Screening stable and metastable ABO3 perovskites using machine learning and the materials project

[Display omitted] •9 crucial features are identified to train the machine learning model.•10-fold cross-validation is used to evaluate the performance of all models.•Grid search method is employed to optimize hyper-parameters of all models.•Gradient Boosting Decision Tree is the best classification model in this work.•37 stable and 13 metastable ABO3 perovskite candidates are screened out. Machine learning and Materials Project are used to investigate stable and metastable perovskite materials based on a dataset of 397 ABO3 compounds. The best performance classification model Gradient Boosting Decision Tree (GBDT) can classify 397 compounds into 143 non-perovskites and 254 perovskites with a 94.6% accuracy over 10-fold cross-validation, which indicates that 9 descriptors are outstanding features for formability of perovskite: tolerance factor, octahedral factor, radius ratio of A to O, A-O and B-O bond length, electronegativity difference for A-O (B-O) multiplied by the radius ratio of A (B) to O, the Mendeleev numbers for A and B. Among 891 ABO3, the GBDT model predicts that 331 have perovskite structure and the top-174 within a probability ≥ 85%. Furthermore, based on the energy above the convex hull (Ehull), 37 thermodynamically stable ABO3 perovskites with 0≤Ehull