Materials informatics for dielectric loss tangent in the millimeter wave region

Abstract

For the prediction of the dielectric loss tangent in the millimeter wave region, machine-learning approaches based on the first-principles calculations were carried out. The data set was prepared by the first-principles calculations considering the anharmonicity of lattice vibrations. The two-phonon density of states, which is correlated with the dielectric loss tangent, was calculated and confirmed the connection with the difference in crystal structures. Machine-learning models to predict the dielectric loss tangent were created considering both atomic compositions and crystal structures as descriptors. In addition, transfer-learning models, in which a pretrained model for the two-phonon density of states was used as the new descriptor, were compared with models from scratch. The transfer-learning model showed 25% higher prediction accuracy than the scratch model.