PVDF Ultrafiltration

Jun 7, 2017 - ACS Combinatorial Science. Rajarathinam, Kumaravel, and Vasuki. 2017 19 (7), pp 455–463. Abstract: Synthesis of the 2-aza analogues of...
0 downloads 0 Views 3MB Size
Research Article pubs.acs.org/acscombsci

Modeling and Optimization of NLDH/PVDF Ultrafiltration Nanocomposite Membrane Using Artificial Neural Network-Genetic Algorithm Hybrid Samira Arefi-Oskoui,† Alireza Khataee,*,† and Vahid Vatanpour‡ †

Research Laboratory of Advanced Water and Wastewater Treatment Processes, Department of Applied Chemistry, Faculty of Chemistry, University of Tabriz, 51666-16471 Tabriz, Iran ‡ Faculty of Chemistry, Kharazmi University, 15719-14911 Tehran, Iran ABSTRACT: In this research, MgAl-CO32− nanolayered double hydroxide (NLDH) was synthesized through a facile coprecipitation method, followed by a hydrothermal treatment. The prepared NLDHs were used as a hydrophilic nanofiller for improving the performance of the PVDF-based ultrafiltration membranes. The main objective of this research was to obtain the optimized formula of NLDH/PVDF nanocomposite membrane presenting the best performance using computational techniques as a cost-effective method. For this aim, an artificial neural network (ANN) model was developed for modeling and expressing the relationship between the performance of the nanocomposite membrane (pure water flux, protein flux and flux recovery ratio) and the affecting parameters including the NLDH, PVP 29000 and polymer concentrations. The effects of the mentioned parameters and the interaction between the parameters were investigated using the contour plot predicted with the developed model. Scanning electron microscopy (SEM), atomic force microscopy (AFM), and water contact angle techniques were applied to characterize the nanocomposite membranes and to interpret the predictions of the ANN model. The developed ANN model was introduced to genetic algorithm (GA) as a bioinspired optimizer to determine the optimum values of input parameters leading to high pure water flux, protein flux, and flux recovery ratio. The optimum values for NLDH, PVP 29000 and the PVDF concentration were determined to be 0.54, 1, and 18 wt %, respectively. The performance of the nanocomposite membrane prepared using the optimum values proposed by GA was investigated experimentally, in which the results were in good agreement with the values predicted by ANN model with error lower than 6%. This good agreement confirmed that the nanocomposite membranes prformance could be successfully modeled and optimized by ANN-GA system. KEYWORDS: ultrafiltration, nanocomposite, Mg−Al nanolayered double hydroxide, artificial neural network, genetic algorithm



INTRODUCTION In recent years, ultrafiltration (UF) technology has attracted a great deal of attention in the field of food industry, biotechnology, water and wastewater treatment, and medical industry.1 The membranes used in this technology, which are able to retain the solutes with diameter of 1−100 nm, can be mainly classified to two groups including ceramic membranes and polymeric membranes.2,3 Polymeric membranes provide a lower total operating cost compared with ceramic membranes, so polymeric membranes make a large part of the commercial ultrafiltration membranes.4 Among the different commercial polymers which can be used for fabrication of UF membranes, polyvinylidene fluoride (PVDF) is of great importance owning to its remarkable properties such as high mechanical strength, excellent chemical resistance and thermal stability.1,5 However, low hydrophilicity and poor wettability of the PVDF membranes result in low permeability and high fouling, and limit the widespread industrial application of these membranes.6 One of the most common methods for overcoming this limitation is to using additives in the casting solution of membrane.7,8 Number of additives such as polymeric additives, © 2017 American Chemical Society

which are known as high molecular weight additives and the hydrophilic nanostructured inorganic materials can improve the membrane performance by improving the morphology of the membrane and enhancing the hydrophilicity of the membrane.5 Polyvinylpyrrolidone (PVP) is a known hydrophilic polymeric additive to produce highly porous membranes.5,9 Hydrophilic nanostructured inorganic and organic materials, for example, TiO2, CuO, SiO2, WS2, and graphene oxide have also received attention as additives to improve the performance of the polymeric membranes in recent years.10−17 Nanolayered double hydroxides (NLDH) with two-dimensional structure are a main group of the anionic clays. These materials can be represented by the general chemical formula of [M2+1−xM3+x(OH)2]x+[An−x/n·mH2O]x−, where the M2+ and M3+ shows divalent and trivalent metal cations, respectively. An− and m represents the anions located in the interlayer space of NLDH, and number of the water molecules in the interlayer Received: March 11, 2017 Revised: May 19, 2017 Published: June 7, 2017 464

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science gallery, respectively.18,19 The ratio of M3+/(M2+ + M3+) determines the value of the x in this formula. Thermal stability, low cytotoxicity, high surface area and especially numerous hydroxyl groups on the layers which results in high hydrophilicity, make these materials appropriate additive for improving the performance of polymeric membranes.18 The membrane characteristics including pore morphology, pore size, porosity, surface roughness and hydrophilicity, which significantly affect the performance of the membranes (e.g., permeability and fouling resistance), are mainly controlled by the membrane fabrication conditions and the composition of the membrane. Therefore, development of a model, which can simulate the filtration behavior of the membrane and represent the influence of the affecting parameters on the performance of the membrane is of great importance for economical process design and scaling-up the membrane system.20,21 Artificial neural networks (ANNs) which are based on the structure and functions of biological neural networks have received great attention in the field of the computing systems.22,23 These systems are known as efficient instruments for modeling nonlinear systems. Moreover, to reach an optimum membrane, the affecting parameters should be recognized and optimized. It should be noted that the experimentally optimization of membrane formulation is a time-consuming and expensive procedure. On this basis, computer-based techniques, such as genetic algorithm (GA), have recently caught increasing attention owning to their capability to analyze and optimize the complex problems. The literature review shows that the GA has been successfully used to solve multiobjective optimization problems.24−26 The main objective of this study was to obtain the optimized formula of NLDH/PVDF nanocomposite membrane with the best performance using computational techniques as a costeffective method. For this aim, MgAl-CO32− NLDH was synthesized using a facile coprecipitation method followed by a hydrothermal treatment. The prepared NLDHs were used as hydrophilic nanofiller for improving the performance of the PVDF-based ultrafiltration membranes. An artificial neural network model was developed for modeling the fabrication process of the nanocomposite membrane. Finally, the developed ANN model was introduced to genetic algorithm as a bioinspired optimizer to determine the optimum values of input parameters leading to high pure water flux, protein flux, and flux recovery ratio.

