recover the instantaneous depth/height-dependent transverse velocity vector from a combination of intensitygrams, magnetograms and/or Dopplergrams of the solar surface. Through this supervised learning approach, the neural network attempts to emulate the synthetic flows, and by extension the physics, from the numerical simulation it was presented during its training, i.e. its outputs are model-dependent and may be subjected to biases. Although simulations have become increasingly realistic, the validity of flows inferred by DeepVel or DeepVelU is subject to debate when using real observational data as input. As a test, we use white light images of the Quiet Sun photosphere (optical depth tau=1) produced by the Interferometric BIdimensional Spectropolarimeter (IBIS) installed at the Dunn Solar Telescope to infer plasma motions approx. 150-200 km above the surface (i.e., near the transition between the photosphere and the chromosphere) using DeepVel. We discuss work in progress comparing the neural network estimates to the optical flows determined from a time series of observational data formed near 150-200 km above the surface. Optical flows do not directly track actual transverse plasma motions, but are correlated with physical flows over certain spatial and temporal scales.