Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm

G Model

JTICE-624; No. of Pages 8 Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

Contents lists available at SciVerse ScienceDirect

Journal of the Taiwan Institute of Chemical Engineers journal homepage: www.elsevier.com/locate/jtice

Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm Ming Tan, Gaohong He *, Fei Nie, Lingling Zhang, Liangping Hu State Key Laboratory of Fine Chemicals, R&D Center of Membrane Science and Technology, School of Chemical Engineering, Dalian University of Technology, 2 Linggong Road, Dalian 116024, China

A R T I C L E I N F O

A B S T R A C T

Article history: Received 16 December 2012 Received in revised form 26 March 2013 Accepted 11 April 2013 Available online xxx

Hybrid models based on backpropagation neural network (BPNN) and genetic algorithm (GA) were constructed to optimize the fabrication of polyetherimide (PEI) ultrafiltration (UF) membrane via dry/ wet phase inversion. BPNN was employed to capture the detailed relationships between the preparation conditions and the UF membrane performances, and GA was used to choose the initial connection weights and biases of BPNN to avoid convergence at suboptimal solutions. The excellent agreements between the model predictions and the testing data indicate that the hybrid models have sufficient accuracy. The effects of preparation conditions on membrane performances were predicted by the hybrid models successfully, which indicate that PEI/N,N-dimethylacetamide (DMAc)/1,4-butyrolactone (GBL) is the best membrane casting system investigated in this study. Furthermore, the optimal preparation conditions were forecasted, and membranes with desired performances, for instance, higher pure water flux (PWF) and bovine serum albumin (BSA) rejection ratio (RR) 80–90% were fabricated with the standard deviation between the predicted performances and validation experimental values less than 10%. The hybrid models can contribute to collaborative optimization of multiple parameters and designing the preparation conditions to obtain desired UF membrane performances and avoiding large experimental data scattering in the fabrication of phase inversion membranes. ß 2013 Taiwan Institute of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

Keywords: Membrane Backpropagation neural network Genetic algorithm Phase inversion Preparation condition

1. Introduction Ultrafiltration (UF) membrane is extensively applied to the product recovery and pollution control in chemical, electronic, food, pharmaceutical, water treatment and biotechnological industries [1,2]. Pure water flux (PWF) and rejection ratio (RR) are the most important performances to characterize UF membranes [3,4]. To optimize the fabrication of UF membrane via dry/ wet phase inversion, traditional orthogonal method is the most widely used owing to its sufficient accuracy [5,6]. However, the method cannot get a function expression between preparation conditions and membrane performances and hence it is difficult to find out the optimal factor combination [7,8].

Abbreviations: BPNN, backpropagation neural network; BSA, bovine serum albumin; BuOH, n-butanol; Da, Dalton; DE, diethyl ether; DMAc, N,N-dimethylacetamide; GA, genetic algorithm; GBL, 1,4-butyrolactone; PEG400, polyethylene glycol with average molecular weight of 400 Da; PEI, polyetherimide; PDMS, polydimethylsiloxane; PVP, polyvinylpyrrolidone; PWF, pure water flux; RR, rejection ratio; SSE, sum square error; UF, ultrafiltration. * Corresponding author. Tel.: +86 411 84707892; fax: +86 411 84707700. E-mail address: [email protected] (G. He).

Backpropagation neural network (BPNN) due to its good robustness and fault tolerance is widely used in optimization and function approximation [9–13]. The biggest problem involved in the application of BPNN is easily convergent to a local solution [14,15]. To overcome this problem, several global search techniques including genetic algorithm (GA) have been developed. Up until now, GA has mainly been used to search the optimal solution of BPNN function [16–20]. In the field of membranes, BPNN models have been used frequently in membrane filtration process, to predict the evolution of membrane fouling [21–25] or membrane performances under different separation parameters, for instance, concentration of solute, solution viscosity, transmembrane pressure difference, temperature and filtration time [10,11,26–30]. The application of BPNN to the membrane fabrication has rarely been reported yet. In our previous work [31], we successfully constructed hybrid models which employed GA to choose the initial connection weights of BPNN to predict the effects of preparation conditions on pervaporation performances of polydimethylsiloxane (PDMS)/ceramic composite membranes. However, the UF membrane which is porous is totally different from the pervaporation membrane that is nonporous. Therefore, it is necessary to construct hybrid models for optimization of UF membrane fabrication.

1876-1070/$ – see front matter ß 2013 Taiwan Institute of Chemical Engineers. Published by Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jtice.2013.04.004

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 2

M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

Nomenclature Symbols b1 b2 Cadd Cf Cp CPEI Fn Fp IW logsig LW max(ni) min(ni) ni nˆ i P Q Rp S t T tansig tadd tp

biases of input/hidden layer biases of hidden/output layer the concentration of additive BSA concentration of the feed BSA concentration of the permeate the concentration of PEI casting solution output of the PWF prediction model predicted PWF (de-normalized of Fn) connection weights of input/hidden layer log-sigmoid transfer function connection weights of hidden/output layer maximum value of ni minimum value of ni numerical value of each preparation condition normalized value of ni model input vector composed of the normalized value of ni volume of the permeate pure water (m3) predicted RR (output of the RR prediction model) effective area of the membrane (m2) evaporation time temperature of the water coagulation bath tan-sigmoid transfer function the type of additive permeation time (h)

Subscripts PWF F RR R

In this study, the fabrication of polyetherimide (PEI) UF membrane via dry/wet phase inversion was optimized by the hybrid models. Effects of PEI concentration, temperature of water coagulation bath, additive type and concentration on membrane performances were predicted by the hybrid models and verified by experimental data. According to the hybrid models, the best membrane casting system of the six was determined. Furthermore, the optimal preparation conditions were forecasted, and membranes with desired performances were fabricated.