autoclave with a Teflon lining. Finally, the Mg−Al NLDH crystallites were collected and dried at 50 °C. Preparation of NLDH/PVDF Nanocomposite Membranes. The flat sheet NLDH/PVDF membranes were fabricated using the simple phase inversion method. The polymeric dope solutions were prepared by dispersing appropriate amounts of Mg−Al NLDH in NMP as solvent under ultrasonic irradiation for 30 min, and subsequently appropriate amounts of PVP 29000 as pore former and PVDF were dissolved. The obtained suspensions were stirred magnetically at 50 °C for 24 h to form homogeneous dope solutions. The resultant dope solutions were kept in oven at 50 °C for 5 h to remove the air bubbles. Then, the prepared degassed dope solutions were casted on the polyester nonwoven fabrics using a casting knife with thickness of 150 μm. The cast membranes, immediately after casting, were merged in a bath containing tap water at room temperature for immersion precipitation. The prepared membranes were stored in distilled water until using them. Pure Water Permeation, BSA Filtration, Antifouling Performance, and Rejection of the Membranes. In this research, dead-end filtration cells with an effective area of 19.6 cm2 were applied to determine the permeation of pure water and BSA solutions and also to determine the rejection parameter of membranes. To reach a steady permeation flux and to reduce the compaction effects, the membranes were precompacted by passing pure water through the membranes under 0.3 MPa for 30 min. The permeation flux of the membranes were determined at operating pressure of 0.2 MPa using eq 1

J=

V At

(1)

where J shows the permeate flux (L/m h), V stands for permeate volume (L), and the A and t represent the membrane surface area (m2) and time (h), respectively. In this study, the average of at least three membranes were calculated and reported. The flux recovery ratio (FRR) of the membranes was determined to study the antifouling property of the prepared membranes. For this purpose, after filtering the pure water for 90 min and determining the pure water flux (Jw,1) according to eq 1, the solution of BSA as foulant model with the concentration of 0.5 g/L was replaced with pure water. The BSA solution filtered at operating pressure of 0.2 MPa for 90 min, and the flux of BSA solution (Jp) was determined using eq 1. Then, the fouled membranes were brought out from cells, and the surface of the fouled membranes were cleaned via washing with distilled water. The membranes remained in distilled water for 20 min. Then, the pure water was filtered at operating pressure of 0.2 MPa for 90 min, and the pure water flux (Jw,2) was determined using eq 1. The FRR parameter was determined using eq 2. 2



EXPERIMENTAL PROCEDURES Chemicals. N-methyl-2-pyrrolidone (NMP), AlCl3, MgCl2· 6H2O, Na2CO3, and NaOH were purchased from Merck Co., Germany. PVDF polymer was purchased from Alfa Aesar, Germany. Bovine serum albumin (BSA) and polyvinylpyrrolidone (PVP, MW = 29 000 g/mol) were purchased from SigmaAldrich Co. Synthesis of Mg−Al NLDH. Mg−Al NLDH was prepared through facile coprecipitation-hydrothermal method. For this purpose, 70 mL of solution containing AlCl3 and MgCl2·6H2O with the concentrations of 0.1 and 0.2 mol/L, respectively, were added into solution containing NaOH (0.15 mol/L) and Na2CO3 (0.013 mol/L) under vigorous stirring. The resultant mixture was stirred for 30 min, and the slurry of the NLDH was then centrifuged and washed three times with distilled water. In the next step, the obtained precipitate was redispersed in 280 mL of distilled water, and the suspension was treated hydrothermally at 100 °C for 16 h using stainless steel

FRR (%) =

Jw,2 Jw,1

× 100 (2)

The protein rejection parameter of the fabricated membranes was calculated using eq 3 ⎛ Cp ⎞ R (%) = ⎜1 − ⎟ × 100 Cf ⎠ ⎝ 465

(3) DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science where Cp represents the concentration of protein in permeate (g/L) and Cf represents the concentration of protein in feed (g/L). In this research, the Bradford method was applied for determining the concentration of protein in permeate and feed. Characterization Analysis. The crystallite structure of the synthesized Mg−Al NLDH was investigated applying X-ray diffraction D5000 diffractometer (Germany), with monochromatic high intensity Cu Kα radiation (l = 1.5406 Å), an accelerating voltage of 40 kV and an emission current of 40 mA. The width size and the morphology of the prepared Mg−Al NLDH and, also, the morphology of the surface and the crosssection of the fabricated membranes were investigated applying a Tescan SEM Model Vega (Czech Republic). The energy dispersive X-ray (EDX) analysis was performed employing INCA (England) instrument to study the distribution of the Mg−Al NLDH in the matrix of the nanocomposite membranes. The hydrophilicity property of the membrane surface was investigated by determining the contact angle parameter. For this purpose, the static contact angle between the distilled water drop and the membrane surface was determined. To minimize the experimental error in this analysis, the average of the four contact angle measurements was reported. A Nanosurf Mobile S (version 1.8) scanning probe-optical microscope (Switzerland) was applied to determine the roughness parameter of the membrane surface and to prepare the three-dimensional image of them. The average roughness (Sa), the mean difference between the highest peaks and the lowest valleys (Sy), and the root-mean-square of the Z data (Sq) were determined to investigate the surface properties of prepared membranes.

hidden layer. In the next step, more neurons were added to the hidden layer and the model was trained. The addition of the neurons was continued until reaching to the best model without overfitting. The standard deviation of error value was calculated to determine the statistical fitness of various neural networks. In this study, hyperbolic tangent sigmoid (tan-sigmoid, eq 4) and linear function (eq 5) were applied as the activation function for hidden and output layers, respectively. f (x ) =

1 1 + exp( −x)

(4)

f (x ) = x

(5)

In the training process, to optimize the weights of the neurons the gradient descend algorithm (traingdm), which is a kind of Standard back-propagation (BP) algorithm, was applied as training algorithm to minimize the error function. Moreover, the relative importance of the input parameters including polymer concentration, pore former concentration and the NLDH concentration was determined using eq 6.28 The value of the relative importance determine the effect of input parameters on the responses. n

Ss =

m

∑ j (wsjujk /∑i wij) m

n

m

∑i ∑ j (wijujk /∑i wij)

(6)

where Ss represents the relative importance of the parameter s. wij shows the weight connecting the parameter i (input parameter) to the neuron j in the hidden layer. ujk shows the weight connecting the neuron j of the hidden layer to the neuron k in the output layer. m, n, and k represent the neurons in the input layer, hidden layer and output layer, respectively. Before training the neural network, all the data were normalized in the range of −1 to +1 using eq 7.



