Physics-Informed Descriptor Engineering and Explainable Machine Learning Reveal the Quantum Mechanical Origins of Electronic Dielectric Response in Oxide Materials

Abstract

A physics-informed machine-learning model was created to forecast the dielectric constant of oxide materials. The study explores the physicochemical descriptors that govern the response of the materials. A set of 3304 non-metallic oxide compounds was built using the Materials Project database. A comprehensive feature space with 23 original and composition-derived variables as well as 10 physics-informed engineered variables representing electronic, thermodynamic, structural, magnetic, and chemical descriptors was constructed. This workflow included model benchmarking, hyperparameter optimization, repeated training (10 times) and testing, consensus feature selection, and SHapley Additive exPlanations (SHAP) analysis. Optimized XGBoost model was the model with the best overall predictive ability. A reduced representation of 25 descriptors was obtained by consensus-based feature selection that performed close to the complete 33-descriptor representation (MAE = 0.611, RMSE = 1.052, and R2 = 0.699). The SHAP analysis revealed that the descriptors associated with chemical properties, electronic excitation and atomic packing were key to the model predictions. The results indicate that the electronic dielectric response is related to a combination of a few physicochemical aspects instead of a single descriptor. The proposed framework offers an intuitive, data-centric method for examining the evolution of dielectric properties, and it can be used to screen and hypothesize oxide materials with desired electronic dielectric response.