Computational simulation of molecular separation in liquid phase using membrane systems: Combination of computational fluid dynamics and machine learning

A new methodology was developed in this work based on combination of machine learning and computational fluid dynamics (CFD) for hybrid modeling of membrane separation processes. The developed methodology was tested for a separation case in removing organic molecules from aqueous solution using a polymeric membrane in cylindrical shape. The CFD simulation was carried out via finite element technique. The results of the CFD outputs were inputted several machine learning techniques to build the hybrid model of the process. The data of mass transfer in form of concentration was extracted which are more than 2000 data points. The dataset includes two inputs r and z, in addition to one output (C). A number of different models are selected for this purpose, including Random Forests, Gradient Boosting models and Extremely Randomized Trees (Extra Trees). During the course of this research, it was confirmed that these models are capable of performing very well in this dataset when their hyper-parameters are optimally chosen. The R2 parameter for all three methods was higher than 0.99, indicating that the performance was very favorable. It is noteworthy that the RMSE metric decreased to 2.43 and the MAE metric has also decreased to 1.06 by optimization.