MODELING AND OPTIMIZATION Artificial Neural Network. In this research, a multilayer perceptron (MLP) neural network was developed for modeling the ultrafiltration membrane preparation process. The MLP neural network is one of the feed forward type networks, in which the signals are transmitted only from a lower layer to higher layer. The MLP neural network is composed of number of layers including an input layer, one or more hidden layer between the input and output layers and an output layer. It should be noted that there is not any certain rule in order to determine the number of the hidden layers in the MLP neural networks. However, in theory, a MLP with one hidden layer can model any process that MLP with two or more layers can model.27 In light of this, a MLP neural network with one hidden layer applied for developing the model. Each layer contains one or more neurons. The number of the neurons in the input and output layers equal to the number of the parameters affecting the process (independent variables) and the number of the responses (dependent variables), respectively. In this research, the input layer consisted of three neurons including the concentration of the PVDF polymer, concentration of PVP 29 000 which was used as pore former and the concentration of the NLDH. There were three neurons in the output layer including the pure water flux, BSA protein solution flux and the flux recovery ratio. The number of the neurons in the hidden layer should be optimized to obtain the best model simulating the experimental data. It is obvious that a neural network with a few number of neurons in the hidden layer cannot model the process in good way. However, a neural network with a lot of neurons in the hidden layer may result in overfitting. Therefore, as aforementioned, optimum number of neurons should be used in the hidden layer. For this purpose, at first step, the model was trained with a few neurons in the

xi =

2Xi −1 X max − X min

(7)

where Xi is the real value of the parameter i, xi represents the dimensionless coded value of the parameter i, and Xmin and Xmax are the lowest and the highest value of the parameter i. Genetic Algorithm. Genetic algorithm is a class of evolutionary algorithms, which can be used successfully to solve the complex optimization problems via simulating the biological evolution.29,30 The algorithm of this method consists of (a) coding the possible solutions, (b) randomly creating the initial population, (c) rating the possible solutions applying evaluation function according to their fitness, (d) changing the compositions of the gens through crossover and mutation during the population reproduction, and (e) determining the optimized parameters. In this study, the resultant ANN model was introduced to GA to optimize the parameters affecting the fabrication process of PVDF/NLDH nanocomposite ultrafiltration membrane. The GA was used for determining the optimum values of the affecting parameters including polymer concentration, NLDH concentration and the pore former concentration to obtain the maximum values for pure water flux, protein solution flux, and flux recovery ratio.



RESULTS AND DISCUSSION Characterization of the Synthesized Mg−Al NLDH. As shown in Figure 1, the crystallite structure of the prepared 466

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 1. XRD pattern of Mg−Al NLDH.

Figure 2. SEM image of Mg−AL NLDH.

in Figure 2, homogeneous nanosheets with average diameter of 41.2 nm are formed. Topology Selection of the ANN. Different network topologies with different numbers of neurons in hidden layer, were investigated to determine the best neural network, which can model the process in the best way. In the first step, a network with two neurons in the hidden layer was selected, and then the number of neurons in hidden layer was increased step by step. The network standard deviation of the error (STD) versus the number of the neurons in hidden layer is represented

MgAl-NLDH was studied using XRD analysis. The characteristic peaks of the (003), (006), (009), (015), (018), (110), and (113) centered at 2θ of 11.4°, 23°, 35.12°, 38.7°, 45.5°, 60.5°, and 62°, respectively, represent the typical NLDH structure of the prepared MgAl-NLDH.31,32 The average crystallite size of the synthesized sample was determined to be 14.6 nm according to the sharpest peak and using Debye−Sherrer’s equation.33 The SEM analysis was applied to study the morphology of the synthesized MgAl-NLDH. As can be seen 467

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 3. Standard deviation of the error (STD) versus number of neurons in the hidden layer.

Figure 4. Model prediction versus experimental values for optimum topology for (a) training, (b) validation, (c) test, and (d) all data (optimum topology = 3, 5, 3).

in Figure 3. The networks with low number of neurons in the hidden layer have large STD, indicating that these networks are not able to properly model the relationship between the input and output parameters. On the other hand, the risk of the overfitting increases for the networks with too many neurons in the hidden layer. Therefore, the network with sufficiently low standard deviation of the error is the best network for modeling. As represented in Figure 3, the STD decreases by enhancing the number of neurons in hidden layer. However,

after reaching a certain number of neurons in hidden layer, the STD does not decrease significantly by addition of more neurons. Furthermore, after this number of neurons, not only the STD does not decrease, but also the risk of the overfitting increases. In this research, the optimum number of neurons in hidden layer is 5 which results in the best model. Validation of the Model. To develop the model, the experimental data were randomly divided into three sets including training data (70%), test data (15%), and the 468

