Evaluation of performance of machine learning methods in mining structure-property data of halide perovskite materials

With the rapid development of artificial intelligence and machine learning (ML) methods, materials science is rapidly entering the era of data-driven materials informatics. ML models serve as the most crucial component, closely bridging material structure and material properties. There is a considerable difference in the prediction performance of different ML methods for material systems. Herein, we evaluated three categories (linear, kernel, and nonlinear methods) of models, with twelve ML algorithms commonly used in the materials field. In addition, halide perovskite was chosen as an example to evaluate the fitting performance of different models. We constructed a total dataset of 540 halide perovskites and 72 features, with formation energy and bandgap as target properties. We found that different categories of ML models show similar trends for different target properties. Among them, the difference between the models is enormous for the formation energy, with the coefficient of determination (R-2) range 0.69-0.953. The fitting performance between the models is closer for bandgap, with the R-2 range 0.941-0.997. The nonlinear-ensemble model shows the best fitting performance for both the formation energy and the bandgap. It shows that the nonlinear-ensemble model, constructed by combining multiple weak learners, effectively describes the nonlinear relationship between material features and target property. In addition, the extreme gradient boosting decision tree model shows the most superior results among all the models and searches for two new descriptors that are crucial for formation energy and bandgap. Our work provides useful guidance for the selection of effective machine learning methods in the data-mining studies of specific material systems. The dataset that supported the findings of this study is available in Science Data Bank, with the link .