Prediction of Nanofiltration and Reverse-Osmosis-Membrane Rejection of Organic Compounds Using Random Forest Model

Abstract

Data-driven membrane prediction models including multiple linear regression and neural network have been widely applied to analyze reveal rejection interactions between various compounds and membranes. While the random forest algorithm is an ensemble learning method widely applied in science, engineering, and many other fields, few studies apply a random forest model for predicting membrane rejection. Thus, this study proposes a random forest model for predicting nanofiltration/reverse-osmosis-membrane rejection of emerging organic contaminants. The original membrane rejection dataset was collected from multiple studies and included 701 points of 84 organic compounds. This study (1) examined prediction performance between random forest and neural network models; (2) compared identified important features from random forest feature importance, principal component analysis, and Pearson correlation coefficient; and (3) analyzed hyper-parameter tuning process and results between random forest and neural network. The findings of this study suggest that random forest is feasible for modeling membrane rejection prediction (determination coefficient≥0.9). The integrated function of feature importance can reliably identify important features, and the random forest model required less efforts for turning parameters than the neural network model.