Convolutional Neural Networks and Stokes Response Functions

TL;DR

This paper investigates how convolutional neural networks learn to map Ca II spectral data to atmospheric temperature profiles, revealing that the CNN captures physically meaningful relationships between spectral features and atmospheric height.

Contribution

It demonstrates that CNNs trained on spectral data inherently learn to associate spectral regions with their corresponding atmospheric heights based on response functions.

Findings

CNN error inversely related to spectral line response functions

Spectral regions provide better temperature predictions at their formation heights

CNN captures physically meaningful wavelength-height mappings

Abstract

In this work, we study the information content learned by a convolutional neural network (CNN) when trained to carry out the inverse mapping between a database of synthetic Ca II intensity spectra and the vertical stratification of the temperature of the atmospheres used to generate such spectra. In particular, we evaluate the ability of the neural network to extract information about the sensitivity of the spectral line to temperature as a function of height. By training the CNN on sufficiently narrow wavelength intervals across the Ca II spectral profiles, we find that the error in the temperature prediction shows an inverse relationship to the response function of the spectral line to temperature, this is, different regions of the spectrum yield a better temperature prediction at their expected regions of formation. This work shows that the function that the CNN learns during the training process contains a physically-meaningful mapping between wavelength and atmospheric height.