used without further purification. All the chemicals were of analytical grade. 2.2. Membrane preparation The PEI UF membranes were prepared by dry/wet phase inversion method. Polymer PEI without or with a kind of additive (BuOH, DE, PEG400, PVP and GBL) was dissolved in solvent DMAc by mechanical stirring for 8 h at 90 8C to form a homogenous membrane casting solution. Air bubbles in the casting solution were removed by vacuum degassing for 30 min. The casting solution was cast on a non-woven fabric (Ahlstrom, Finland). After exposed to air with the relative humidity of around 55% for a few seconds, the cast films were immersed in a water coagulation bath for 24 h, where the polymer precipitation occurred due to the exchange of solvent in the cast film and non-solvent (water) in the coagulation bath, and then the membrane was formed. 2.3. Membrane characterization A home made UF cell was used to measure pure water flux (PWF) and rejection ratio (RR), with the effective membrane area of 33.18  104 m2 and transmembrane pressure difference of 0.1 MPa. Prior to the experiments, the as-prepared membrane was compacted in the cell by deionized water for 30 min under the transmembrane pressure difference of 0.15 MPa. The PWF (m3/(m2 h)) is defined by Eq. (1). PWF ¼

Q S  tp

(1)

where Q is the volume of the permeate pure water (m3), S is the effective area of the membrane (m2), and tp is the permeation time (h). The RR, tested with 0.5 kg/m3 BSA solution, is calculated by Eq. (2). RR ¼ 1 

Cp Cf

(2)

where Cp and Cf is the BSA concentration of permeate and feed, respectively, which is determined by ultraviolet–vis spectrophotometer (Shanghai xinmao instrument Co., Ltd., China) at 280 nm. 3. Modeling schemes of the hybrid models

2. Experimental 2.1. Materials PEI purchased from General Electric Plastics (USA) in pellet form and polyvinylpyrrolidone (PVP) supplied by Shanghai Chemical Reagent Station (China), were dried in a vacuum oven at 105 8C to constant weight. 1,4-Butyrolactone (GBL) purchased from Tianjin Guangfu Fine Chemical Research Institute (China) and N,N-dimethylacetamide (DMAc) supplied by Beijing Chemical Plant (China) were dried over molecular sieve beads (50 nm, Dalian Liaodong Chemical Reagent Co., China) before used. Bovine serum albumin (BSA) with average molecular weight of 67,000 Dalton (Da) bought from Beijing Aoboxing Biological Product Company (China), n-Butanol (BuOH) purchased from Shenyang Reagent Plant 3 (China), Diethyl Ether (DE) supplied by Tianjin Tanggu Industrial and Commercial Industry Co. (China), and polyethylene glycol with average molecular weight of 400 Da (PEG400) bought from Guangdong Province Xilong Chemical Factory (China) were

A commercially available software program (MATLAB Version 7.0.0.19920, genetic algorithm and direct search toolbox v. 1.0.1, neural network toolbox v. 4.0.3, the Math Works Inc.) was used to implement GA and BPNN on a personal computer. 3.1. Training/testing data In the dry/wet phase inversion process, membrane performances are mainly determined by the concentration of PEI casting solution (CPEI), the type of additive (tadd), the concentration of additive (Cadd), evaporation time (t), temperature of the water coagulation bath (T) and relative humidity of air [32]. In our preparation processes, relative humidity of air remained constant at about 55%. Therefore, the variables that influenced the membrane performances were CPEI, tadd, Cadd, t and T, which were considered as the model inputs. Model inputs must be normalized to avoid numerical overflows due to very large or very small weights [33,34]. The normalization

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

equation is shown as Eq. (3). nˆ i ¼ 2 

ni  minðni Þ  1; maxðni Þ  minðni Þ

i ¼ 1; 2 . . . 5

(3)

where nˆ i is the normalized value of ni, ni is the numerical value of each preparation condition, and min(ni) and max(ni) is the minimum and maximum value of ni, respectively. After the normalization, all of the model inputs were between 1 and +1. As model inputs, the type of additive must be quantified. Here, without additive, additive of BuOH, DE, PEG400, PVP, GBL was evenly represented by 1.0, 0.6, 0.2, +0.2, +0.6 and +1.0, respectively. The code which represented additive type is not unique. For example, additive BuOH can also be expressed as +1.0. However, after the hybrid models were established, the code must be constant. If PWF and RR were predicted by a hybrid model simultaneously, the standard deviations between the predictions and the experiments were even more than 100%. Therefore, two hybrid models were constructed to predict PWF and RR, respectively. The original experimental data were divided into two subsets [35,36]. One was the training data, and the other was the testing data. The training data were summarized in Table 1 which shows that the experimental fluxes range from 0.01 to 1.45 m3/(m2 h). To make the PWF prediction model be convergent and have certain ability in extrapolation, the output fluxes were also normalized by Eq. (3), whose minimal and maximal values were set at 0 and 1.60 m3/(m2 h).

