A machine learning model with crude estimation of property strategy for performance prediction of perovskite solar cells based on process optimization

Perovskite solar cells (PSCs) have attracted significant attention due to their high power conversion efficiency (PCE) and affordability. However, optimizing the preparation parameters for PSCs is crucial. This study establishes a machine learning model incorporating a crude estimation of property (CEP) strategy to enhance prediction accuracy and precisely control process parameters. The model's evaluation metrics improved by utilizing excess non-stoichiometric components (Ensc) and perovskite additive compounds (Pac) as CEP. Notably, the coefficient of determination (R-2) on the test set increased by 16.14%, while the root mean square error decreased by 20.44%, respectively. Nine machine learning algorithms, including decision tree (DT), random forest (RF), CatBoost, LassoLarsCV, histogram gradient boosting, extreme gradient boosting (XGBoost), K nearest neighbor, ridge regression (Ridge), and linear regression (Linear R), were employed to optimize PSC preparation and assess its impact on device performance. The best-performing models, DT and RF, were combined to create a stacking model demonstrating the most stable overall performance on training and test sets. The study identified key process parameters affecting PCE based on the stacking model. Among these, adding Ensc was the most critical factor, followed by perovskite thickness, thermal annealing time (Ta-ti), perovskite deposition solvent (Pds), solvent mixing ratio, and Pac. Experimental verification showed that PSCs with a 10% excess of PbI2 exhibited higher PCE compared to those with 5% excess, confirming that adding Ensc can effectively enhance PCE. These findings offer a valuable reference for optimizing PSC process parameters and improving performance, thereby saving time and labor costs.