Application of quantitative structure–property relationships (QSPRs) to predict the rejection of organic solutes by nanofiltration

Abstract

Recent interest in quantitative structure property/activity relationship (QSPR/QSAR) models to predict the removal of organic contaminants by membrane processes has highlighted the need to develop models applicable to different operating conditions, such as flux. In this study, two types of QSPR models were developed to predict removal of nonionic organic compounds by a nanofiltration membrane (NF270) including: (1) QSPR models with flux as an independent (or predictor) variable; and (2) QSPR models to predict fitting parameters of a fundamental model (i.e., Spiegler–Kedem model). Rejection data for 67 nonionic organic compounds and an NF membrane at five solvent fluxes (10–120Lm−2h−1) were generated and used for model development and validation. In order to select the best statistical method to develop QSPRs, several models were developed using multiple linear regression (MLR), partial least squares (PLS), and artificial neural network (ANN) with molecular descriptors selected by different methods (i.e., filtering, hybrid, and wrapper methods). The most effective linear and nonlinear models were developed using PLS and ANN with molecular descriptors selected by variable importance plot and feed forward selection. The rejection of nonionic compounds with minimal solute-membrane affinity could be described by permeate flux, and a compound’s molecular depth and diffusion coefficient. Fitting parameters of Spiegler–Kedem model (reflection coefficient and permeability coefficient) could be best described by size parameters and diffusivity. These results indicate that the rejection of the majority of the nonionic organic compounds evaluated was mainly due to size exclusion. Generally, both QSPR methods developed during this study were found to be effective methods for predicting solute rejection by the nanofiltration membrane, and nonlinear models provided better fits (based on statistical parameters) of both training and validation rejection datasets.