Machine learning-assisted ZnO-doped photocatalytic degradation of rhodamine B: Prediction, analysis, and optimization

Photocatalysts have been extensively investigated for their applications in dye degradation; however, determining the optimal combination of reaction parameters remains a substantial challenge. Here, machine learning (ML) is employed to predict and optimize the photocatalytic degradation efficiency of rhodamine B (RhB) by ZnO catalysts doped with different elements. A dataset of 905 entries with 19 doped elements (In, Ba, Fe, Ni, Rb, Sb, Sr, Ce, V, Eu, Ag, Mn, Ho, Sm, N, B, Au, Pt, C) is constructed, initially containing 16 input variables, later expanded to 23 features grouped into catalyst properties, reaction conditions, and preparation conditions through feature engineering. Eleven ML models are compared using four evaluation metrics, revealing that CatBoost provided the best predictive performance ( R 2 =0.96, MAE =0.0359, RMSE =0.0579, MAPE =35.04 %). Feature importance analysis showed that reaction time, doped element, calcination temperature, and light wavelength significantly affected degradation efficiency. Reaction conditions contributed the most (70.7 %), followed by catalyst properties (22.1 %) and preparation conditions (7.2 %). Furthermore, the CatBoost model’s generalization is validated using magnesium (Mg)—unseen during training—to predict a parameter space of 1036,800 data points. Optimal reaction parameters predicted by the model are successfully confirmed by photocatalytic experiments. • Curated 905 data entries with 19 doped elements and expanded features. • Achieved high predictive accuracy (R 2 = 0.96) using CatBoost. • Constructed 1036,800 data points for Mg-doped ZnO and verified.