3

hidden layer were proposed. It was found that there is no significant change in the model accuracy. Therefore, our models were equipped with one hidden layer and 11 neurons. The most widely used transfer functions are the log-sigmoid (logsig) and the tan-sigmoid function (tansig) [38], which is expressed by Eqs. (4) and (5), respectively. Function logsig produces outputs in the range of 0 to +1, and function tansig produces outputs in the range of 1 to +1. logsigðxÞ ¼

1 1 þ ex

(4)

tansigðxÞ ¼

ex  ex ex þ ex

(5)

For the PWF prediction model, PWF was normalized by Eq. (3) and hereby ranged from 1 to +1, and therefore the transfer function of hidden/output layer must be tansig. Meanwhile, for the RR prediction model, RR ranged from 0 to 100% and hence the transfer function of hidden/output layer had to be logsig. Next step was to determine the transfer function of input/ hidden layer. PWF and RR prediction models with input/hidden layer transfer function of logsig were compared with the models with input/hidden layer transfer function of tansig, and the mean standard deviations between the predictions and the experiments were listed in Table 2. From the comparison, it can be deduced that the input/hidden layer transfer function of both the PWF and RR prediction model should be logsig. 3.3. Hybrid models

3.2. Determination of the model architecture The universal approximation theory suggested that a neural network with a single hidden layer and (2N + 1) neurons was able to approximate any continuous function converting the Ndimensional input vector into the M-dimensional output vector [37]. Moreover, hybrid models with 11–16 neuron numbers in the

Fig. 1 shows the flowchart of the hybrid models, and the optimization scheme is similar to our previous study [31]. The parameters of GA used in this study were as follows: (I) population size was 100, (II) the maximum number of generations was 500, (III) crossover probability was 0.90 and (IV) mutation probability was 0.05. The sum square error (SSE) objectives of BPNN training

Table 1 Training data of the hybrid models. CPEI (g/g)

Preparation conditions Additive type

0.12 0.14 0.17 0.21 0.17 0.17 0.17 0.17 0.17 0.17 0.15 0.15 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.15 0.15 0.15

tadd

Representation

Without Without Without Without BuOH BuOH BuOH BuOH BuOH DE DE DE PEG400 PEG400 PEG400 PVP PVP PVP GBL GBL GBL GBL GBL GBL GBL GBL GBL

1 1 1 1 0.6 0.6 0.6 0.6 0.6 0.2 0.2 0.2 0.2 0.2 0.2 0.6 0.6 0.6 1 1 1 1 1 1 1 1 1

Membrane performance Cadd (g/g)

T (8C)

t (s)

PWF (m3/(m2 h))

RR (%)

0 0 0 0 0.02 0.06 0.08 0.04 0.04 0.07 0.07 0.07 0.02 0.07 0.11 0.03 0.09 0.15 0.04 0.13 0.21 0.07 0.07 0.07 0.1 0.1 0.1

24 24 24 24 24 24 24 14 44 24 24 24 24 24 24 24 24 24 24 24 24 24 24 24 14 24 44

5 5 5 5 5 5 5 5 5 5 10 60 5 5 5 5 5 5 5 5 5 10 30 60 5 5 5

1.45 0.95 0.523 0.13 0.89 0.648 0.4 0.649 0.789 0.29 0.721 0.01 0.524 0.834 1.021 0.411 0.288 0.134 0.767 0.66 0.473 0.577 0.659 0.015 0.705 0.943 1.021

50 75 92.4 99.5 92.5 68 46.7 99 20.1 99.1 85.6 99.8 93 75.9 30.3 88.9 92.8 99.1 97 99 99.5 82.2 33.9 99.6 84.5 78 83.6

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

1.50

0

20

40

60

tansig

3.6% 4.1%

3.37% 6.30%

for both PWF and RR prediction models were set at 0.05 to minimize the error and preclude overfitting. 4. Results and discussion

1.00

4.1. Detailed relationships between the preparation conditions and the membrane performances

80 0.75 60 0.50 40

2

R =0.9662

0.25

20

0.00

0.00

0.25

0.50

0.75

1.00 3

-2

Experiment PWF/m .m .h According to the optimal connection weights and biases listed in supplementary materials, the hybrid models to predict PWF and RR are described as Eqs. (6)–(8). F n ¼ tan sigfLWF  ½logsigðIWF  P þ b1F Þ þ b2F g

1.50

0

1.25

20

40

-1

where, P is the model input vector composed of the normalized value of ni. IW and b1 are the connection weights and biases of input/hidden layer. LW and b2 are the connection weights and biases of hidden/output layer. The subscripts F and R represent PWF and RR, respectively. Fn is the output of the PWF prediction model. Fp, de-normalized of Fn, and Rp are predicted PWF and RR. Eqs. (6)–(8) are then used in the prediction. Fig. 2 shows the comparisons between the experimental data and the model predictions. Linear correlation coefficients of the training and total data (training and testing data) for the PWF is 0.9899 and 0.9752, respectively, while those for the RR are 0.9662 and 0.9632. These linear correlation coefficients, close to 1, reveal that the constructed hybrid models are statistically accurate [39].

80

100

PWF 100

2

R =0.9752

1.00

-2 3

(8)

Predicted PWF/m .m .h

Rp ¼ log sigfLWR  ½log sigðIWR  P þ b1R Þ þ b2R g

0 1.50

-1

60

RR

(7)

1.25

Experiment RR/%

(b)

(6)

F p ¼ ðF n þ 1Þ  ð1:6  0Þ  0:5

Predicted RR/%

logsig

13.64% 9.5%

-1

tansig

100

2

R =0.9899

RR

logsig

100

PWF

1.25

6.6% 7.4%

80

-2

Training data Testing data

RR prediction model

3

PWF prediction model

Experiment RR/%

(a)

80 0.75 60 0.50 40

2

R =0.9632

0.25

20

0.00

0.00

Predicted RR/%

Table 2 Model comparison of the mean standard deviation with different input/hidden transfer functions.