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science validation data (15%). The training data were used for learning the network, and the test data were applied to validate the predictability of the model. In addition, the overfitting of the model was tested using validation data. The predicted data versus experimental data for training, test, validation and all data are represented in Figure 4a−d. As can be observed in Figure 4a, the training data are distributed well around the line y = x with high R2 value (R2 = 0.981), indicating that the developed model with the selected topology is successfully able to predict the data. The high R2 value of the testing data (R2 = 0.986) confirms that the developed model is able to predict the responses (including the pure water flux, BSA protein solution concentration and flux recovery ratio) for the conditions that are not used in training process (see Figure 4b). In addition, the high R2 value (R2 = 0.990) for the validation data (Figure 4c) proves that the developed model does not overfit the data used for training. Effect of the Parameters on the NLDH-PVDF Nanocomposite Performance. Figure 5 represents the predicted pure water flux (L/m2 h), protein flux (L/m2 h), and the flux recovery ratio (FRR) versus affecting parameters in 3D plots. As can be seen in Figure 5a, the pure water flux decreases significantly by increasing the concentration of the polymer at the range of 16−24 wt %. It should be noted that the dope solution with the PVDF concentration lower than 16 wt % is too dilute, resulting in formation of membrane with low protein rejection (lower than 90%), which is not appropriate for ultrafiltration processes. On the other hand, the dope solution with PVDF concentration higher than 24% is so viscous, and the dope solution with so high viscosity have a negligible pouring capability. Accordingly, the casting of this dope with so high viscosity solution is impossible. Therefore, in this research, the effect of the PVDF concentration was studied in the range of 16−24 wt %. The observed decrease in pure water flux by increasing the polymer concentration could be attributed to the rheologic feature improvement. Indeed, increasing the polymer concentration leads to an increment in viscosity of the dope solution, which consequently results in developing membranes with low porosity and dense structure.34 The cross-sectional and surface SEM images of the membranes prepared using different polymer concentrations with 1 wt % of PVP 29000 are shown in Figure 6a−j. From the figure, all the membranes show an asymmetric structure with a dense layer at top and a sublayer in which the finger-like pores were developed. The finger-like pores can be seen in all membranes with different concentrations of polymer, indicating that the mechanism of the membrane fabrication was not changed during the phase inversion process by changing the polymer concentration in dope solution.35 The surface SEM images of the membranes indicate that the number of the pores on the surface of the membranes decreases by increasing the polymer concentration. The pure water flux is mainly controlled by membrane surface porosity. Therefore, the observed decline in pure water flux at high concentration of polymer can be attributed to the decrease in the surface porosity. As shown in Figure 5a, the pure water flux significantly enhances by increasing the concentration of PVP 29000 as pore former. Among the high molecular weight additives, the PVP is known as most commonly used additive for fabrication of highly porous membranes. PVP is hydrophilic in nature, so the presence of the PVP in the dope solution can enhance the liquid- liquid demixing process during the precipitation, which results the formation of membrane with high porosity.5 The

Figure 5. Contour plots of (a) pure water flux−PVP 29000/polymer, (b) protein flux−PVP 29000/polymer, (c) FRR (%)−PVP 29000/ polymer, (d) pure water flux−NLDH/polymer, (e) protein flux− NLDH/polymer, (f) FRR (%)−NLDH/polymer, (g) pure water flux− NLDH/polymer, (h) protein flux−NLDH/polymer, and (i) FRR (%)−NLDH/polymer.

cross-sectional and surface SEM images of the membranes containing 18 wt % of PVDF with different concentrations of the PVP 29000 are represented in Figure 7a−l. As can be seen, the sponge-like and very short finger-like pores are formed respectively in the low and up parts of the cross-section of the 469

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 6. Cross-sectional and surface SEM images of the ultrafiltration membranes containing 1 wt % of PVP 29000 and different concentration of PVDF.

facilitating the movement of water molecules via the membrane. So the pure water flux enhancement at high concentration of the PVP 29000 can be ascribed to both hydrophilicity and the surface porosity improvement. On the other hand, the slope of the pure water increment by increasing the PVP 29000 concentration at high concentration of polymer is lower than the low polymer concentration (Figure 5a). Totally, at high concentration of polymer, the rheological feature is dominant, and favors the formation of dense structure. Therefore, increasing the PVP 29000 concentration with high concentrations of the polymer cannot significantly enhance the surface porosity of the membranes. The effect of the polymer and PVP 29000 concentrations on the flux of BSA protein is represented in Figure 5b. As can be seen, the protein flux trend is similar to pure water flux, considering the aforementioned explanations. The predicted contour plot of the FRR versus the PVP 29000 and polymer concentration is represented in Figure 5c. In this contour plot, the content of NLDH is fixed in 0.5 wt %. As can be seen, the FRR value enhances by increasing the

membrane without any pore former. Comparing the crosssectional SEM images of membranes without PVP 29000 and containing 0.5 wt % of PVP 29000, reveals that the morphology of the membrane is significantly changes by embedding PVP 29000 in dope solution, and the sponge-like pores changes to large finger-like pores (Figure 7a and Figure 7b). Therefore, the observed increment in pure water flux by adding PVP 29000 in dope solution could be assigned to the significant morphology change in the membrane. Finger-like pores can be observed in all membranes containing PVP 29000 with different concentrations. However, from the Figure 7g−l and Table 1, the number of pores in the surface of membranes and the surface porosity are significantly enhanced by increasing the concentration of the PVP 29000, resulting in the increment in pure water flux. In addition, the presence of the PVP 29000 in the matrix of the nanocomposite membrane as the hydrophilic additive can improve the hydrophilicity feature of the PVDFbased membranes (Table 1). The hydrophilicity improvement can enhance the permeability of the membranes through attracting the water molecules inside the membrane matrix and 470

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 7. Cross-sectional and surface SEM images of the ultrafiltration membranes containing 18 wt % of PVDF and different concentration of PVP 29000.

concentration of the polymer. The observed trend can be explained based on the surface roughness of the membranes. The AFM images of the membranes containing 1 wt % of PVP 29000 with different concentrations of the polymer is represented in Figure 8a−e. Some bright and dark areas can be observed in the AFM images of the membranes, representing the highest point and the pores or valleys in the surface of the membranes, respectively. The average roughness parameters of the membranes with different concentrations of the polymer were determined using Nanosurf Mobile S software in the scanning area of 8 μm × 8 μm, and the results are reported in Table 2. As can be seen, increasing the polymer concentration leads to a decrease in the average roughness, which can be attributed to the surface porosity decrement as shown in surface SEM images in Figure 6f−j. It is well-known that the fouling of the membranes with smooth surface is significantly lower compared with membranes with rough surface. This is because the pollutants accumulate in the valleys of the rough surface membranes.36 Therefore, the FRR

Table 1. Surface Roughness Parameters, Surface Porosity, and Contact Angle of Ultrafiltration Membrane Containing 18 wt.% PVDF and Different Concentrations of the PVP 29000 roughness parameters membrane 0 wt % PVP 29000 0.5 wt % PVP 29000 1 wt % PVP 29000 2 wt % PVP 29000 4 wt % PVP 29000 6 wt % PVP 29000

Sa (nm)

Sq (nm)

Sy (nm)

surface porosity (%)

contact angle (deg)

