Artificial neural network models based on QSAR for predicting rejection of neutral organic compounds by polyamide nanofiltration and reverse osmosis membranes

Prediction of the rejection of neutral organic compounds by polyamide nanofiltration (NF) and reverse osmosis (RO) membranes was tested in this study, using artificial neural network (ANN) models. The ANN models were based on a quantitative structure-activity relationship (QSAR) equation that defined an appropriate set of solute and membrane variables able to represent and describe rejection. The QSAR equation was defined using principal component analysis and multiple linear regression. The QSAR model combines interactions between solute properties, membrane characteristics and operating conditions. Solute size descriptors (molecular length and equivalent width) and the hydrophobicity descriptor represented by log K-ow, were the main physicochemical properties of the neutral organic compounds able to describe rejection. Regarding the membrane characteristics, it was found that salt rejection (in combination with the aforementioned solute properties) can be practically applied to predict organic solute rejection. An experimental database produced by the authors in combination with data collected from literature comprising a total of 161 rejections of 50 organic neutral compounds (including pharmaceuticals, endocrine disrupting compounds, pesticides, alcohols, phenols and solvents) by six NF and nine RO membranes was used to produce the ANN models. Sixty percent of the data were used to generate the model, and validation was performed with 20% of the total data. Independent rejection predictions were calculated for the remaining 20% of the data. For the most promising ANN models, the independent predicted rejection values were compared to measured rejections and good correlations were found (R-2 = 0.97) and predicted rejections presented and standard deviation of error of ca. 5%. (C) 2009 Elsevier B.V. All rights reserved.