Mimicking spectropolarimetric inversions using convolutional neural networks

¹ Department of Physics, University of Colorado, Boulder, CO 80309, USA
e-mail: ivan.milic@colorado.edu
² Laboratory for Atmospheric and Space Physics, University of Colorado, Boulder, CO 80303, USA
³ National Solar Observatory, Boulder, CO 80303, USA
⁴ Instituto Astrofisica de Andalucia (CSIC), Apartado de Correos 3004, 18080 Granada, Spain
e-mail: gafeira@iaa.es
⁵ Centre for Earth and Space Research of University of Coimbra, Department of Physics, University of Coimbra, 3000-456 Coimbra, Portugal

Received: 20 August 2019
Accepted: 2 March 2020

Abstract

Context. Interpreting spectropolarimetric observations of the solar atmosphere takes much longer than the acquiring the data. The most important reason for this is that the model fitting, or “inversion”, used to infer physical quantities from the observations is extremely slow, because the underlying models are numerically demanding.

Aims. We aim to improve the speed of the inference by using a neural network that relates input polarized spectra to the output physical parameters.

Methods. We first select a subset of the data to be interpreted and infer physical quantities from corresponding spectra using a standard minimization-based inversion code. Taking these results as reliable and representative of the whole data set, we train a convolutional neural network to connect the input polarized spectra to the output physical parameters (nodes, in context of spectropolarimetric inversion). We then apply the neural network to the various other data, previously unseen to the network. As a check, we apply the referent inversion code to the unseen data and compare the fit quality and the maps of the inferred parameters between the two inversions.

Results. The physical parameters inferred by the neural network show excellent agreement with the results from the inversion, and are obtained in a factor of 10⁵ less time. Additionally, substituting the results of the neural network back in the forward model, shows excellent agreement between inferred and original spectra.

Conclusions. The method we present here is very simple for implementation and extremely fast. It only requires a training data set, which can be obtained by inverting a representative subset of the observed data. Applying these (and similar) machine learning techniques will yield orders of magnitude acceleration in the routine interpretation of spectropolarimetric data.