Deep Learning in Searching the Spectroscopic Redshift of Quasars
Abstract
Studying the cosmological sources at their cosmological rest-frames is crucial to track the cosmic history and properties of compact objects. In view of the increasing data volume of existing and upcoming telescopes/detectors, we here construct a 1--dimensional convolutional neural network (CNN) with a residual neural network (ResNet) structure to estimate the redshift of quasars in Sloan Digital Sky Survey IV (SDSS-IV) catalog from DR16 quasar-only (DR16Q) of eBOSS on a broad range of signal-to-noise ratios, named FNet. Owing to its 24 convolutional layers and the ResNet structure with different kernel sizes of 500, 200 and 15, FNet is able to discover the "local" and "global" patterns in the whole sample of spectra by a self-learning procedure. It reaches the accuracy of 97.0\% for the velocity difference for redshift, ||< 6000~ km/s and 98.0\% for ||< 12000~ km/s. While QuasarNET, which is a standard CNN adopted in the SDSS routine and is constructed by 4 convolutional layers (no ResNet structure), with kernel sizes of 10, to measure the redshift via identifying seven emission lines (local patterns), fails in estimating redshift of 1.3\% of visually inspected quasars in DR16Q catalog, and it gives 97.8\% for ||< 6000~ km/s and 97.9\% for ||< 12000~ km/s. Hence, FNet provides similar accuracy to QuasarNET, but it is applicable for a wider range of SDSS spectra, especially for those missing the clear emission lines exploited by QuasarNET. These properties of FNet, together with the fast predictive power of machine learning, allow FNet to be a more accurate alternative for the pipeline redshift estimator and can make it practical in the upcoming catalogs to reduce the number of spectra to visually inspect.