X-ray spectra are used for characterization of materials from astronomy to industry. In this work we will develop machine-learning-based methods to interpret these spectra for structural information. The developed algorithm filters out structural variations irrelevant to spectra, and thus allows for a drastic dimensionality reduction for the problem to be solved. After identification of the spectrally relevant degrees of freedom, the inverse problem posed by X-ray spectra will be solved in terms of this solvable part. The work is based on training neural networks with data from supercomputer simulations. The results of the developed reconstruction method will be checked against experimental evidence. The proposed method is not specific to spectroscopy, but can be applied to general inverse problems found in a wealth of scientific and technological questions. In the work the approach will be generalized to UV/visible region.

2025

2029