Predicted PWF/m .m .h

4

0.25

0.50

0.75

1.00 3

-2

Experiment PWF/m .m .h

1.25

0 1.50

-1

Fig. 2. Experimental PWF and RR versus model predictions for the (a) training data and (b) total data.

4.2. Prediction of the effects of preparation conditions on membrane performances The effects of preparation conditions on membrane performances are shown in Figs. 3–5. The testing data which have never been used in the model training were employed to verify the model accuracy.

Input training data to GA Generate initial random population

1.5

Evaluate GA fitness value of each population

100 1.2

Crossover

80

Mutation 0.9

60

3

Population of the next generation No Maximal generation (GA) reached Yes

0.6

0.3

GA solution: initial connection weights of BPNN BPNN training

0.0 11

40 Predicted PWF Testing PWF Predicted RR Testing RR

RR/%

-2

PWF/m .m .h

-1

Selection

20 0

13

15

17

19

21

Concentration of PEI/wt.% Final solution: final connection weights and bias Fig. 1. Flowchart of optimization scheme based on the hybrid models.

Fig. 3. Effects of PEI concentration on membrane performances predicted by the hybrid models (without additive, t = 5 s and T = 24 8C).

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

1.2 1.0

100

GBL GBL

80

0.6

60

3

BuOH Predicted PWF Testing PWF (BuOH) Testing PWF (GBL) Predicted RR Testing RR (BuOH) Testing RR (GBL)

0.4 0.2

BuOH

40 20

0.0 9

14

19

24

RR/%

-2

PWF/m .m .h

-1

0.8

29

34

39

0 49

44

Temperature of water coagulation bath/ºC Fig. 4. Effects of coagulation bath temperature on membrane performances predicted by the hybrid models with t of 5 s (for BuOH additive, CPEI = 17 wt.%, Cadd = 4 wt.%; for GBL additive, CPEI = 15 wt.%, Cadd = 10 wt.%).

Fig. 3 shows the effects of PEI concentration on membrane performances. As the PEI concentration increases, the PWF decreases and the RR increases. Increasing the polymer concentration leads to closer packing of polymer segments, thereby rendering the membrane skin layer denser [40]. The predictions of the hybrid models accord with the experimental testing data. For example, when the PEI concentration is 18%, the predicted PWF and RR are 0.432 m3/(m2 h) and 95.1%, and the experimental

(a)

1.2 100 1.0

60

3

0.6 0.4

40

Predicted PWF Testing PWF Predicted RR Testing RR

0.2

RR/%

-2

PWF/m .m .h

-1

80 0.8

20

0.0

0 2

4

6

8

10

12

Concentration of PEG400/wt. %

1.0

100

0.8

80

0.6

60

0.4

40

Predicted PWF Testing PWF Predicted RR Testing RR

3

0.2

RR/%

1.2

-2

PWF/m .m .h

-1

(b)

20

0.0

0 3

5

7

9

11

13

15

17

19

21

Concentration of GBL/wt.% Fig. 5. Effects of additive concentration on membrane performances predicted by the hybrid models (CPEI = 17 wt.%, t = 5 s and T = 24 8C) with additive of (a) PEG400 and (b) GBL.

5

testing PWF and RR are 0.375 m3/(m2 h) and 99.5%, as the points in Fig. 3, with the standard deviation of 15.2% and 4.4%, respectively. Fig. 4 shows the effects of coagulation bath temperature on membrane performances. As coagulation bath temperature increases, the PWFs increase, while RR of the PEI/DMAc/BuOH membrane decreases and that of the PEI/DMAc/GBL membrane keeps constant. It is known that fast diffusion velocity of the solvent DMAc, additive and the nonsolvent water at high coagulation bath temperature results in high porosity of the membrane [41,42], which is responsible for the increase of PWF. Besides, since molecular size of BuOH (0.66 nm, calculated by materials studio) is larger than DMAc (0.49 nm), fast leaching of additive BuOH leads to larger pore size, which brings on lower RR of PEI/DMAc/BuOH membrane. For PEI/DMAc/GBL membrane, due to approximately equal molecular size of GBL (0.45 nm) and DMAc (0.49 nm), fast leaching of additive GBL is similar to that of DMAc. Therefore, pore size almost keeps constant, which accounts for the constant RR. The predictions agree with the testing data very well. For BuOH additive, when the coagulation bath temperature is 24 8C, the predicted PWF and RR are 0.722 m3/(m2 h) and 88.1%. And the experimental testing PWF and RR are 0.686 m3/(m2 h) and 89%, with the standard deviation of 5.2% and 1.1%, respectively. For GBL additive, when the coagulation bath temperature is 34 8C, the predicted PWF and RR are 0.994 m3/(m2 h) and 85.2%, while the experimental testing PWF and RR are 0.986 m3/(m2 h) and 86.7%, with the standard deviation of 0.8% and 1.7%. Table 3 shows the effects of additive type and concentration on membrane performances. DE volatilizes easily during the evaporation time in air, and thus the PEI concentration on the surface of the cast film increases which brings about a dense skin layer [6,43– 45]. Therefore, membrane PWF with additive of DE decreases and RR increases. PVP has two contradictory influences on membrane performances. On the one hand, since it is soluble in water coagulation bath, it is a pore-forming agent [46,47]. On the other hand, the addition of PVP sharply increases the viscosity of the casting solution thereby decreases the inlet velocity of nonsolvent and outlet velocity of solvent and additive and consequently makes a denser skin layer [48–50]. The latter is the primary. Hence, membrane PWF with additive of PVP decreases and RR increases. Pore former additive of membrane, PEG400 [51,52], results in a more porous structure which brings on the increase of PWF and decrease of RR. BuOH, typical nonsolvent additive [53,54], enhances the formation of macrovoids, and induces higher porosity of the membrane. Moreover, molecular size of BuOH is higher than DMAc. Hence, diffusion of BuOH into the water coagulation bath causes the increase of pore size. However, when additive BuOH increases to a certain value, the greatly increased viscosity of the membrane casting solution hinders the diffusion of the solvent DMAc, additive and the nonsolvent water hereby causes a thicker skin layer. As a result, membrane PWF with additive BuOH first increases then decreases, and RR decreases. Fig. 5 which shows the effects of PEG400 and GBL (example additive) concentrations on membrane performances validates the powerful prediction ability of the hybrid models. Table 3 and Fig. 5 show that membrane performances can be predicted by the hybrid models with most of the deviations less than 10%, although they have various trends. In terms of the comparisons between the experimental testing data and model predictions, as well as the linear correlation coefficients, it is deduced that the hybrid models have sufficient accuracy. In additive, compared to other additives, membrane PWFs with additive GBL are relatively high and corresponding RRs change slightly. Therefore, PEI/DMAc/GBL is the best of the six membrane casting systems investigated in this study, including systems of PEI/DMAc, PEI/DMAc/BuOH, PEI/DMAc/DE, PEI/DMAc/ PVP, PEI/DMAc/PEG400 and PEI/DMAc/GBL.

