Skip to main content
SearchLogin or Signup

FeLoNet: Using Synthetic Training Data and a Convolutional Neural Network to Classify Quasar Spectra

Published onJun 01, 2020
FeLoNet: Using Synthetic Training Data and a Convolutional Neural Network to Classify Quasar Spectra

There is a well-known correlation between the mass of the supermassive black hole at the center of a galaxy and the mass of that galaxy’s spheroidal component. However, the sphere of gravitational influence for the supermassive black hole is small relative to the size of the galaxy. This means that there must be some method of feedback between the supermassive black hole and its host. A promising potential avenue for such feedback is quasar outflows. Quasar outflows manifest themselves unambiguously through blue-shifted absorption lines in Broad Absorption Line Quasars (BALQs). The rarest subclass of BALQs with the most morphologically diverse spectra are known as FeLoBALs. These objects are named for the presence of absorption lines from Fe+, observed in the rest frame UV. The presence of thousands of observable Fe+ transitions allows for excellent constraints on the physical conditions of the outflowing gas. In particular, FeLoBALs provide strong density constraints, which allow us to accurately estimate the distance of the outflow from the central engine. FeLoBALs have the thickest outflows of any BALQ subclass. As outflow energy is directly proportional to column density this makes them important feedback candidates. A difficulty is that FeLoBALs are extremely rare, making up roughly 0.1% of all quasars. On top of this, they are significantly under-identified in recent SDSS releases. We present FeLoNet, a Convolutional Neural Network (CNN) designed to identify FeLoBAL quasars in the SDSS DR14 Quasar Catalog. As there were fewer than 200 identified FeLoBALs prior to this work, we trained our CNN on synthetic data produced with our novel spectral synthesis tool, SimBAL. We have tripled the number of identified FeLoBAL quasars, allowing for statistical analysis of their outflow properties and feedback potential. The model's success at classifying real objects with synthetic training data speaks to the potential for deep learning techniques for spectral classification.

Comments
0
comment

No comments here