Alternative assessment of machine learning to polynomial regression in response surface methodology for predicting decolorization efficiency in textile wastewater treatment.

Abstract

This study investigated the potential of machine learning (ML) as a substitute for polynomial regression in conventional response surface methodology (RSM) for decolorizing textile wastewater via a UV/H₂O₂ process. While polynomial regression offers limited adaptability, ML models provide superior flexibility in capturing nonlinear responses but are prone to overfitting, particularly with constrained RSM datasets. In this study, we evaluated decision tree (DT), random forest (RF), multilayer perceptron (MLP), and extreme gradient boosting (XGBoost) models with respect to a quadratic regression model. Our observations indicated that the ML models achieved higher R² values, demonstrating better adaptability. However, when provided with additional data, the polynomial regression displayed a moderate predictability, whereas MLP and XGBoost exhibited indications of overfitting, while DT and RF remained robust. Both ANalysis Of VAriance (ANOVA) and SHapley Additive exPlanations (SHAP) analyses consistently emphasized the significance of operational factors (H₂O₂ concentration, reaction time, UV light intensity) in decolorization. The findings underscore the need for cautious validation when substituting ML models in RSM and highlight the complementary value of ML (particularly SHAP analysis) alongside conventional ANOVA for analyzing factor significance. This study offered significant insights into replacing polynomial regression with ML models in RSM.