Interpretable Machine Learning to Discover Perovskites with High Spontaneous Polarization

Machine learning can accelerate the design of new materials by screening large quantities of materials. We investigated the spontaneous polarization intensity of inorganic perovskite ferroelectrics using a machine learning approach. The machine learning model covers the entire structure type of perovskite ferroelectrics. We make a large number of predictions for perovskite materials based on our model and screen 20 perovskite materials that have high spontaneous polarization intensity. We employ the SHAP (Shapley additive explanations) technique to qualitatively explain the machine learning model's correctness from a physical point of view. The results show that the larger the average atomic radius and the smaller the electronegativity of the metal atoms of the perovskite, the easier it is to find greater spontaneous polarization intensity. We also screen and verify the reasonableness of descriptors based on the model interpretation to improve the reliability of the model. By utilizing an interpretable machine learning approach, we can predict and optimize the properties of ferroelectrics, which facilitates the evaluation and application of materials.