Data-driven COD prediction during TiO ₂ –P25 sonophotocatalysis of herbicide-contaminated wastewater: a neural-network benchmark

Abstract Herbicide-contaminated wastewater is difficult to treat, and predicting Chemical Oxygen Demand (COD) removal in advanced oxidation processes remains challenging. This work develops an artificial neural network (ANN) framework to predict COD removal during sonophotocatalytic degradation of two commercial herbicides (Alazine and Gesaprim) using TiO 2 –P25. A data set of 500 experiments was generated by varying reaction time, pH, TiO 2 concentration, UV power, ultrasound frequency, and herbicide concentration. Five backpropagation training algorithms were benchmarked in an identical 6-10-1 ANN architecture: Gradient Descent, Conjugate Gradient Polak–Ribiere, Scaled Conjugate Gradient, Quasi-Newton BFGS, and Levenberg–Marquardt. The Levenberg–Marquardt algorithm achieved the highest accuracy ( R 2 = 0.9999; RMSE = 0.014 mg/L), and statistical tests confirmed its superiority over the other methods. Sensitivity analysis showed that reaction time was the dominant factor, followed by herbicide and TiO 2 concentrations. A user-friendly graphical interface was developed to load data, train models, compare algorithms, and perform “what-if” scenario analysis in real time. The proposed ANN-plus-GUI toolkit reduces experimental effort and offers a practical decision-support tool for designing and optimizing sustainable treatment of herbicide-contaminated wastewater.