Estimation of redshift and associated uncertainty of Fermi/LAT extra-galactic sources with Deep Learning

TL;DR

This paper presents a deep learning approach to estimate the redshift of gamma-ray loud AGNs from Fermi-LAT data, including uncertainty quantification, to aid in understanding the universe's structure.

Contribution

It introduces a simple deep learning model with variational inference methods for redshift prediction and uncertainty estimation of AGNs from Fermi-LAT data.

Findings

Correlation coefficient of 0.784 on test set

Redshift predictions with mean around 0.4

Uncertainty estimates with standard deviations below 0.3

Abstract

With the advancement of technology, machine learning-based analytical methods have pervaded nearly every discipline in modern studies. Particularly, a number of methods have been employed to estimate the redshift of gamma-ray loud active galactic nuclei (AGN), which are a class of supermassive black hole systems known for their intense multi-wavelength emissions and violent variability. Determining the redshifts of AGNs is essential for understanding their distances, which, in turn, sheds light on our current understanding of the structure of the nearby universe. However, the task involves a number of challenges such as the need for meticulous follow-up observations across multiple wavelengths and astronomical facilities. In this study, we employ a simple yet effective deep learning model with a single hidden layer having $64$ neurons and a dropout of 0.25 in the hidden layer, on a sample of AGNs with known redshifts from the latest AGN catalog, 4LAC-DR3, obtained from Fermi-LAT. We utilized their spectral, spatial, and temporal properties to robustly predict the redshifts of AGNs as well quantify their associated uncertainties, by modifying the model using two different variational inference methods. We achieve a correlation coefficient of 0.784 on the test set from the frequentist model and 0.777 and 0.778 from both the variants of variational inference, and, when used to make predictions on the samples with unknown redshifts, we achieve mean predictions of 0.421, 0.415 and 0.393, with standard deviations of 0.258, 0.246 and 0.207 from the models, respectively.