Development of hybrid mechanistic-artificial intelligence computational technique for separation of organic molecules from water in polymeric membranes

Abstract

We have investigated the application of machine learning model for description of a physical process in separation of liquids. The considered process is a membrane system for removal of an organic component. Multiple machine learning models were taken into account, while the inputs are the results of a simulation conducted by computational fluid dynamics (CFD). Indeed, the coordinates (r and z) were used as the inputs to the models, and the models predicted the single output which was the content of organic component in the feed section of membrane (C). In the generated dataset, there are more than 8K rows of data included. Each row has two inputs (r and z) and one output (C). Among the models selected for this study are the Decision Tree, the Poisson regression, and the K-nearest neighbor. The hyperparameters were tuned using the grid-search method, and the final models were obtained. By R2-score criteria, Decision Tree, Poisson regression, and K-nearest neighbors had scored 0.9969, 0.9797, and 0.9841, respectively. According to this and other visual and numerical comparisons, Decision Tree is the most accurate model of this study. The maximum error by this model is 5.5 × 102 and its RMSE error rate is 4.829 × 101.