A comparison between semi-theoretical and empirical modeling of cross-flow microfiltration using ANN

• Published in: Desalination 2011-08, Vol.277 (1), p.348-355
• Main Authors: Ghandehari, Sara, Montazer-Rahmati, Mohammad Mehdi, Asghari, Morteza
• Format: Article
• Language: English
• Subjects: algorithms; ambient temperature; Applied sciences; Artificial neural networks; Bovine serum albumin; cellulose; Chemical engineering; Classic mechanisms of fouling; Cross-flow microfiltration; desalination; Exact sciences and technology; Filtration; Flux; fouling; hydrophobicity; Learning theory; Mathematical models; Membrane separation (reverse osmosis, dialysis...); Membranes; Microfiltration; Neural networks; Pollution; Porosity; prediction
• ISSN: 0011-9164; 1873-4464
• DOI: 10.1016/j.desal.2011.04.057

The applicability of semi-empirical and artificial neural network (ANN) modeling techniques for predicting the characteristics of a microfiltration system was assessed. Flux decline under various operating parameters in cross-flow microfiltration of BSA (bovine serum albumin) was measured. Two hydrophobic membranes were used: PES (polyethersulfone) and MCE (mixed cellulose ester) with average pore diameters of 0.22 μm and 0.45 μm, respectively. The experiments were carried out to investigate the effect of protein solution concentration and pH, trans-membrane pressure (TMP), cross-flow velocity (CFV), and membrane pore size on the trend of flux decline and membrane rejection at constant trans-membrane pressure and ambient temperature. Subsequently, the experimental flux data were modeled using both classical pore blocking and feed forward ANN models. Semi-empirical models based on classic mechanisms of fouling have been proposed. It was shown that these mechanisms could predict the microfiltration flux for a specified period of processing time; while through appropriate selection of ANN parameters such as the network structure and training algorithm, the ANN-based models are competent in modeling membrane filtration systems for all operating conditions and the entire filtration time with desired accuracy. ► The Intermediate blocking model predicts the flux decline in the experimental range. ► The standard blocking model can only predict the flux decline at the beginning of filtration. ► The Levenberg–Marquart training algorithm yields the lowest MSE and the highest R 2. ► An ANN with a 5:6:8:1 structure accurately predicts the filtration flux. ► An ANN with a 4:6:5:1 structure accurately predicts the membrane rejection.