We present SRGz - a programming package for doing effective optical cross-match, photometric classification and probabilistic photo-z measurements of SRG extragalactic sources. SRGz is based on competitive empirical machine learning (ML) algorithms: quantile random forest, gradient boosting, deep neural networks. ML-models were trained on SDSS spectral samples of quasars, galaxies and stars and samples of optical sources in the vicinity of X-ray sources from XMM-Newton Serendipitous Source Catalog and can provide accurate photometric prior probabilities for optical counterparts of SRG sources detected during the all-sky survey, their STAR/QUASAR/GALAXY classification scores and photometric redshifts in various forms (photo-z point predictions, confidence intervals and full probability distribution functions PDZ). The proposed methods allow one to precisely identify optical counterparts for the majority of X-ray sources in the SDSS footprint and accurately measure their redshifts with low fraction of catastrophic outliers. Quality of optical cross-match and photo-z prediction models were intensively tested on the spectroscopic X-ray samples in the Stripe82X and XMM-XXL-N extragalactic fields.
|Presenter status||eROSITA/RU consortium member|