Paper detail

Deep solar ALMA neural network estimator for image refinement and estimates of small-scale dynamics

The contrasts and magnitude of observable signatures of small-scale features degrade as angular resolution decreases. High-cadence time-series of synthetic observable maps at 1.25 mm were produced from 3D magnetohydrodynamic Bifrost simulations of the solar atmosphere and degraded to the angular resolution corresponding to observational data with the Atacama Large Millimeter/sub-millimeter Array (ALMA). The Deep Solar ALMA Neural Network Estimator (Deep-SANNE) is an artificial neural network trained to improve the resolution and contrast of solar observations. This is done by recognizing dynamic patterns in both the spatial and temporal domains of small-scale features at an angular resolution corresponding to observational data and correlated them to highly resolved nondegraded data from the magnetohydrodynamic simulations. A second simulation, was used to validate the performance. Deep-SANNE provides maps of the estimated degradation of the brightness temperature, which can be used to filter for locations that most probably show a high accuracy and as correction factors in order to construct refined images that show higher contrast and more accurate brightness temperatures than at the observational resolution. Deep-SANNE reveals more small-scale features and estimates the excess temperature of brightening events with an average accuracy of 94.0% relative to the highly resolved data, compared to 43.7% at the observational resolution. By using the additional information of the temporal domain, Deep-SANNE can restore high contrasts better than a standard two-dimensional deconvolver technique. Deep-SANNE is applied on observational solar ALMA data. The Deep-SANNE refined images are useful for analysing small-scale and dynamic features. They can identify locations in the data with high accuracy for an in-depth analysis and allow a more meaningful interpretation of solar observations.