34.2

47.2

342.6

1.3 ± 0.12

77.4 ± 1.4

53.1

67.9

440.7

1.8 ± 0.09

75.1 ± 0.8

67.3

94.8

511.5

2.9 ± 0.18

72.3 ± 1.2

80.4

101.3

546.9

3.6 ± 0.11

70.6 ± 1.6

92.6

124.3

589.1

4.2 ± 0.21

68.4 ± 1.1

117.1

135.8

622.7

4.9 ± 0.19

66.2 ± 0.9

471

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 8. AFM images of the membranes containing 1 wt % of the PVP29000 with different concentration of the polymer.

value of the PVDF-based nanocomposite ultrafiltration membranes containing 0.5 wt % of NLDH is represented in Figure 5c. As can be seen, the FRR value is enhanced significantly by increasing the concentration of the PVP 29000 up to optimum content. As aforementioned, the PVP 29000 is hydrophilic in nature, so its combination with to PVDF-based membranes increases the hydrophilicity of the membranes (as reported in Table 1). The hydrophilicity improvement resulted from the addition of PVP 29000 leads to the formation of a thin layer of water molecules on the surface of the membrane. This decreases the adsorption of BSA molecules on the surface of the membrane and results in antifouling improvement of the membranes. However, from Figure 5c, the FRR value declines by increasing the PVP 29000 concentration beyond the optimum content. This result can be ascribed to the surface roughness increment at high concentration of PVP 29000.

Table 2. Surface Roughness Parameters and Surface Porosity of Ultrafiltration Membrane Containing 1 wt.% PVP 29000 and Different Concentration of the PVDF roughness parameters membrane 16 18 20 22 24

wt wt wt wt wt

% % % % %

PVDF PVDF PVDF PVDF PVDF

Sa (nm)

Sq (nm)

Sy (nm)

85.1 68.3 54.9 48.1 34.6

107.7 98.8 84.1 79.6 67.9

595.8 501.5 487.3 398.3 376.8

surface porosity (%) 3.6 2.9 2.1 1.7 1.2

± ± ± ± ±

0.09 0.18 0.33 0.07 0.12

improvement at high concentrations of the polymers could be attributed to the reduction in the average roughness resulted from the enhanced viscosity at high concentration of the polymer. The effect of PVP 29000 concentration on the FRR

Figure 9. AFM images of the membranes containing 18 wt % of the PVDF with different concentration of PVP 29000. 472

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

inorganic materials in each dope solution. As aforementioned, the presence of the nanomaterials below the threshold concentration results in an enhancement in coagulation rate owning to thermodynamic feature improvement, which favors the formation of the high porous surface. On the other hand, presence of the nanomaterials beyond the threshold results in rheological feature improvement, leading to the formation a dense structure with low surface porosity. Therefore, the low pure water flux at high NLDH concentration can be attributed to the low surface porosity resulted from the viscosity enhancement in the presence of high concentration of NLDH. In addition, the EDAX mapping of the nanocomposite membranes demonstrate that the NLDHs were distributed homogeneously in the matrix of the membranes (see Figure 10b, e, h, k, n, and q). PVDF with dipole momentum of 2.1 D is known as polar molecule. In addition, the presence of the hydroxyl groups with unshared electrons make the layers of the MgAl-NLDH polar. So there is an electrostatic attraction between the layers of the NLDH and the PVDF chains resulting in homogeneous dispersion of the NLDH in the PVDF polymeric dope solution. The homogeneous dispersion of nanofiller in the dope solution prevents the aggregation of the NLDH in the fabricated nanocomposite membrane. Moreover, the observed trend for pure water flux can be explained based on the surface roughness of the prepared membranes. The roughness parameters of the membranes containing 18 wt % of the PVDF and 1 wt % of PVP 29000 were determined using AFM images with the diameter of 5 μm × 5 μm (Figure 11a−f), and the results are reported in Table 3. The results of roughness parameters demonstrate that the surface roughness of the nanocomposite membranes enhances by increasing the NLDH concentration up to 0.5 wt % and then decreases by further increase in the NLDH concentration. This trend can be explained based on the surface porosity of the membranes (Table 3). As aforementioned, by increasing the NLDH up to 0.5 wt % the surface porosity increases, which results in surface roughness enhancement. Whereas, the surface porosity decreases by further increasing the NLDH concentration which brought about observed decrement in roughness parameters. High surface roughness can result in an enhancement in pure water flux because of large available area for water transport. Therefore, the enhancement in pure water flux by increasing NLDH concentration up to 0.5 wt % can be ascribed to the surface roughness enhancement. As can be seen in Figure 5d, the effect of the NLDH on pure water flux is not remarkable at high concentration of polymer. Generally, the phase inversion mechanism is rheologically affected by high viscosity of the dope solution at high concentration of polymer. This results in the formation of membranes with low surface porosity and dense structure. The PVP 29000 and the polymer concentration affect the protein flux similar to pure water flux (Figure 5e). Figure 5f plots represent the effect of NLDH and polymer concentration on the FRR value of nanocomposite membranes containing 1 wt % of PVP 29000. As can be seen, there is an optimum NLDH concentration in which the nanocomposite membranes represent the best antifouling performance. The FRR improvement can be ascribed to the hydrophilicity enhancement of the nanocomposite membranes at optimum NLDH concentration. Similar to pure water flux, the effect of NLDH on the FRR is negligible at high concentrations of polymer compared with low concentrations. This confirms that

The AFM images of the membranes containing different concentrations of PVP 29000 are represented in Figure 9a−f, and the surface roughness parameters are reported in Table 1. The surface roughness of membranes enhances by increasing the PVP 29000 concentration, attributing to the surface porosity enhancement. Besides, the effect of the hydrophilicity is dominant at low concentration of the PVP 29000, which results in antifouling performance improvement. While at high concentrations of PVP 29000, the surface roughness effect is dominant which leads to decreasing the FRR value. The effect of the polymer and NLDH concentration on the pure water flux of nanocomposite membranes containing 1 wt % of PVP 29000 is shown in Figure 5d. There is an optimum concentration for NLDH at low concentration of polymer. The pure water flux of the nanocomposite membrane enhances by increasing the NLDH concentration up to about 0.5 wt %, and then decreases by further increment the NLDH concentration. Nanolayered double hydroxides are hydrophilic materials, owning to high content of hydroxyl groups on the surface of layers.37 Therefore, the presence of NLDH as hydrophilic nanomaterial in the matrix of the nanocomposite membranes can improve the hydrophilicity of the membranes, as given in Table 3. The hydrophilicity improvement could result in Table 3. Surface Roughness Parameters, Surface Porosity, and Contact Angle of Ultrafiltration Membrane 18 wt.% of PVDF, 1 wt.% PVP 29000, and Different Concentration of NLDH roughness parameters membrane 0 wt % NLDH 0.1 wt % NLDH 0.25 wt % NLDH 0.5 wt % NLDH 0.75 wt % NLDH 1 wt % NLDH

