Interpretable machine learning model of effective mass in perovskite oxides with cross-scale features

Abstract

The interpretability of machine learning reveals associations between input features and predicted physical properties in models, which are essential for discovering new materials. However, previous works were mainly devoted to algorithm improvement, while the essential multi-scale characteristics are not well addressed. This paper introduces distortion modes of oxygen octahedrons as cross-scale structural features to bridge chemical compositions and material properties. Combining model-agnostic interpretation methods, we are able to achieve interpretability even using simple machine learning schemes and develop a predictive model of effective mass for a widely used material type, namely perovskite oxides. With this framework, we reach the interpretability of the model, understanding the trend of the effective mass without any prior background information. Moreover, we obtained the knowledge only available to experts, i.e., the interpretation of effective mass from the s–p orbitals hybridization of B-site cations and O²⁻ in ABO₃ perovskite oxides.