Predicting and optimizing forward osmosis membrane operation using machine learning

Forward osmosis (FO) utilizes a draw solution to transport water across a semipermeable membrane, offering energy-efficient water treatment and resource recovery. This study explores machine learning models to predict FO performance at pilot scale, overcoming the limitations of traditional mathematical models in terms of computational load and time. By analyzing data from FO pilot experiments, we compare various algorithms, including multiple linear regression (MLR), support vector machine (SVM), k-nearest neighbor (KNN), decision tree, and artificial neural networks (ANNs). Among these, ANN was evaluated as most suitable and further optimized with input features of the permeate flux, membrane area, feed and draw solution flow rates, and feed and draw solution concentrations. The optimized ANN model demonstrated high accuracy for water flux prediction, with R² values of 0.9886 and RMSE values of 0.3498 Lm⁻² h⁻¹. Additionally, an ANN model is developed to predict operating pressures under various FO operation conditions. By integrating FO water flux and operating pressure predictions, our model identifies optimal operating conditions that balance specific energy consumption and water recovery. Our findings offer insights and practical guidance for process engineers to efficiently design and operate FO systems to minimize energy consumption and maximize recovery.