A machine learning model for screening thermodynamic stable lead-free halide double perovskites

• Published in: Computational materials science 2022-03, Vol.204, p.111172, Article 111172
• Main Authors: Liang, Gui-Qin, Zhang, Jian
• Format: Article
• Language: English
• Subjects: Halide double perovskite; Machine learning; Model interpretation; SHapley Additive exPlanations; Thermodynamic stability
• ISSN: 0927-0256; 1879-0801
• DOI: 10.1016/j.commatsci.2021.111172

[Display omitted] •XGBoost algorithm achieves the best performance in classification and regression for predicting the thermodynamic stability of halide double perovskites with A2B′ BX6 structure.•Machine learning models with Input features only from periodic table can effectively predict thermodynamic stability of halide double perovskite compounds.•Shannon’s revised effective ionic radii of four elements used for constructing A2B′ BX6 compounds are the most important features for their thermodynamic stability prediction.•Shapley Additive exPlanations(SHAP) was employed to reveal the relationship between features and the output of prediction models.•Proposed XGBoost models were validated with known experimental compounds excluded from training dataset. Halide double perovskites have garnered significant interest due to their outstanding photovoltaic properties. The thermodynamic stability of compounds is one of the most significant properties for materials screening, which can be well indicated by the energy above the convex hull (Ehull). The Ehull of compounds can be calculated from Density Functional Theory (DFT) requiring significant computational time and cost, making it almost impossible to be utilized for screening large numbers of possible compounds. To address this challenge, a data-driven approach implemented by machine learning (ML) algorithms has been employed to obtain the optimal model for predicting the thermodynamic phase stability of Lead-free halide double perovskite through a dataset containing 469 A2B′ BX6 double perovskites with DFT-calculated Ehull values and 24 primary features from periodic table. The results indicate that XGBoost algorithm provides more excellent predictions in both classification and regression by comparing performances of various ML algorithms under 5-folds cross validation. Furthermore, the optimal model was utilized to predict the stability of 22 completely new A2B′ BX6 compounds with known experimental results, and the experiment results demonstrated that our proposed model can be effective method for screening halide double perovskite with thermodynamic stability. Finally, we employ SHapley Additive exPlanations (SHAP) for feature analysis for ML models to reveal the relationships between the feature values and target properties, which provides important guidance for material design and screening with thermodynamic stability in the future.