Sa (nm)

Sq (nm)

Sy (nm)

surface porosity (%)

contact angle (deg)

24.80

31.81

234.10

2.98 ± 0.33

72.0 ± 2.1

28.21

32.31

214.56

3.82 ± 0.24

67.3 ± 2.5

30.56

39.23

254.21

4.13 ± 0.32

64.7 ± 1.9

40.83

50.72

365.01

5.68 ± 0.14

61.1 ± 2.3

35.42

44.15

261.72

4.01 ± 0.28

59.8 ± 1.4

33.57

42.32

243.76

3.24 ± 0.19

58.1 ± 1.7

permeability enhancement via absorbing water molecules and facilitating their movement across the membrane. Moreover, by adding the hydrophilic NLDH to dope solution, the thermodynamic feature improves, leading to acceleration of the coagulation which consequently result in formation of membranes with high surface porosity.38,39 The cross-sectional, EDAX mapping and the surface SEM images of the membranes containing 18 wt % PVDF, 1 wt % PVP 29000, and different concentrations of NLDH are represented in Figure 10a−r. The finger-like pores can be seen in all membranes with different concentrations of NLDH. The surface SEM images demonstrate that the surface pores enhances by increasing the concentration of NLDH up to 0.5 wt % and then decreases by further increasing the NLDH concentration. It should be noted that the presence of the nanostructured inorganic materials in the dope solution can influence the phase inversion mechanism both thermodynamically and rheologically.38,39 At low and high concentrations of the nanostructured inorganic materials, the thermodynamic and rheology features are dominant, respectively. Therefore, there is a threshold for nanostructured 473

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 10. Cross-sectional, EDAX mapping, and surface SEM images of the ultrafiltration membranes containing 18 wt % of PVDF, 1 wt % PVP 29000, and different concentration of NLDH. 474

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

Figure 11. AFM images of the membranes containing 18 wt % of PVDF, 1 wt % PVP 29000, and different concentration of NLDH.

the viscosity is the main controlling parameter at high polymer concentration and the effect of NLDH is negligible. Figure 5(g) shows the contour plot of the effect of NLDH concentration and PVP 29000 on pure water flux of nanocomposite membranes containing 18 wt % of PVDF. As can be seen, the effect of NLDH is related to the concentration of PVP 29000. There is an optimum NLDH concentration in which the maximum pure water flux can be obtained at low concentration of PVP 29000. However, the effect of NLDH is negligible at high concentration of PVP 29000. Although NLDH is a hydrophilic material that affect the phase inversion method for formation of porous surface, the effect of NLDH is negligible at high PVP29000 concentrations. On the other hand, membranes with high surface porosity are formed at high PVP 29000 concentration because of the high concentration of pore former and the effect of NLDH is not remarkable. Accordingly, the effect of PVP 29000 is dominant to the hydrophilicity of the NLDH at high concentration of PVP 29000. Figure 5h shows the similarity between the protein flux and pure water flux as well as the similar effects of NLDH and PVP 29000 on the flux of BSA. Figure 5i represents the effect of the NLDH and PVP 29000 concentration on the FRR of the nanocomposite membranes containing 18 wt % of PVDF. As can be seen, at low concentration of PVP 29000, the membranes present the best antifouling performance at optimum NLDH concentration due to the hydrophilicity improvement of nanocomposite membranes. However, the presence of NLDH in the matrix of the nanocomposite membranes could not improve the antifouling performance of the membrane significantly at high concentration of the PVP 29000. Besides, there is a lot of pores on the surface of the membrane at high PVP 29000 concentration which results in high surface roughness. Therefore, the probability of fouling increased by increasing the surface roughness of the membranes. Consequently, the surface roughness effect is dominant compared with hydrophilicity of the NLDH at high concentration of the PVP 29000. The results of the protein rejection tests prove that all the prepared membranes can reject the BSA proteins more than 98%, indicating that the prepared membranes in the studied

range of PVD, PVP 29000 and NLDH concentrations act as ultrafiltration membranes without any defects in their structure. Relative Significance of the Input Parameters. The relative significance of the input parameters namely polymer concentration, pore former concentration, and the NLDH concentration was determined via the obtained ANN model using the neuron weights in hidden and output layers, according to Figure 12. All the parameters have remarkable

Figure 12. Relative significance of input parameters.

influence on the performance of the polymeric membrane, thus, none should be underestimated (Figure 12). Furthermore, the results confirmed the high importance of nanomaterials equaling to the polymer and pore former concentration in nanocomposite membranes. Optimization of Nanocomposite Membrane Performance Using GA. The obtained ANN model which was developed for modeling the relationship between the input parameters (polymer concentration, PVP 29000 concentration as pore former and the NLDH concentration) and nanocomposite membrane performance (pure water flux, protein flux, and FRR) was used as an objective function for optimization by genetic algorithm. For this purpose, the GA was used to determine the optimum values of input parameters that maximize pure water flux, protein flux, and FRR. The 475

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

able to predict the data successfully. In addition, the influence of the affecting parameters and the interaction between the parameters were investigated using the contour plots predicted with the developed model. Finally, the developed ANN model was introduced to genetic algorithm as a bioinspired optimizer to determine the optimum values of input parameters leading to high pure water flux, protein flux and flux recovery ratio. The optimum values obtained by GA for polymer, PVP 29000, and NLDH concentration were found to be 18, 1, and 0.54 wt %, respectively. The performance of the nanocomposite membrane prepared by the GA optimum values was experimentally investigated. The good agreement between the experimental values and the predicted values of pure water flux, protein flux and FRR confirmed that the performance of the nanocomposite membrane could be successfully modeled and optimized using ANN-GA hybrid system. This research open a new way for achieving optimum polymeric membranes using a cost-effective computational method.

minimum and maximum values of the input parameters are shown in Table 4, and the GA parameters used for optimization Table 4. Constraints for Input Parameters of GA

min max optimum point (ANNGA)

PVDF (wt %)

