Speaker
Description
A complete census of supermassive black holes (SMBH) increases our understanding of the role of Active Galactic Nuclei (AGN) in the formation and evolution of galaxies. As AGN detection is less affected by obscuration effects in the X-ray window, eROSITA offers an increased likelihood of detecting these kinds of objects. That being said, a substantial fraction of spectroscopic redshifts (spec-z) for AGN identified by eROSITA will be available only in 2-3 years from now at best. In the meantime, we must rely on photometric redshifts (photo-z). For wide-area surveys, the quality of current estimates of photo-z for AGN (both via SED fitting and ML techniques) using broad-band photometry is poor because the limited number of photometric bands is insufficient to disentangle the relative AGN/host-galaxy contribution, thus producing an undesired high fraction of outliers. More recent efforts to compute photo-z for AGN using a set of aperture photometry (as provided by Legacy Survey DR9 and DR10) via ML, have shown promising improvement. Catalogs, however, depend on parameters established by the source detection and flux estimate algorithms, where they are usually fine-tuned for galaxies and not AGN. For this reason, we provide a novel image-based Machine Learning (ML) algorithm, namely a convolutional neural network (CNN), to alleviate previous empirical approaches and decrease the fraction of objects with predicted redshifts classified as outliers. Special attention is given to creating a clean training sample and ML optimisation techniques for redshift determination specific to AGN. In my talk, I will show how our preliminary work already outperforms previous results.
