Machine learning applications for the study of AGN physical properties using photometric observations

We investigate the physical nature of active galactic nuclei (AGNs) using machine learning (ML) tools. We show that the redshift, $z$, bolometric luminosity, $L_{\rm Bol}$, central mass of the supermassive black hole (SMBH), $M_{\rm BH}$, Eddington ratio, $\lambda_{\rm Edd}$, and AGN class (obscured or unobscured) can be reconstructed through multi-wavelength photometric observations only. We trained a random forest regressor (RFR) ML-model on 7616 spectroscopically observed AGNs from the SPIDERS-AGN survey, which had previously been cross-matched with soft X-ray observations (from ROSAT or XMM), WISE mid-infrared photometry, and optical photometry from SDSS \textit{ugriz} filters. We built a catalog of 21050 AGNs that were subsequently reconstructed with the trained RFR; for 9687 sources, we found archival redshift measurements. All AGNs were classified as either type 1 or type 2 using a random forest classifier (RFC) algorithm on a subset of known sources. All known photometric measurement uncertainties were incorporated via a simulation-based approach. We present the reconstructed catalog of 21050 AGNs with redshifts ranging from $0 < z < 2.5$. We determined $z$ estimations for 11363 new sources, with both accuracy and outlier rates within 2%. The distinction between type 1 or type 2 AGNs could be identified with respective efficiencies of 94% and 89%. The estimated obscuration level, a proxy for AGN classification, of all sources is given in the dataset. The $L_{\rm Bol}$, $M_{\rm BH}$, and $\lambda_{\rm Edd}$ values are given for 21050 new sources with their estimated error. These results have been made publicly available. The release of this catalog will advance AGN studies by presenting key parameters of the accretion history of 6 dex in luminosity over a wide range of $z$.