Detection of absorption systems in quasar spectra using deep learning

The thesis presents the development of convolutional neural networks designed to search for absorption systems of atomic and molecular hydrogen in the spectra of quasars from the SDSS DR14 survey, as well as to identify some of their key properties. The neural networks were trained on artificially generated absorption systems of atomic and molecular hydrogen. At the same time, a non-standard approach to the analysis of comparable one-dimensional data with two-dimensional models was proposed, which made it possible to significantly improve the quality of determining the properties of the desired fragments (profiles of the desired systems). It is shown that for reasonable values of the signal-to-noise ratio and column density in the spectra, the probability of false-positive and false-negative identification may not exceed one percent.