PVP 29000 (wt %)

Mg-Al NLDH (wt %)

16 24 18

0 6 1

0 1 0.54

Table 5. Genetic Algorithm Parameters Used for Optimization population creation function elite count fitness selection crossover mutation migration stopping criteria

double vector constraint dependent 0.05 × population size scaling rank stochastic uniform constraint dependent fraction = 0.8 constraint dependent forward fraction = 0.2 function tolerance = 10−6



AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Tel: 0098 41 33393165. ORCID

Alireza Khataee: 0000-0002-4673-0223 Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS We sincerely thank University of Tabriz and Kharazmi University for all the support. We also acknowledge the support of Iran Science Elites Federation and Iran National Science Foundation (INSF). We sincerely thank Dr. Shima Rahim Pouran for the language edition.

of the ANN model are reported in Table 5. The optimization process was started by creating a random initial population with size of 50 and applying GA operators including selection, crossover, mutation and migration as reported in Table 5. The optimization process reached to stopping criteria and terminated after 60 generation. The optimum values for polymer, PVP 29000 and NLDH concentration were found to be 18, 1, and 0.54 wt %, respectively. Finally, the performance of the nanocomposite membrane prepared by optimum values was experimentally investigated, and the values for pure water flux, protein flux and the FRR were found to be 654 L/m2 h, 84 L/m2 h and 65%, respectively. In addition, the protein rejection test confirmed that the optimum prepared membrane can reject more than 99% of BSA molecules. The obtained experimental values were in good agreement with the values predicted by ANN model. This good agreement with low error (lower than 6%) confirmed that the performance of the nanocomposite membranes can be successfully modeled and optimized using ANN-GA system.



REFERENCES

(1) Ayyaru, S.; Ahn, Y.-H. Application of sulfonic acid group functionalized graphene oxide to improve hydrophilicity, permeability, and antifouling of PVDF nanocomposite ultrafiltration membranes. J. Membr. Sci. 2017, 525, 210−219. (2) Mulder, J. Basic Principles of Membrane Technology; Springer Science & Business Media, 2012. (3) Li, N. N.; Fane, A. G.; Ho, W. W.; Matsuura, T. Advanced Membrane Technology and Applications; John Wiley & Sons, 2011. (4) Chen, X.; Su, Y.; Shen, F.; Wan, Y. Antifouling ultrafiltration membranes made from PAN-b-PEG copolymers: Effect of copolymer composition and PEG chain length. J. Membr. Sci. 2011, 384 (1), 44− 51. (5) Liu, F.; Hashim, N. A.; Liu, Y.; Abed, M. M.; Li, K. Progress in the production and modification of PVDF membranes. J. Membr. Sci. 2011, 375 (1), 1−27. (6) Kang, G.-d.; Cao, Y.-m. Application and modification of poly (vinylidene fluoride) (PVDF) membranes−A review. J. Membr. Sci. 2014, 463, 145−165. (7) Zheng, L.; Wu, Z.; Wei, Y.; Zhang, Y.; Yuan, Y.; Wang, J. Preparation of PVDF-CTFE hydrophobic membranes for MD application: Effect of LiCl-based mixed additives. J. Membr. Sci. 2016, 506, 71−85. (8) Zhang, X.; Lang, W.-Z.; Xu, H.-P.; Yan, X.; Guo, Y.-J.; Chu, L.-F. Improved performances of PVDF/PFSA/O-MWNTs hollow fiber membranes and the synergism effects of two additives. J. Membr. Sci. 2014, 469, 458−470. (9) Kong, J.; Li, K. Preparation of PVDF hollow-fiber membranes via immersion precipitation. J. Appl. Polym. Sci. 2001, 81 (7), 1643−1653. (10) Zhang, F.; Zhang, W.; Yu, Y.; Deng, B.; Li, J.; Jin, J. Sol−gel preparation of PAA-g-PVDF/TiO2 nanocomposite hollow fiber



CONCLUSION In this research, the MgAl-CO32− NLDH was used as a hydrophilic nanofiller to improve the performance of PVDFbased ultrafiltration membrane. The NLDHs were synthesized using a facile coprecipitation method followed by hydrothermal treatment. Subsequently, the prepared NLDH was characterized using XRD and SEM analysis. The main objective of this research was to obtain the optimized formula of the NLDH/PVDF nanocomposite ultrafiltration using computational technique. A three layer perceptron ANN model was developed for expressing the relationship between the performance of the nanocomposite membrane and the affecting parameters including the NLDH, PVP 29000 and polymer concentration. The high R2 value (0.98) for the training data confirmed that the developed model with selected topology is 476

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477

Research Article

ACS Combinatorial Science

(28) Garson, D. G. Interpreting neural network connection weights. AI Expert 1991, 6 (4), 46−51. (29) Biglarijoo, N.; Mirbagheri, S. A.; Bagheri, M.; Ehteshami, M. Assessment of effective parameters in landfill leachate treatment and optimization of the process using neural network, genetic algorithm and response surface methodology. Process Saf. Environ. Prot. 2017, 106, 89−103. (30) Woodley, S.; Battle, P.; Gale, J.; Catlow, C. R. A. The prediction of inorganic crystal structures using a genetic algorithm and energy minimisation. Phys. Chem. Chem. Phys. 1999, 1 (10), 2535−2542. (31) Li, B.; Zhang, Y.; Zhou, X.; Liu, Z.; Liu, Q.; Li, X. Different dye removal mechanisms between monodispersed and uniform hexagonal thin plate-like MgAl−CO32‑-LDH and its calcined product in efficient removal of Congo red from water. J. Alloys Compd. 2016, 673, 265− 271. (32) Pérez-Ramírez, J.; Ribera, A.; Kapteijn, F.; Coronado, E.; Gómez-García, C. J. Magnetic properties of Co−Al, Ni−Al, and Mg− Al hydrotalcites and the oxides formed upon their thermal decomposition. J. Mater. Chem. 2002, 12 (8), 2370−2375. (33) Patterson, A. The Scherrer formula for X-ray particle size determination. Phys. Rev. 1939, 56 (10), 978. (34) Akbari, A.; Hamadanian, M.; Jabbari, V.; Lehi, A. Y.; Bojaran, M. Influence of PVDF concentration on the morphology, surface roughness, crystalline structure, and filtration separation properties of semicrystalline phase inversion polymeric membranes. Desalin. Water Treat. 2012, 46 (1−3), 96−106. (35) Yan, L.; Li, Y. S.; Xiang, C. B.; Xianda, S. Effect of nano-sized Al2O3 particle addition on PVDF ultrafiltration membrane performance. J. Membr. Sci. 2006, 276 (1), 162−167. (36) Oh, S. J.; Kim, N.; Lee, Y. T. Preparation and characterization of PVDF/TiO2 organic−inorganic composite membranes for fouling resistance improvement. J. Membr. Sci. 2009, 345 (1), 13−20. (37) Wang, Q.; O’Hare, D. Recent advances in the synthesis and application of layered double hydroxide (LDH) nanosheets. Chem. Rev. 2012, 112 (7), 4124−4155. (38) Mehrnia, M. R.; Mojtahedi, Y. M.; Homayoonfal, M. What is the concentration threshold of nanoparticles within the membrane structure? A case study of Al2O3/PSf nanocomposite membrane. Desalination 2015, 372, 75−88. (39) Yu, L. Y.; Xu, Z. L.; Shen, H. M.; Yang, H. Preparation and characterization of PVDF−SiO2 composite hollow fiber UF membrane by sol−gel method. J. Membr. Sci. 2009, 337 (1), 257−265.

