Analysis of the performance of polyester, polypropylene and polyester geosynthetic mats in an anaerobic bioreactor for sanitary sewage treatment using neural models

Abstract

This study evaluated the performance of polyester, polypropylene, and geosynthetic polyester membranes applied in anaerobic membrane bioreactors for domestic wastewater treatment, using artificial neural networks (ANNs) as a modeling tool. Experimental data from the literature were used to train ANNs with the Levenberg-Marquardt and Bayesian Regularization algorithms, focusing on two critical process variables: permeate flux and organic matter removal efficiency (ECOD). The results showed that both algorithms were able to satisfactorily reproduce operational behaviors, achieving high correlation coefficients across all datasets. However, LM-trained networks exhibited signs of overfitting, with excellent adjustment to training data but lower generalization capacity for validation and test sets. Conversely, BR-trained networks demonstrated greater robustness, with improved global performance and reduced sensitivity to outliers. In general, for permeate flux prediction, the Bayesian Regularization ANN with seven hidden neurons achieved a high global correlation coefficient (R = 0.983), with regression slopes close to unity and negligible intercepts, and for ECOD prediction, the Bayesian Regularization ANN with nine hidden neurons yielded a global correlation coefficient of 0.965, with outstanding predictive performance on the testing dataset (R = 0.998). Error analysis revealed stable mean squared error behavior and well-centered error distributions, indicating minimal overfitting and strong generalization. These findings confirm the potential of ANNs as an efficient tool for modeling wastewater treatment processes, with Bayesian Regularization emerging as a promising strategy for complex systems subject to experimental variability.