Possibilities of Neural Network Powder Diffraction Analysis Crystal Structure of Chemical Compounds
Author:
Alexander N. Zaloga
Vladimir V. Stanovov
Oksana E. Bezrukova
Petr S. Dubinin
Igor S. Yakimov
Corporate Contributor:
Институт цветных металлов и материаловедения
Кафедра композиционных материалов и физико-химии металлургических процессов
Date:
2019Journal Name:
Journal of Siberian Federal University. ChemistryJournal Quartile in Web of Science:
без квартиляBibliographic Citation:
Alexander N. Zaloga. Possibilities of Neural Network Powder Diffraction Analysis Crystal Structure of Chemical Compounds [Текст] / Alexander N. Zaloga, Vladimir V. Stanovov, Oksana E. Bezrukova, Petr S. Dubinin, Igor S. Yakimov // Journal of Siberian Federal University. Chemistry. — 2019.Abstract:
Some possibilities of using convolutional artificial neural networks (ANN) for powder diffraction
structural analysis of crystalline substances have been investigated. First, ANNs are used to classify
crystalline systems and space groups according to calculated full-profile diffractograms calculated from
the crystal structures of the ICSD database (2017 year). The ICSD database contains 192004 structures,
of which 80% was used for in-depth network training, and 20% for independent testing of recognition
accuracy. The accuracy of classification by a network of crystalline systems was 87.9%, and that of
space groups was 77.2%. Secondly, the ANN is used for a similar classification of structural models
generated by the stochastic genetic algorithm in the search processes for triclinic crystal structures of
test compound K4SnO4 according to their full-profile diffraction patterns. The classification criterion
was the entry of one or several atoms into their crystallographic positions in the structure of a substance.
Independent deep network training was performed on 120 thousand structural models of the K4PbO4
triclinic structure generated in several runs of the genetic algorithm. The accuracy of the classification
of K4SnO4 structural models exceeded 50%. The results show that deeply trained convolutional ANNs
can be effective for classifying crystal structures according to the structural characteristics of their
powder diffraction patterns.