Hybrid mechanistic approach in the estimation of flow properties in cylindrical membrane modules

Abstract

A hybrid computational approach was employed for simulation of molecular separation using polymeric membranes. The considered system is a cylindrical membrane module in which the mass transfer equations were solved numerically using CFD (Computational Fluid Dynamics) to obtain the concentration of the species, and then the simulation results were used in machine learning models. Indeed, the CFD simulation results were used as the inputs for several machine learning models to obtain the hybrid model. We have a dataset with more than 2000 data points and two input features (r and z). Also, the only output is C which is the concentration of the species in the feed channel of membrane module. KNN (K nearest neighbor), PLSR (Partial Least Square Regression), and SGD (Stochastic Gradient Descent) are the models employed in this research to analyze the mentioned data set. Models were optimized with their hyper-parameters and finally evaluated with different statistical metrics. MAE error metric is 3.4, 5.1, and 5.5 for KNN, SGD, and PLSR. Also, they have 0.998, 0.997, 0.896 coefficient of determination (R2) respectively. Finally, based on the overall results, KNN with K = 8 is selected as the best model in this study for simulation of the membrane system. The final maximum error is also 1.35E+02.

Keywords

• CFD
• Computational simulation
• Fluid flow
• Separation
• Machine learning
• Environmental