Machine learning and deep learning for predicting photocatalytic degradation efficiency of metronidazole via TiO2/ZnO nanocomposites: a response surface methodology approach

Abstract

Abstract Inappropriate disposal of antibiotics in aquatic and soil environments makes bacteria resistant, which is a potential threat to humans and other organisms. Photocatalysis is a simple, inexpensive, and eco-friendly process that is considered an attractive option for degrading antibiotics. Here, we synthesized the TiO 2 /ZnO nanocomposite as a photocatalyst by the sol-gel method and characterized by EDX, TEM, FTIR, SEM, and XRD. Using the response surface methodology based on central composite design, the effect of different variables: pH, irradiation time, metronidazole (MNZ) concentration, and catalyst dose on the photodegradation of MNZ was investigated and optimized. Under most advantageous conditions, the synthesized nanocomposite is capable of degrading MNZ by a significant 94.92%. In this study, the recyclability, mechanism, and effect of light source intensity were also investigated. Machine learning and deep learning models were employed to predict the photocatalytic degradation efficiency of MNZ, guided by a response surface methodology (RSM). Among the evaluated models-support vector regression (SVR), artificial neural network (ANN), fully connected neural network (FCNN), and random forest (RF)- SVR demonstrated the highest predictive accuracy, SVR demonstrated the highest predictive accuracy, achieving R  = 0.9495, R 2  = 0.8502, and the lowest MSE (7.2560) and RMSE (2.6937) among the evaluated models. Feature importance analysis revealed that reaction time and pH were the most influential parameters, followed by MNZ concentration, while catalyst dose had minimal impact. These findings underscore the effectiveness of kernel-based SVR in modeling complex photocatalytic systems with limited data and highlight the critical role of reaction conditions in optimizing degradation performance.