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

6

Table 3 Membrane performances with different additives. CPEI (g/g)

tadd

Cadd (g/g)

t (s)

T (8C)

Testing PWF (m3/(m2 h))

Testing RR (%)

Predicted PWF (m3/(m2 h))

PWF error (%)

Predicted RR (%)

RR error (%)

0.13 0.16 0.18 0.2 0.17 0.17 0.17 0.17 0.15 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.15 0.15 0.15 0.15 0.16 0.16 0.17 0.17

Without Without Without Without BuOH BuOH BuOH BuOH DE PEG400 PEG400 PEG400 PEG400 PVP PVP GBL GBL GBL GBL GBL GBL GBL GBL GBL GBL

0 0 0 0 0.04 0.04 0.04 0.04 0.07 0.03 0.05 0.08 0.10 0.06 0.12 0.09 0.17 0.1 0.1 0.1 0.13 0.09 0.13 0.07 0.15

24 24 24 24 5 5 5 5 30 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

5 5 5 5 19 24 34 39 24 24 24 24 24 24 24 24 24 19 34 39 24 24 24 24 24

1.08 0.75 0.375 0.175 0.659 0.686 0.721 0.791 0.621 0.54 0.721 0.842 0.889 0.38 0.192 0.69 0.5 0.788 0.986 1 0.709 0.817 0.664 0.626 0.562

66.6 88 99.5 100 98.7 89 48.8 28.6 92.6 100 88.5 75.5 45.6 91.6 95.5 98 97 83.9 86.7 84.1 88.2 91.7 96.4 97.1 98.9

1.213 0.669 0.432 0.196 0.705 0.722 0.722 0.743 0.626 0.571 0.667 0.857 0.97 0.35 0.225 0.705 0.547 0.853 0.994 0.998 0.837 0.817 0.713 0.733 0.586

12.3 10.9 15.2 11.8 7.0 5.2 0.2 6.1 0.8 5.7 7.4 1.8 9.1 7.8 17.1 2.2 9.4 8.3 0.8 0.2 18.1 0 7.4 17.1 4.2

62.5 87.5 95.1 98.2 95.8 88.1 50.9 32.1 95.2 99.7 97.4 67.4 40.7 90.1 88.8 94.6 99.6 81.8 85.2 85.6 93.4 89.5 97.3 92.3 99.4

6.1 0.5 4.4 1.8 2.9 1 4.2 12.2 2.9 0.3 10.0 10.7 10.7 1.6 7 3.5 2.7 2.5 1.7 1.8 5.9 2.4 1.0 4.9 0.5

4.3. Optimization of the preparation conditions It is convenient to prepare membranes with evaporation time of 5 s and coagulation bath temperature of 24 8C and PEI/DMAc/GBL is the best of the six membrane casting systems. Consequently, under above preparation conditions, membrane performances with different PEI and GBL concentration were predicted and plotted in Fig. 6. Firstly, additive GBL suppresses the growth of macrovoids. Therefore, the membrane changes from a finger-like to a spongelike structure and surface porosity of the membrane increases [55]. Secondly, GBL greatly increases the viscosity of the casting solution [55,56] and hereby reduces the inlet velocity of nonsolvent and outlet velocity of the solvent DMAc. Lastly, GBL is a poor solvent for polymer PEI, and the mutual affinity between GBL and water is weaker than that between DMAc and water [57]. The latter two

lead to the decrease of phase separation velocity. The interactions among the three factors induce the decrease of PWF and nonuniform RR with the increasing GBL concentration. It can be deduced from Fig. 6 that the hybrid models can be adopted to predict performances of the PEI UF membranes under various preparation conditions. Hence, the hybrid models can contribute to collaborative optimization of multiple parameters and avoiding large experimental data scattering in the fabrication of phase inversion UF membranes. Usually, there is a trade-off between PWF and RR for UF membrane. Membrane which has high PWF often has low RR and vice versa. In practical application of UF membrane, higher PWF and RR of 90% are commonly required. That is the reason why the RRs of the experiments to train and test the hybrid models are high. However, when molecular size rejected in filtration solution is

Fig. 6. Membrane performances with different PEI and GBL concentration predicted by the hybrid models (GBL additive, t = 5 s, T = 24 8C).

