A Statistical Modelling and Machine Learning Approach for Textile Wastewater Treatment: Response Surface Methodology, Random Forest Regression and Monte Carlo Analysis

Aerobic ceramic membrane bioreactors (AeCeMBR) have shown great potential in treating wastewater (WW) from the textile industry; however, their operation faces challenges such as process variability, membrane contamination, and the need for accurate prediction of treated water quality under varying conditions. In this study, chemical oxygen demand (COD) and turbidity were selected as key indicators, as they directly reflect organic load removal and solids separation efficiency in MBR systems. The effect of four operational parameters: hydraulic retention time (HRT), organic loading rate (OLR), mixed liquor suspended solids (MLSS), and transmembrane pressure (TMP), was investigated using a response surface methodology (RSM) based on a Box–Behnken design. A random forest (RF) model coupled with Monte Carlo simulation (MC) was also developed using 174 experimental data points to enhance predictive power and quantify uncertainty. The RSM model showed strong agreement with experimental results (coefficient of determination (R2) > 0.95), achieving approximately 96% removal for both COD and turbidity, with validation errors of less than 2%. MC simulation (10,000 iterations) was applied to assess the effect of ±10% variance under operating conditions, providing a probabilistic view of system performance. The RF-MC framework demonstrated high predictive accuracy, with strong correlations between predicted and observed values (R2 = 0.92 for COD and 0.97 for turbidity) and low uncertainty. Overall, this study proposes an integrated RSM, RF–MC approach for AeCeMBR systems, providing a robust and uncertainty-aware framework for process optimization and performance prediction under changing operating conditions.