Machine Learning Assisted Predictions of Intrinsic Dielectric Breakdown Strength of ABX3Perovskites

Abstract

New and improved dielectric materials with high dielectric breakdown strength are required for both high energy density electric energy storage applications and continued miniaturization of electronic devices. Despite much practical significance, accurate ab initio predictions of dielectric breakdown strength for complex materials are beyond the current state-of-the art. Here we take an alternative data-enabled route to address this design problem. Our informatics-based approach employs a transferable machine learning model, trained and validated on a limited amount of accurate data generated through laborious first-principles computations, to predict intrinsic dielectric breakdown strength of several hundreds of chemical compositions in a highly efficient manner. While the adopted approach is quite general, here we take up a specific example of perovskite materials to demonstrate the efficacy of our method. Starting from several thousands of compounds, we systematically downselect 209 insultors which are dynamically stable in a perovskite crystal structure. After making predictions on these compounds using our machine learning model, the intrinsic dielectric breakdown strength was further cross-validated using first-principles computations. Our analysis reveals that boron-containing compounds are of particular interest, some of which exhibit remarkable intrinsic breakdown strength of almost 2 GV/m.