Machine learning-based prediction of biofouling and performance decline in reverse osmosis membranes treating urban wastewater

ABSTRACT

Flowchart illustrating a machine learning-based framework for predicting outcomes in a water-related system. The diagram shows input data sources (such as environmental or operational variables) feeding into data preprocessing steps, followed by model development using machine learning algorithms. The trained model generates predictions, which are evaluated and applied for decision-making or system optimization. Arrows indicate the sequential workflow from data input to final prediction and application.

Biofouling in reverse osmosis (RO) membranes treating urban wastewater progressively reduces permeate flux, deteriorates salt rejection, and increases operational costs. To address this challenge, this study presents a novel machine learning-based surrogate modelling framework to predict RO membrane performance under non-disinfected wastewater conditions – an underexplored scenario with high environmental and industrial relevance, where biofouling is the primary performance driver. Using experimental data from a pilot-scale RO system, three algorithms were evaluated: Random Forest, Gradient Boosting, and Support Vector Regression. Gradient Boosting achieved the best results for flux prediction (R2 = 0.97, MAE = 0.56 L/m2·h), while Random Forest performed best for salt rejection (R2 = 0.65, MAE = 0.105). Operating time, temperature, and COD were identified as the most influential variables. The proposed models offer a practical tool for proactive maintenance and real-time monitoring in water reuse facilities, contributing to the sustainability of wastewater treatment processes.