Too Many Gaussians, not Enough Grad Students
Extragalactic surveys of galaxies have often used single spectrums on the centre of galaxies, e.g. SDSS, GAMMA. Now, surveys are using Integral Field Spectroscopy (IFS) where hundreds of spectra are observed for a single galaxy. This means instead of a few thousand spectrums (one for each galaxy) is now millions of spectra that need to be reduced and analysed.
Modelling the emission line spectra of galaxies is made possible with Gaussians. Done with both single spectra and IFS, Gaussians are 'overlaid' onto the emission line spectra in combinations of 1, 2 or 3 Gaussians for each emission line peak. The peaks are associated with particular element transitions which inform us on the physical processes ongoing in our galaxies.
Figure from Hampton et al., 2017: Displaying the spectra from two different positions within a galaxy.
The above figure displays two example spectra from the same galaxy. The peaks represented are taken from the IFS observations of S7 (Dopita et al., 2015). To note is that the peaks in the top panels are not similar in shape to the bottom panels. Stating that spectrums from different parts of the same galaxy need to be treated differently.
The top panels in the Figure are complicated peaks that require a combination of Gaussians to model the physical processes producing them. While the bottom panels are simpler and could be modelled with a single Gaussian. Modelling of using different numbers of Gaussians is readily done with programs like LZIFU (Ho et al., 2016b). But we ran into a problem with determining what number of Gaussians fits the spectra best.
Previously F-tests have been tried, but didn't match what Astronomer expected, in making the decision on the best number of Gaussians. Using teams of Astronomers has also been tried but it is a time consuming process that doesn't scale to the large IFS surveys like SAMI (Croom et al., 2012) where it would take years to completely eye-ball every spectrum.
To solve this problem we developed an Artificial Neural Network (LZComp; Hampton et al., (2017)). The ANN is trained by Astronomers with a small subset of galaxies and then let-loose on the remainder of the survey. The results showed that this ANN could match the selections of using Astronomers, pick out physically motivated components, and do it all in less than a week for a survey the size of ~3,000 IFS observations (including training).
To find out more then please check out my paper describing the methods, and testing we did to develop LZComp. Archive link.