07 Feb 2020 — Conor MacBride
Accurately constraining velocity information from spectral imaging observations using machine learning techniques
Conor D. MacBride1 and David B. Jess1
1Astrophysics Research Centre, School of Mathematics and Physics, Queen’s University Belfast, Belfast, UK
Determining accurate plasma velocities from spectroscopic measurements is a challenging endeavour, especially when considering weak chromospheric absorption lines are often rapidly evolving and contain multiple plasma profiles in their composition. Here, Mr MacBride presents a novel method that employs machine learning to identify the underlying components present within an observed spectral line, before constraining the constituent profiles through Gaussian and/or Voigt fits, alongside minimisation tests to validate the reliability of the results. With this method, automatic adjustments can be made to the models fitted such that active and quiescent components present in each particular spectrum can be identified accurately. Lastly, Mr MacBride utilises a Ca II 8542 Å spectral imaging dataset of a sunspot as a proof-of-concept study to show the potential of his team’s method for reliably extracting two-component atmospheric profiles that are commonly present in dynamic sunspot umbral chromospheres.
Having accurate velocity information for plasma in the solar atmosphere is important for studying the properties of waves present within. Using spectroscopic measurements of the Sun, velocities can be found by calculating the Doppler shift of a particular absorption line core.
Different spectral lines are formed across different atmospheric heights, therefore, by choosing a particular line, waves at a particular atmospheric height can be studied[1,2]. Some spectral lines, including Ca II 8542 Å, include an active emission component as well as the quiescent absorption component. Separate absorption and emission profiles must be fitted to each spectrum. The fitted profiles are then used to find the Doppler velocities of the quiescent/active atmospheric components.
Preliminary profile fitting methods struggled to accurately fit a profile across every region of a sunspot’s umbra. This was due to the spectra having a variety of different profiles due to the active component that was often present among the quiescent component.
Even if emission was not present in a particular spectrum, the fit would still “improve” if the algorithm fitted a significant non-zero emission profile as it could filter out some noise in the absorption profile.
Using machine learning, spectra can be accurately classified into discrete categories based on the ratio of their active component to their quiescent component. This allows the fitting method to be tailored to the physics that is present in each spectrum.
For a description of the method used please download the poster here.
 David Jess et al. 2019, Nat. Astron.
 Krishna Prasad et al. 2017, ApJ, 847, 5
 Mofreh Zaghloul 2007, Mon. Not. R. Astron. Soc., 375, 1-6