Identifying descriptors for perovskite structure of composite oxides and inferring formability via low-dimensional described features

[Display omitted] •A framework of descriptors identified based on feature selection and compressed sensing method is proposed.•Two descriptors were developed and identified to analyze the perovskite structure.•Analyzed and interpreted the relationship between the main features and constructed descriptors by the shapley additive explanation and the decision boundary.•GBDT classification model was selected from several machine learning algorithms to predicts the formability.•The proposed strategy can shed light on the development of new materials with targeted perovskite structures. As potential perovskite candidates, ABO3 compounds have been explored to determine whether they can have perovskite structures. To address this, in this study, a comprehensive set of features was established based on chemical composition and physical structure from a raw dataset of 435 ABO3 compounds. First, considering the application of compressed sensing method to reduce high dimensional features, two accurate and easily interpretable new descriptors were created and identified, which combined with tolerance factor t, octahedral factor u, B-site element Mendeleev number M_B and B-site volume to predict the formability of perovskite structure from unknown material. Additionally, the relationship between the main features and constructed descriptors was analyzed and interpreted using the shapley additive explanation (SHAP) and the decision boundary. On the basis of the selected GBDT classification model with the best performance from several machine learning algorithms, 591 novel ABO3-type compounds were predicted for the formability and screened out as perovskite candidates with high forming probability. This approach provides a practical method for rapidly and effectively screening and identifying potential perovskite candidates.