Machine learning approach to automated analysis of atomic configuration of molecular dynamics simulation
T Fukuya and Y Shibuta, COMPUTATIONAL MATERIALS SCIENCE, 184, 109880 (2020).
Three-dimensional convolutional neural network (3D-CNN) is employed for automated analysis of atomic configuration of molecular dynamics (MD) simulation. Solid and liquid atoms in the solid-liquid biphasic system of various elements at high temperature are identified by a 3D-CNN architecture. Accuracy of 3D-CNN successfully achieves more than 90% independent of crystal structure, whereas accuracy of common neighbor analysis (CNA) is approximately 50% at most for the same system. 3D-CNN can extract the morphology of solid-liquid interface very clearly including roughness at atomistic scale. Moreover, 3D-CNN trained by the data set of a certain element (iron) can be applied for another element (tungsten) of same crystal structure without further training. It is significant in this study to shed light on a high potential of machine learning (ML)-based approach for automated analysis of atomistic configuration since it is not straightforward to develop an identifier of atomic configuration manually when we face a new problem out of existing methodologies.
Return to Publications page