QuasarNET: Human-level spectral classification and redshifting with Deep Neural Networks

TL;DR

QuasarNET introduces a convolutional neural network that treats quasar spectral classification and redshift estimation as a feature-detection task, matching human-expert performance on BOSS data. The model detects seven emission-line features across multiple wavelength intervals, yielding purity and completeness around 99.5% for quasar selection and a redshift precision with Δv ~ 8 km/s and catastrophic failures < 0.2%. It extends to BAL quasar detection, achieving high BAL/non-BAL recognition while maintaining redshift reliability. Compared with automatic PCA-based classifiers, QuasarNET delivers superior sample quality with reduced need for visual inspection, and its training/inference efficiency supports application to upcoming surveys like DESI and 4MOST. The approach shows robust performance across redshift and motivates using QuasarNET outputs as priors for more precise spectral modeling.

Abstract

We introduce QuasarNET, a deep convolutional neural network that performs classification and redshift estimation of astrophysical spectra with human-expert accuracy. We pose these two tasks as a \emph{feature detection} problem: presence or absence of spectral features determines the class, and their wavelength determines the redshift, very much like human-experts proceed. When ran on BOSS data to identify quasars through their emission lines, QuasarNET defines a sample $99.51\pm0.03$\% pure and $99.52\pm0.03$\% complete, well above the requirements of many analyses using these data. QuasarNET significantly reduces the problem of line-confusion that induces catastrophic redshift failures to below 0.2\%. We also extend QuasarNET to classify spectra with broad absorption line (BAL) features, achieving an accuracy of $98.0\pm0.4$\% for recognizing BAL and $97.0\pm0.2$\% for rejecting non-BAL quasars. QuasarNET is trained on data of low signal-to-noise and medium resolution, typical of current and future astrophysical surveys, and could be easily applied to classify spectra from current and upcoming surveys such as eBOSS, DESI and 4MOST.

Figures (11)

• Figure 1: Neural Network architecture. Layers 1-4 are convolutional layers of 100 filters of size 10 and strides of 2 and rectified linear unit (ReLU) activations. Layer 5 is a fully connected layer of size 100 and a sigmoid activation. Each line finder consists of 13 sigmoid units for the coarse-grained confidences and 13 sigmoid units for the fine-grained line position.
• Figure 2: An example of a high-redshift quasar spectrum from BOSS (identified by the plate id, the modified julian day (mjd) and the fiber id of the observation) annotated as a quasar and with high-confidence emission-line detections by QuasarNET. The broad absorption line annotations, BAL_FLAG_VI and BI_CIV are also shown (see § \ref{['ssec:results_bal']}). The flux has been renormalized as explained in the text. Also indicated are the positions of the lined as found by QuasarNET (black dashed lines) and the confidences. Low-confidence lines are grayed out.
• Figure 3: Purity (blue) and completeness (red) of the predicted quasar sample as a function of threshold confidence for one 1 line (solid) or two lines (hatched). Bands represent the root-mean-square calculated over ten random 80/20 training/validation data splits. The horizontal lines correspond to the purity (blue) and completeness (red) for the automatic procedure from BOSS (solid) and the automatic procedure improved by visual inspections from eBOSS (dashed).
• Figure 4: Velocity difference implied by the predicted and annotated redshifts (blue), and Z_VI and Z_PCA redshifts from Paris:2016xdm (black dashed line).
• Figure 5: pseudo-purity (blue) and pseudo-completeness (red) for the detection of quasar spectra exhibiting a BAL feature bluewards of the CIV emission line as a function of threshold confidence.