Prediction of the formability and stability of perovskite oxides via multi-label classification

Abstract

Perovskite oxides are promising candidates for diverse applications due to their versatile physical and chemical properties. However, their structural and compositional flexibility significantly delay the traditional methods of screening formable and thermodynamically stable perovskite oxides. Single-label machine learning methods have been extensively used to solve this challenge, but these often result in the misselection of unstable perovskite oxides by formability prediction models and non-formable perovskite oxides by stability prediction models. Here, multi-label classification (MLC) methods are employed to simultaneously screen for both formable and stable perovskite oxides. We investigate the label dependency of formability and stability labels, finding significant unconditional dependency but little conditional dependency. Using a recursive feature addition method, 10 features are selected from an initial set of 159. SHapley Additive exPlanations (SHAP) analysis reveals that the atomic weight of B-site elements and the ionic radii ratio of the A-site to the B-site cations are the most important features. Among the eight MLC methods evaluated, the classifier chains (CC) model outperforms its counterparts. An optimized CC model achieves outstanding performance with a subset accuracy of 0.932 and a Hamming loss of 0.0342. This model is further generalized on 2226 virtual perovskite combinations, identifying 42 formable and stable perovskite oxides for future investigation. This work presents an effective approach for screening potential perovskite oxides, which can be further extended to other fields that involve predicting multiple properties concurrently.