Artificial neural network model for desalination by sweeping gas membrane distillation M. Khayet ⁎, C. Cojocaru Department of Applied Physics I, Faculty of Physics, University Complutense of Madrid, Avda. Complutense s/n 28040, Madrid, Spain abstract article info Article history: Received 10 February 2012 Received in revised form 22 June 2012 Accepted 24 June 2012 Available online 17 July 2012 Keywords: Sweeping gas membrane distillation Desalination Artificial neural network Optimization Sweeping gas membrane distillation process (SGMD) has been used for desalination and its performance index, defined as the product of the distillate flux and the salt rejection factor, has been modeled using artificial neural network (ANN) methodology. A feed-forward ANN has been developed for prediction of the performance index based on a set of 53 different experimental SGMD tests. A feed solution of 30 g/L sodium chloride was used in all experiments and the salt rejection factors were found to be greater than 99.4%. The individual and interaction effects of the input variables, namely the feed inlet temperature, the feed flow rate or the feed circulation veloc- ity, and the air flow rate or the air circulation velocity, on the SGMD performance index have been investigated. The optimum point was determined by means of Monte Carlo simulation. The obtained optimal conditions were a feed inlet temperature of 69 °C, an air flow rate of 34.5 L/min (i.e. 2.02 m/s air circulation velocity) and a feed flow rate of 160 L/h (i.e. 0.155 m/s liquid circulation velocity). Under these operating conditions a performance index of 1.493×10 -3 kg/m 2 .s has been achieved experimentally being the maximal SGMD performance index obtained inside the region of experimentation. © 2012 Elsevier B.V. All rights reserved. 1. Introduction During the last years artificial neural network (ANN) modeling was used frequently in various separation and technological applications, mainly due to their powerfulness for solving complex multiple regres- sion problems [1]. The ability of ANN for mapping non-linear relation- ships between the inputs and outputs of a system or a process has extended the field of applications of this modeling tool. Several studies have been considered for the application of ANN in modeling of various processes in membrane technology [1–10].A feed-forward ANN was developed by Abbas and Al-Bastaki [1] for the prediction of a reverse osmosis (RO) performance using a FilmTec SW30 membrane for desalination of various salt solutions ranging between brackish water and seawater salinities. Purkait and collabora- tors [2] employed two ANN models for prediction of the permeate flux when treating a leather plant effluent by nanofiltration (NF) process followed by reverse osmosis (RO). Zhao and co-workers [3] have per- formed a comparison between a modified solution diffusion model and ANN to predict RO/NF water quality effluent. Yangali-Quintanilla et al. [4] also used ANN to predict the rejection of neutral organic compounds by NF and RO using polyamide membranes. Libotean and collaborators [5] proposed an ANN with back-propagation to forecast the performance of an RO plant and for potential use in operational diagnostics. Al-Abri and Hilal [6] developed an ANN model for simula- tion of a combined humic substance coagulation and membrane filtration. Huaiqun and Kim [7] used a radial basis function neural net- work approach for prediction of permeate flux decline in crossflow membrane filtration of a colloidal suspension. Darwish, et al. [8] used ANN for simulation of NF of sodium chloride and magnesium chloride solutions. Sahoo and Ray [9] worked on the prediction of permeate flux decline in crossflow membranes using ANN and genetic algorithms. Mhurchú and Foley [10] employed the dead-end filtration of yeast suspensions by correlating specific resistance and permeate flux data using artificial neural networks. In our previous paper [11] we compared the ANN model with response surface model (RSM) in terms of prediction and optimization of desalination by RO process. The ANN model was found to be more adequate in prediction of the RO performance index than the RSM em- pirical model in a wide range of salt concentrations. It is worth quoting that the application of neural network modeling in membrane distillation (MD) is very limited. MD is a thermally driven process mainly dealing with water vapor transport through non-wetted porous hydrophobic membranes [12]. This process demonstrates to be successfully applied in desalination of seawater or brackish waters. Various MD configurations can be considered to apply the driving force, which is the transmembrane vapor pressure, i.e. direct contact membrane distillation (DCMD), air gap membrane distillation (AGMD), vacuum membrane distillation (VMD) and sweeping gas membrane distillation (SGMD) [12]. Recently we have reported on the development and application of an ANN model to predict the AGMD desalination performance index [13]. The ANN model was used for optimization of the AGMD process and the following optimal conditions were obtained, an air gap Desalination 308 (2013) 102–110 ⁎ Corresponding author. Tel.: +34 91 3945185; fax: +34 91 3945191. E-mail address: khayetm@fis.ucm.es (M. Khayet). 0011-9164/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.desal.2012.06.023 Contents lists available at SciVerse ScienceDirect Desalination journal homepage: www.elsevier.com/locate/desal