Accurate prediction of dielectric properties and bandgaps in materials with a machine learning approach

The conventional approach to exploring suitable dielectrics for future logic and memory devices relies on first-principle calculations, which are expensive and time-consuming. In this work, we adopt a data-driven machine learning (ML)-based approach to build a model for predicting these properties. By incorporating structural information into the input descriptors, we achieve record-high accuracy in predicting the dielectric constant, with the coefficients of determination (R2) of 0.886 and root mean square error (RMSE) of 0.083. Additionally, we achieve high predictions for the bandgap, with accuracies of 0.832 and 0.533 for R2 and RMSE, respectively. The features corresponding to specific properties are analyzed to obtain physical insights. Finally, we employ first-principle calculations to validate the feasibility of this model. This work proposes a highly efficient approach for using ML to predict material properties.