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

were much larger than those in literatures with the same BSA rejection, which indicates that the hybrid models are reliable to optimize the fabrication of UF membranes and can contribute to designing the preparation conditions to obtain desired UF membrane performances and avoiding large experimental data scattering in the fabrication of phase inversion membranes.

100 80

RR/%

7

60

5. Conclusion

40 Subsquent experiments based on the model predictions Predicted higher PWF and desired RR 80-90% Experiments to train and test the hybrid models

20 0 0.00

0.25

0.50

0.75 3

1.00 -2

PWF/m .m .h

1.25

1.50

-1

Fig. 7. Higher PWF and BSA RR 80–90% predicted by the hybrid models and subsequent experiments based on the model predictions.

larger than that of BSA which is used to characterize the membrane, the desired BSA RR is lower than 90%. According to the hybrid models, the desired performances, such as higher PWF with desired RR 80–90%, were shown in Fig. 7. In addition, some corresponding preparation conditions were predicted in Table 4. In order to validate the points predicted by the hybrid models, three membranes in Table 4 (No. 3, 7, 10) were prepared consequently. The results in Table 5 show that the experimental performances of the three membranes are in good agreement with the predicted values, with the standard deviation less than 10%. In addition, when BSA RRs are 80%, 85% and 90%, the PWFs in literatures for polyphthalazinone ether sulfone ketone (PPESK) membranes are smaller than 0.52, 0.48 and 0.4 [58], those for polyvinylidene fluoride/polyethersulfone (PVDF/PES) blend membranes are less than 0.06, 0.04, 0.03 [46], and those for PVDF/graphene oxide blend membranes are between 0.3 and 0.4 m3/(m2 h) [59]. Compared with the membrane performances in Table 5, the PWFs of the membranes prepared in this study based on the model predictions

Table 4 Some preparation conditions to fabricate membranes with higher PWF and desired RR 80–90%. No.

CPEI

Cadd

PWF (m3/(m2 h))

RR (%)

1 2 3 4 5 6 7 8 9 10

0.129 0.130 0.131 0.131 0.132 0.132 0.133 0.135 0.136 0.139

0.164 0.158 0.152 0.162 0.156 0.166 0.158 0.160 0.156 0.150

1.162 1.161 1.161 1.117 1.117 1.072 1.081 1.020 1.013 0.968

80.3 80.5 80.1 83.1 83.1 85.2 85.0 88.0 88.3 90.0

Table 5 Experiments for validating the membranes with higher PWF and desired RR 80–90% identified by the hybrid models. No.

Predicted higher PWF (m3/(m2 h))

Experimental PWF (m3/(m2 h))

PWF error

Predicted designed RR (%)

Experimental RR (%)

RR error

3 7 10

1.161 1.081 0.968

1.09 1.03 0.999

6.51% 4.95% 3.10%

80.1 85.0 90.0

81.5 78 83.3

1.72% 8.97% 8.04%

Hybrid models based on BPNN and GA were constructed to optimize the fabrication of PEI UF membrane via dry/wet phase inversion. BPNN was employed to capture the detailed relationships between the preparation conditions and the UF membrane performances, GA was used to choose the initial connection weights and biases of BPNN to avoid convergence at suboptimal solutions, and trial-and-error method was used to determine the architectures of the hybrid models. The consistencies between the predictions and the testing data indicate that the hybrid models have sufficient accuracy. The effects of PEI concentration, temperature of water coagulation bath, additive type and concentration on membrane performances were predicted by the hybrid model successfully, which indicate that PEI/DMAc/GBL is the best of the six membrane casting systems investigated in this study. Furthermore, the optimal preparation conditions were forecasted, and membranes with desired performances, for instance, higher PWF and BSA RR 80–90% were fabricated with the standard deviation between the predicted performances and validation experimental values less than 10%. The hybrid models can contribute to collaborative optimization of multiple parameters and designing the preparation conditions to obtain desired UF membrane performances and avoiding large experimental data scattering in the fabrication of phase inversion membranes. Acknowledgements The authors thank the financial support of National Science Fund for Distinguished Young Scholars of China (21125628), the National High Technology Research and Development Program of China (2012AA03A611), Science and Technology plan projects of Liaoning Province of China (2011224005) and the Fundamental Research Funds for the Central Universities (DUT11ZD112). Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.jtice.2013.04.004. References [1] Dave YG, Reddy AVR. Preparation, characterization, acid stability and organic fouling of poly(acrylonitrile-co-methacrylic acid) ultrafiltration membranes. Desalination 2011;282:9–18. [2] Lin T, Li L, Chen W, Pan S. Effect and mechanism of preoxidation using potassium permanganate in an ultrafiltration membrane system. Desalination 2012;286:379–88. [3] Wu C, Zhang S, Liu C, Yang D, Jian X. Preparation, characterization and performance of thermal stable poly(phthalazinone ether amide) UF membranes. J Membr Sci 2008;311:360–70. [4] Liao C, Yu P, Zhao J, Wang L, Luo Y. Preparation and characterization of NaY/ PVDF hybrid ultrafiltration membranes containing silver ions as antibacterial materials. Desalination 2011;272:59–65. [5] Ge J, Cui Y, Yan Y, Jiang W. The effect of structure on pervaporation of chitosan membrane. J Membr Sci 2000;165:75–81. [6] Dai Y, Jian X, Zhang S, Guiver MD. Thermostable ultrafiltration and nanofiltration membranes from sulfonated poly(phthalazinone ether sulfone ketone). J Membr Sci 2001;188:195–203. [7] Xiangli F, Wei W, Chen Y, Jin W, Xu N. Optimization of preparation conditions for polydimethylsiloxane (PDMS)/ceramic composite pervaporation membranes using response surface methodology. J Membr Sci 2008;311:23–33.

