Author：Tatsuya Yokoi
Affiliation：Department of Materials Physics, Graduate School of Engineering
Abstract：This work developed a trainable descriptor for machine learning potentials to predict atomic structures of grain boundaries in silicon. Our descriptor exhibited lower errors than commonly used descriptors based on analytic functions for both training and test datasets, demonstrating excellent transferability. Using the trainable descriptor, energetically favorable atomic structures of asymmetric tilt grain boundaries were systematically explored by varying misorientation angles between two grains. Complex structural units were formed in a certain range of angles. This is probably because asymmetric tilt grain boundaries in this range have {111} planes that make small angles relative to the interfacial planes, and hence the {111} planes partly form twin boundaries with notably low grain boundary energies by altering the original structural units.