Integrating NSGA-II and TOPSIS for Stacking Model Optimization in Pursuit of Halide Double Perovskite Screening

Halide double perovskite materials have been used for various applications; their bandgap (E_g) and heat of formation (ΔH_f) are their key properties. They can be obtained through calculations based on high-throughput density functional theory (DFT), but such calculations are computationally expensive and time-consuming. Machine learning (ML) has proved to be an effective tool for screening potential materials. The prediction accuracy of ML models strongly depends on both input features and ML algorithms. However, there is no unified feature set with which ML models can effectively distinguish halide double perovskite materials. Although it has been proven that stacking ML models can achieve higher prediction accuracy than individual ML models, little attention has been paid to the optimization of stacking models. To solve these problems, we constructed a new feature set obtained from periodic tables for predicting the E_g and ΔH_f of halide double perovskites, and we further proposed a method integrating the nondominated sorting genetic algorithm (NSGA-II) and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) decision-making tool for stacking model optimization to predict the E_g and ΔH_f of 540 compounds of halide double perovskites. Experimental results from 40 runs of 5-fold cross-validation demonstrate that our proposed new feature set enables ML models to achieve better performance than the original feature set. Moreover, the stacking model optimized by our proposed method yields better predicting performance than that of any individual single model and stacking regression models without optimization, with average improvements of 5.02%, 2.70%, 3.72% and 0.28% in MSE, RMSE, MAE and R², respectively, in E_g prediction, thus providing more effective guidance for screening potential compounds for solar cells from a large quantity of materials.