Machine Learning-Assisted Prediction of Brilliant Green Removal from Water Using CaTiO 3 Photocatalysis: Bayesian Optimization, SHAP Explainability, and Robustness Assessment

Brilliant Green (BG)-contaminated wastewater threatens aquatic ecosystems by limiting light penetration, disrupting photosynthesis, and exposing organisms to potentially toxic compounds, raising broader environmental and human-health concerns. This study presents an explainable machine-learning framework for predicting BG removal during CaTiO 3 -assisted photocatalytic water treatment. Using 63 published experimental observations for the 2:7 precursor composition, six regression models Decision Tree, Random Forest, Extra Trees, Gradient Boosting, XGBoost, and CatBoost were evaluated through Bayesian hyperparameter optimization and repeated cross-validation. Although Extra Trees achieved marginally higher holdout accuracy, CatBoost was selected for its superior repeated-validation performance, achieving a test R 2 of 0.9884, RMSE of 3.4629 percentage points, and mean cross-validation R 2 of 0.9697. SHAP analysis identified degradation time (60.57%), initial BG concentration (32.12%), and catalyst dosage (7.31%) as the principal contributors to model predictions. Feature-combination analysis reinforced this hierarchy, while CatBoost retained a test R 2 of 0.9778 under 10% Gaussian noise. Predicted BG removal (89.32%) closely matched the reported experimental value (88.46%). These findings provide a promising foundation for future experimental validation, broader catalyst screening, and interpretable optimization of photocatalytic wastewater treatment.