Possibilities for neural network powder diffraction analysis of the crystal structure of chemical compounds

Some possibilities of using convolutional artificial neural networks (ANNs) for powder diffraction structural analysis of crystalline substances are investigated. First, ANNs were used to classify crystal systems and space groups of symmetry based on the full-profile diffractograms calculated from the crystal structures of the ICSD 2017 database. The ICSD database contains 192004 structures, of which 80% were used for deep network learning, and 20% for independent testing of recognition accuracy. The accuracy of classification by a network of crystalline systems was 87.9%, and space groups was 77.2%. Second, ANN was applied for a similar classification of structural models generated by a stochastic evolutionary algorithm in the search for triclinic crystal structures of test compounds K4SnO4 and Rb4SnO4 from their full-profile diffraction patterns. The classification criterion was the hit of one or more atoms in their correct crystallographic positions in the structure of the substance. Independent deep learning of the network was performed on 100 thousand structural models of the triclinic structure of K4SnO4, generated in several runs of the evolutionary algorithm. The classification accuracy of the structural models of these substances was 72.2% and 76.9%, respectively.