Machine learning models for predicting the dielectric constants of oxides based on high-throughput first-principles calculations

Prediction models of both the electronic and ionic contributions to the static dielectric constants have been constructed using data from density functional perturbation theory calculations of approximately 1200 metal oxides via supervised machine learning. We developed two types of random forest regression models for oxides with the ground-state crystal structures: one model requires only compositional information and the other model also uses structural information. Although the training data included various atomic frameworks, the prediction models performed well even when only compositional information was used as feature descriptors. In prediction of the electronic contributions to the dielectric constants, the accuracies of the regression models with and without structural information were comparable, while the structural descriptors more clearly improved the prediction accuracy for the ionic contributions. We also analyzed the feature importance for prediction of the dielectric constants. The mean atomic mass and mass density were determined to be significant features in prediction of the electronic contributions without and with structural information, respectively. The standard deviation of the principal quantum number and mean neighbor distance variation were found to be important for the respective prediction models of the ionic contributions. The correlations between the dielectric constants and these features are discussed, along with the underlying physical mechanisms.