Machine learning and atomistic origin of high dielectric permittivity in oxides

Abstract

Discovering new stable materials with large dielectric permittivity is important for future energy storage and electronics applications. Theoretical and computational approaches help design new materials by elucidating microscopic mechanisms and establishing structure-property relations. Ab initio methods can be used to reliably predict the dielectric response, but for fast materials screening, machine learning (ML) approaches, which can directly infer properties from the structural information, are needed. Here, random forest and graph convolutional neural network models are trained and tested to predict the dielectric constant from the structural information. We create a database of the dielectric properties of oxides and design, train, and test the two ML models. Both approaches show similar performance and can successfully predict response based on the structure. The analysis of the feature importance allows identification of local geometric features leading to the high dielectric permittivity of the crystal. Dimensionality reduction and clustering further confirms the relevance of descriptors and compositional features for obtaining high dielectric permittivity.