membranes with extremely high water flux and improved antifouling property. J. Membr. Sci. 2013, 432, 25−32. (11) Safarpour, M.; Khataee, A.; Vatanpour, V. Preparation of a Novel Polyvinylidene Fluoride (PVDF) Ultrafiltration Membrane Modified with Reduced Graphene Oxide/Titanium Dioxide (TiO2) Nanocomposite with Enhanced Hydrophilicity and Antifouling Properties. Ind. Eng. Chem. Res. 2014, 53 (34), 13370−13382. (12) Safarpour, M.; Khataee, A.; Vatanpour, V. Effect of reduced graphene oxide/TiO2 nanocomposite with different molar ratios on the performance of PVDF ultrafiltration membranes. Sep. Purif. Technol. 2015, 140, 32−42. (13) Vatanpour, V.; Yekavalangi, M. E.; Safarpour, M. Preparation and characterization of nanocomposite PVDF ultrafiltration membrane embedded with nanoporous SAPO-34 to improve permeability and antifouling performance. Sep. Purif. Technol. 2016, 163, 300−309. (14) Vatanpour, V.; Esmaeili, M.; Farahani, M. H. D. A. Fouling reduction and retention increment of polyethersulfone nanofiltration membranes embedded by amine-functionalized multi-walled carbon nanotubes. J. Membr. Sci. 2014, 466, 70−81. (15) Lin, J.; Zhang, R.; Ye, W.; Jullok, N.; Sotto, A.; Van der Bruggen, B. Nano-WS2 embedded PES membrane with improved fouling and permselectivity. J. Colloid Interface Sci. 2013, 396, 120−128. (16) Lin, J.; Ye, W.; Zhong, K.; Shen, J.; Jullok, N.; Sotto, A.; Van der Bruggen, B. Enhancement of polyethersulfone (PES) membrane doped by monodisperse Stöber silica for water treatment. Chem. Eng. Process. 2016, 107, 194−205. (17) Zhao, Y.; Lu, J.; Liu, X.; Wang, Y.; Lin, J.; Peng, N.; Li, J.; Zhao, F. Performance enhancement of polyvinyl chloride ultrafiltration membrane modified with graphene oxide. J. Colloid Interface Sci. 2016, 480, 1−8. (18) Wan, S.; Wang, S.; Li, Y.; Gao, B. Functionalizing biochar with Mg−Al and Mg−Fe layered double hydroxides for removal of phosphate from aqueous solutions. J. Ind. Eng. Chem. 2017, 47, 246−253. (19) Mohapatra, L.; Parida, K. A review on the recent progress, challenges and perspective of layered double hydroxides as promising photocatalysts. J. Mater. Chem. A 2016, 4 (28), 10744−10766. (20) Al-Abri, M.; Hilal, N. Artificial neural network simulation of combined humic substance coagulation and membrane filtration. Chem. Eng. J. 2008, 141 (1), 27−34. (21) Madaeni, S.; Hasankiadeh, N. T.; Kurdian, A.; Rahimpour, A. Modeling and optimization of membrane fabrication using artificial neural network and genetic algorithm. Sep. Purif. Technol. 2010, 76 (1), 33−43. (22) Samadian, H.; Zakariaee, S. S.; Adabi, M.; Mobasheri, H.; Azami, M.; Faridi-Majidi, R. Effective parameters on conductivity of mineralized carbon nanofibers: an investigation using artificial neural networks. RSC Adv. 2016, 6 (113), 111908−111918. (23) Daood, G. S.; Basri, H.; Stanslas, J.; Fard Masoumi, H. R.; Basri, M. Predicting the optimum compositions of a parenteral nanoemulsion system loaded with azithromycin antibiotic utilizing the artificial neural network model. RSC Adv. 2015, 5 (101), 82654− 82665. (24) Azarhoosh, M.; Farivar, F.; Ale Ebrahim, H. Simulation and optimization of a horizontal ammonia synthesis reactor using genetic algorithm. RSC Adv. 2014, 4 (26), 13419−13429. (25) Askari, S.; Halladj, R.; Azarhoosh, M. J. Modeling and optimization of catalytic performance of SAPO-34 nanocatalysts synthesized sonochemically using a new hybrid of non-dominated sorting genetic algorithm-II based artificial neural networks (NSGA-IIANNs). RSC Adv. 2015, 5 (65), 52788−52800. (26) Kumar, R. V.; Moorthy, I. G.; Pugazhenthi, G. Modelling and optimization of critical parameters by hybrid RSM-GA for the separation of BSA using a tubular configured MFI-type zeolite microfiltration membrane. RSC Adv. 2015, 5 (106), 87645−87659. (27) Oskoui, S. A.; Niaei, A.; Tseng, H.-H.; Salari, D.; Izadkhah, B.; Hosseini, S. A. Modeling preparation condition and composition− activity relationship of perovskite-type Lax Sr1−x Fey Co1−y O3 nano catalyst. ACS Comb. Sci. 2013, 15 (12), 609−621. 477

DOI: 10.1021/acscombsci.7b00046 ACS Comb. Sci. 2017, 19, 464−477