Data-Driven Prediction of SAnMBR Performance under Phenolic Shock Loads: A Comparative Machine Learning Study

Abstract Treating phenolic wastewater remains a major challenge due to the toxicity and inhibitory effects of compounds such as 2,4-dichlorophenol (2,4-DCP), which destabilize biological processes. This study investigates the performance of a pilot-scale submerged anaerobic membrane bioreactor(SAnMBR) treating 2,4-DCP-contaminated wastewater over 189 days and develops a data-driven machine learning framework to predict COD removal efficiency. Based on 189 experimental observations and six key operational parameters, three algorithms, namely multiple linear regression, artificial neural networks, and support vector regression were trained. SVR outperformed the other models, achieving a test R² of 0.952, an RMSE of 2.96, and strong generalization performance. The results indicate that optimal COD removal occurred within a narrow pH range (8-9) and at low 2,4-DCP loads ( 150 mg/L). In addition to COD, SAnMBRs also achieved high removal efficiencies for turbidity (> 75%), TSS (> 90%), and 2,4-DCP (> 80%). These findings demonstrate that integrating long-term experimental data with interpretable machine learning enables reliable performance prediction and provides actionable insights for optimizing the SAnMBR operation in treating toxic industrial wastewater. Graphical Abstract This graphical abstract presents a structured visual pathway linking long-term experimental operation with data-driven predictive modelling for a submerged anaerobic membrane bioreactor (SAnMBR) treating phenolic wastewater. The first segment illustrates the pilot-scale system, highlighting the continuous operation under progressively increasing 2,4-dichlorophenol (2,4-DCP) shock loads over 189 days. Key operational variables, including influent COD, organic loading rate, turbidity, total suspended solids, pH, and 2,4-DCP concentration, are introduced as the primary data inputs. The second segment focuses on the analytical and modelling framework, where experimental data are processed and fed into three machine learning models (MLR, ANN, and SVR). The model comparison is visually emphasized to demonstrate the superiority of the SVR model in capturing non-linear system behaviour. The final segment summarizes the key outcomes by linking model predictions with process understanding. It highlights the existence of a narrow optimal operational window governed by pH and 2,4-DCP concentration, where maximum COD removal efficiency is achieved. Additionally, it distinguishes between biological performance drivers and operational constraints, particularly membrane fouling associated with suspended solids. Overall, the graphical abstract provides a concise integration of experimental operation, machine learning modelling, and process interpretation, enabling rapid understanding of how data-driven approaches can enhance the prediction, optimization, and control of SAnMBR systems treating toxic industrial wastewater.