The prediction of organic solvent reverse osmosis (OSRO) performance is challenging because of the numerous variables influencing the process. Therefore, applying a machine learning model, such as an artificial neural network (ANN), instead of physics-based modeling is helpful because it can consider many variables and learn patterns to generate a prediction. However, limited experimental data, along with the numerous variables that influence the OSRO process, considerably hamper the proper training of an ANN. Moreover, the affinities between membranes and solvents and between solvents play an important role in OSRO performance. Thus, new variables should be considered for appropriately training the ANN. However, this consideration further increases the number of variables. Herein, we proposed new variables that considered all three components of the Hansen solubility parameters for the membranes and the involved solvents to represent their interactions. Principal component analysis (PCA) was performed to simplify the information from multiple variables and extract only the main information to evaluate the importance of the input variables. The results from PCA were used as fundamental inputs in constructing the ANN model. Additionally, some OSRO-relevant variables, including applied pressure, solvent fraction, and solubility parameters, were used to train the multilayer perceptron ANN, and its structure was optimized using a genetic algorithm. To the best of our knowledge, this is the first time that the OSRO performance is predicted using machine learning. The average absolute relative deviation values between the model-calculated and experimental data were 6.81% and 1.84% for flux and rejection, respectively. Our findings demonstrated that the proposed new variables significantly improved the OSRO predictions.

Keywords