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004

G Model

JTICE-624; No. of Pages 8 8

M. Tan et al. / Journal of the Taiwan Institute of Chemical Engineers xxx (2013) xxx–xxx

[8] Yi S, Su Y, Qi B, Su Z, Wan Y. Application of response surface methodology and central composite rotatable design in optimizing the preparation conditions of vinyltriethoxysilane modified silicalite/polydimethylsiloxane hybrid pervaporation membranes. Sep Purif Technol 2010;71:252–62. [9] Wood H, Sourirajan S. The nature of pores in the skin layer of polymeric reverse osmosis/ultrafiltration membranes. Dev Chem Eng Miner Process 1993;1: 32–41. [10] Al-Abri M, Hilal N. Artificial neural network simulation of combined humic substance coagulation and membrane filtration. Chem Eng J 2008;141:27–34. [11] Al-Abri M, Al Anezi K, Dakheel A, Hilal N. Humic substance coagulation: artificial neural network simulation. Desalination 2010;253:153–7. [12] Madaeni SS, Gheshlaghi A, Rekabdar F. Membrane treatment of oily wastewater from refinery processes. Asia-Pac J Chem Eng 2013;8:45–53. [13] Khayet M, Cojocaru C. Artificial neural network model for desalination by sweeping gas membrane distillation. Desalination 2013;308:102–10. [14] Sexton RS, Dorsey RE, Johnson JD. Toward global optimization of neural networks: a comparison of the genetic algorithm and backpropagation. Decis Support Syst 1998;22:171–85. [15] Sexton RS, Gupta F N D. Comparative evaluation of genetic algorithm and backpropagation for training neural networks. Inform Sci 2000;129:45–59. [16] Morimoto T, De Baerdemaeker J, Hashimoto Y. An intelligent approach for optimal control of fruit-storage process using neural networks and genetic algorithms. Comput Electron Agric 1997;18:205–24. [17] Wang L. A hybrid genetic algorithm-neural network strategy for simulation optimization. Appl Math Comput 2005;170:1329–43. [18] Shen C, Wang L, Li Q. Optimization of injection molding process parameters using combination of artificial neural network and genetic algorithm method. J Mater Process Technol 2007;183:412–8. [19] Madaeni SS, Hasankiadeh NT, Kurdian AR, Rahimpour A. Modeling and optimization of membrane fabrication using artificial neural network and genetic algorithm. Sep Purif Technol 2010;76:33–43. [20] Tavakolmoghadam M, Safavi M. An optimized neural network model of desalination by vacuum membrane distillation using genetic algorithm. Proc Eng 2012;42:106–12. [21] Delgrange N, Cabassud C, Cabassud M, Durand-Bourlier L, Laine JM. Modelling of ultrafiltration fouling by neural network. Desalination 1998;118:213–27. [22] Shetty GR, Chellam S. Predicting membrane fouling during municipal drinking water nanofiltration using artificial neural networks. J Membr Sci 2003;217: 69–86. [23] Chellam S. Artificial neural network model for transient crossflow microfiltration of polydispersed suspensions. J Membr Sci 2005;258:35–42. [24] Liu Q-F, Kim S-H, Lee S. Prediction of microfiltration membrane fouling using artificial neural network models. Sep Purif Technol 2009;70:96–102. [25] Aydiner C, Demir I, Keskinler B, Ince O. Joint analysis of transient flux behaviors via membrane fouling in hybrid PAC/MF processes using neural network. Desalination 2010;250:188–96. [26] Bowen WR, Jones MG, Yousef HNS. Dynamic ultrafiltration of proteins – a neural network approach. J Membr Sci 1998;146:225–35. [27] Richard Bowen W, Jones MG, Welfoot JS, Yousef HNS. Predicting salt rejections at nanofiltration membranes using artificial neural networks. Desalination 2000;129:147–62. [28] Shetty GR, Malki H, Chellam S. Predicting contaminant removal during municipal drinking water nanofiltration using artificial neural networks. J Membr Sci 2003;212:99–112. [29] Baek K, Yang JW. Removal of ferriccyanide using micellar enhanced ultrafiltration (MEUF). Dev Chem Eng Miner Process 2005;13:137–46. [30] Hilal N, Ogunbiyi OO, Al-Abri M. Neural network modeling for separation of bentonite in tubular ceramic membranes. Desalination 2008;228:175–82. [31] Tan M, He G, Li X, Liu Y, Dong C, Feng J. Prediction of the effects of preparation conditions on pervaporation performances of polydimethylsiloxane(PDMS)/ ceramic composite membranes by backpropagation neural network and genetic algorithm. Sep Purif Technol 2012;89:142–6. [32] Albrecht W, Weigel T, Schossig-Tiedemann M, Kneifel K, Peinemann KV, Paul D. Formation of hollow fiber membranes from poly(ether imide) at wet phase inversion using binary mixtures of solvents for the preparation of the dope. J Membr Sci 2001;192:217–30. [33] Aydiner C, Demir I, Yildiz E. Modeling of flux decline in crossflow microfiltration using neural networks: the case of phosphate removal. J Membr Sci 2005;248:53–62. [34] Liu Q, Kim SH. Evaluation of membrane fouling models based on bench-scale experiments: a comparison between constant flowrate blocking laws and artificial neural network (ANNs) model. J Membr Sci 2008;310:393–401.

