Modeling the reaction time required for the removal of chemical oxygen demand (COD) from wastewater is important because can be an indicator of harmful pollutants that may pose risks to human health.. Here, we evaluate the use of an inverse artificial neural network (ANNi) to optimize the removal of COD from aqueous herbicide solutions using sonophotocatalysis as the treatment process.. This algorithm takes the operating conditions as input and estimates the optimal reaction time to reach the required effluent COD. This model was validated by comparison with both experimental measurements and simulated analysis, resulting in low error (0.5%), high Pearson correlation (R2 = 0.9956), and short computing time (less than one minute). The results for the proposed methodology (ANNi-GAs) were efficient and attractive in a reasonable way providing a satisfactory picture of the relationships between the optimal input parameters. Such parameters control the degradation process of herbicides online and can be used to predict the behavior of COD removal, thus taking actions to reduce the impact of aqueous treatment with alazine and gesaprim herbicides on aquatic, soil biota, and human health. Therefore, this technique constitutes a promising framework for predicting the environmental impact of herbicides within a tolerable degree of error.

Herbicide-contaminated wastewater poses a critical environmental challenge, largely due to the difficulty of accurately predicting Chemical Oxygen Demand (COD) removal efficiency. This study introduces a systematic Artificial Neural Network (ANN) approach to model COD dynamics during the advanced oxidation (sonophotocatalytic) degradation of two commercial herbicides (Alazine and Gesaprim). Multiple backpropagation algorithms—including Gradient Descent, Conjugate Gradient, Quasi-Newton, and Levenberg-Marquardt—were evaluated to identify the most effective training strategy for COD prediction. The optimal ANN architecture (6-10-1) consisted of six input neurons, ten hidden neurons, and one output neuron, using a hyperbolic tangent sigmoid transfer function in the hidden layer and a linear transfer function in the output layer. Trained on a dataset of 500 samples (70% for training; 30% for validation), this model achieved near-perfect correlation between measured and predicted COD values (R² = 0.9999, RMSE = 0.0140). Key findings include: • Parameter Characterization: Six operational variables (reaction time, pH, TiO₂ concentration, UV intensity, ultrasound frequency, and herbicide concentration) were assessed for their impact on COD removal. • Dominant Variables: Reaction time emerged as the most influential parameter (59.91%), followed by herbicide concentration (8.3%) and TiO₂ concentration (7.5%). • Algorithm Performance: Levenberg-Marquardt provided superior predictive performance (p < 0.01) among the tested backpropagation algorithms. • Robust Validation: Friedman tests confirmed the reliability of the ANN models for COD prediction. Uniqueness of the Present Work While prior studies have underscored the growing importance of ANNs in wastewater treatment and herbicide degradation, none directly match the objectives and depth of this research. Beyond comparing multiple backpropagation algorithms and integrating sensitivity analysis, a novel “Neural Networks for Statistical Modeling” methodology was developed, complete with a comprehensive Graphical User Interface (GUI) and advanced statistical tests. This unique combination of algorithmic evaluation, in-depth statistical analysis, and user-friendly implementation offers new perspectives on optimizing herbicide degradation processes, with potential to reduce operational costs and environmental risks in agricultural and industrial applications. Future studies will focus on practical implementation strategies and scaling up to full-scale operations, further solidifying the model’s efficacy for real-time monitoring and process optimization in advanced oxidation processes.