Performance of a deep convolutional neural network to classify crystal structures using selected area electron beam diffraction patterns containing lattice defect information

Abstract

A deep convolutional neural network (DCNN) architecture ResNet has been tested to verify its ability to handle selected area electron diffraction pattern (SADP) datasets carrying information on lattice defects including strains, thermal lattice vibrations, point defects, dislocations, and twin boundaries. The disordered states of the crystal lattices in the presence of these defects were predicted by ab initio molecular dynamics simulations, first principles geometry optimizations, and lattice manipulation operations in an effort to establish a possible dataset augmentation strategy for the improvement of classification performance of the ResNet. Using the disordered lattice information originating from the defects, test dataset SADPs were generated by simulating electron diffraction in transmission electron microscopy. The ResNet, pre-trained using SADPs from defect-free materials, showed decreasing but acceptable classification accuracies with increasing degrees of lattice disorder regarding the lattice vibrations and point defects. When tested using the diffraction patterns for strained lattices, the ResNet responded to the changing lattice symmetry when strain levels are relatively high suggesting that it has capability to discern different symmetries induced by large strains. However, the ResNet failed to recognize lattice structure when dislocations and twin boundaries were considered. It is suggested that DCNN architectures be trained over various scenarios including changes in the image feature characteristics in the diffraction patterns related to defects in future developments for improved general classification performances.