[35] Razavi SMA, Mousavi SM, Mortazavi SA. Dynamic prediction of milk ultrafiltration performance: a neural network approach. Chem Eng Sci 2003;58: 4185–95. [36] Darwish NA, Hilal N, Al-Zoubi H, Mohammad AW. Neural networks simulation of the filtration of sodium chloride and magnesium chloride solutions using nanofiltration membranes. Chem Eng Res Des 2007;85:417–30. [37] Molga EJ. Neural network approach to support modelling of chemical reactors: problems, resolutions, criteria of application. Chem Eng Process 2003;42: 675–95. [38] Al-Zoubi H, Hilal N, Darwish NA, Mohammad AW. Rejection and modelling of sulphate and potassium salts by nanofiltration membranes: neural network and Spiegler–Kedem model. Desalination 2007;206:42–60. [39] Khayet M, Cojocaru C. Artificial neural network modeling and optimization of desalination by air gap membrane distillation. Sep Purif Technol 2012;86: 171–82. [40] Feng X, Huang RYM. Preparation and performance of asymmetric polyetherimide membranes for isopropanol dehydration by pervaporation. J Membr Sci 1996;109:165–72. [41] Bindal RC, Hanra MS, Misra B.M.. Novel solvent exchange cum immersion precipitation technique for the preparation of asymmetric polymeric membrane. J Membr Sci 1996;118:23–9. [42] Kim I-C, Yoon H-G, Lee K-H. Formation of integrally skinned asymmetric polyetherimide nanofiltration membranes by phase inversion process. J Appl Polym Sci 2002;84:1300–7. [43] Wang G, Zhang S, Yang D, Jian X. Preparation of a novel poly(phthalazinone ether nitrile ketone) ultrafiltration membrane: effect of non-solvent additives. J Funct Polym 2005;18:105–10 [in Chinese]. [44] Hu L, He G, Zhang L, Zhao W, Li Y. Structure characterization and preparation of carbon nitride film using ion beam sputtering technique. Polym Mater Sci Eng 2008;24:169–71+75 [in Chinese]. [45] Sun B. Effect of thermodynamic and kinetic factors on structure and performance of asymmetric membrane formed via L-S phase inversion. Membrane Science and Technology 2011;31:1–11 [in Chinese]. [46] Wu L, Sun J, Wang Q. Poly(vinylidene fluoride)/polyethersulfone blend membranes: effects of solvent sort, polyethersulfone and polyvinylpyrrolidone concentration on their properties and morphology. J Membr Sci 2006;285: 290–8. [47] Rahimpour A, Madaeni SS, Shockravi A, Ghorbani S. Preparation and characterization of hydrophile nano-porous polyethersulfone membranes using synthesized poly(sulfoxide-amide) as additive in the casting solution. J Membr Sci 2009;334:64–73. [48] Cabasso I, Klein E, Smith JK. Polysulfone hollow fibers. I. Spinning and properties. J Appl Polym Sci 1976;20:2377–94. [49] Wienk IM, Olde Scholtenhuis FHA, van den Boomgaard T, Smolders CA. Spinning of hollow fiber ultrafiltration membranes from a polymer blend. J Membr Sci 1995;106:233–43. [50] Cranford RJ, Darmstadt H, Yang J, Roy C. Polyetherimide/polyvinylpyrrolidone vapor permeation membranes. Physical and chemical characterization. J Membr Sci 1999;155:231–40. [51] Xu J, Xu Z-L. Poly(vinyl chloride) (PVC) hollow fiber ultrafiltration membranes prepared from PVC/additives/solvent. J Membr Sci 2002;208:203–12. [52] Arthanareeswaran G, Sriyamuna Devi TK, Mohan D. Development, characterization and separation performance of organic–inorganic membranes: part II. Effect of additives. Sep Purif Technol 2009;67:271–81. [53] Wang D, Li K, Teo WK. Polyethersulfone hollow fiber gas separation membranes prepared from NMP/alcohol solvent systems. J Membr Sci 1996;115: 85–108. [54] Lee K-R, Teng M-Y, Lee H-H, Lai J-Y. Dehydration of ethanol/water mixtures by pervaporation with composite membranes of polyacrylic acid and plasmatreated polycarbonate. J Membr Sci 2000;164:13–23. [55] Kneifel S, Peinemann KV. Preparation of hollow fiber membranes from polyetherimide for gas separation. J Membr Sci 1992;65:295–307. [56] Kim I-C, Lee K-H. Effect of various additives on pore size of polysulfone membrane by phase-inversion process. J Appl Polym Sci 2003;89:2562–6. [57] Li Z, Ren J, Fane AG, Li DF, Wong F-S. Influence of solvent on the structure and performance of cellulose acetate membranes. J Membr Sci 2006;279: 601–7. [58] Yun Y, Tian Y, Shi G, Li J, Chen C. Preparation, morphologies and properties for flat sheet PPESK ultrafiltration membranes. J Membr Sci 2006;270: 146–53. [59] Wang Z, Yu H, Xia J, Zhang F, Li F, Xia Y, et al. Novel GO-blended PVDF ultrafiltration membranes. Desalination 2012;299:50–4.

Please cite this article in press as: Tan, M., et al., Optimization of ultrafiltration membrane fabrication using backpropagation neural network and genetic algorithm. J. Taiwan Inst. Chem. Eng. (2013), http://dx.doi.org/10.1016/j.jtice.2013.04.004