Highly Efficient Screening of Halide Double Perovskite Optoelectronic Materials Based on Machine learning

The photoelectronic properties and corresponding applications of halide perovskites significantly depend on their band gaps and formation energy. However, experiments and density functional theory (DFT) calculations are usually time consuming and laborious to obtain these properties. In this study, the formation energy, band gap, and band gap classification label of halide double perovskites were predicted in terms of material parameters via using the gradient boosting tree combined with the genetic algorithm and grid search algorithm. The coefficients of determination (R²) of GA-GBR_f and GRID-GBR_b were improved to 0.9958 and 0.9206, respectively, and the accuracy of GA-GBC_b was 0.9273. A set of 1515 candidates with stable structure and band gaps (1-4 eV) was screened out from 77,604 halide double perovskites through multistep prediction via optimized models. Forty candidates were randomly selected for density functional theory calculation, which successfully verified the robustness of optimized models. In addition, the relationship between the properties and feature parameters was discussed by SHapley Additive exPlanations (SHAP). Furthermore, a perovskite Cs₂RbBiI₆ obtained from the efficient screening was selected for experimental evaluation as an example, which was successfully applied for photodetection and photocatalysis. This study provides ideas for discovering materials for specific applications at the low cost of time-consuming and experimental resources.