Use of artificial neural network for the prediction of bioelectricity production in a membrane less microbial fuel cell

Fuel 117 (2014) 697–703

Contents lists available at ScienceDirect

Fuel journal homepage: www.elsevier.com/locate/fuel

Use of artificial neural network for the prediction of bioelectricity production in a membrane less microbial fuel cell Ali Tardast a, Mostafa Rahimnejad a,⇑, Ghasem Najafpour a, Ali Ghoreyshi a, Giuliano C. Premier b, Gholamreza Bakeri a, Sang-Eun Oh c a b c

Biotechnology Research Lab., Faculty of Chemical Engineering, Noshirvani University, Babol, Iran Sustainable Environment Research Centre (SERC), Faculty of Advanced Technology, 16 University of Glamorgan, Pontypridd CF37 1DL, United Kingdom Department of Biological Environment, Kangwon National University, 200-701 Chuncheon, Republic of Korea

h i g h l i g h t s  This study demonstrates the use of membrane-less microbial fuel cell. 2

 The maximum generated voltage and power were 850 mV and 80.12 mW m

, respectively.

 Artificial neural network was applied for prediction of bioelectricity production from glucose.  Such a membrane-less microbial fuel cell has high scalability and low material cost.

a r t i c l e

i n f o

Article history: Received 17 December 2012 Received in revised form 16 September 2013 Accepted 17 September 2013 Available online 10 October 2013 Keywords: Microbial fuel cell Bioelectricity Power density Artificial neural network

a b s t r a c t Microbial fuel cells (MFCs) are the most recent bioelectrical devices which convert biodegradable organic matters to bioelectricity in presence of active biocatalyst. This system can generate electrons (e) and protons (H+), in which electrons transfer from anode compartment to cathode chamber through an external circuit. MFC architect is one of important factor that effects on MFC performance. In this study, new membrane-less MFC was fabricated. Mixed culture of anaerobic microorganisms was collected from dairy wastewater effluents (Gella, Amol) as active biocatalysts in anode chamber. Initial open circuit voltage was less than 500 mV. Maximum open circuit voltage of 750 mV was achieved after 95 h of operation time. Maximum obtained power density was 80.12 mW/m2. Artificial neural network was applied for the prediction of bioelectricity production from glucose as electron donors. Fabricated network was presented by multilayer perceptron and had a good ability for prediction with high correlation coefficient (R2average-ANN = 0.99). Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction Fossil sources of energy have several disadvantages such as climate change, environmental pollutions, creating more economical crisis as the costs having increasing trends. Fossil sources while abundant are exhaustible and alternative sources of energy are both socio-economically and environmentally attractive. Relatively recent developments in direct generation of electricity from living bacteria, such as the avoidance of toxic mediators and increasing power densities, have advanced the field of Microbial fuel cells (MFCs). MFCs convert low grade biodegradable organic and inorganic materials to bioenergy directly in the form of electricity and the concept has been extended to include reduction reactions at a cathode, using electrons in, e.g. proton reduction to hydrogen ⇑ Corresponding author. Tel.: +98 111 323 4204; fax: +98 111 321 0975. E-mail addresses: (M. Rahimnejad).

[email protected],

[email protected]

0016-2361/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.fuel.2013.09.047

gas [1–4]. MFCs have the ability to produce electricity from waste materials, thus can simultaneously treat wastes such as wastewaters [5,6]. Different biodegradable substrates such as glucose, sucrose, acetate and biomass material, have been used for electricity production in MFC [6]. Microorganisms in an anode compartment can degrade substrate into electron, proton and CO2 [7]. The operating mechanisms of MFCs are well presented in several studies [2,8–10], with the anodic electron and proton generating reactions catalyzed by living microorganisms as part of their metabolic processes. Anion and cation exchange membranes can be used in MFCs to transfer ions between the anode and cathode chambers [11–13]. Several factors affect on MFC performances, such as temperature, substrate concentration, electrode surface area and materials, internal and external impedances, pH and oxidation–reduction potentials in the cathode and anode chambers [3,14]. Recently, a number of investigators have conducted researches to enhance MFC performances [15]. One of the key factors affecting on MFC’s

698

A. Tardast et al. / Fuel 117 (2014) 697–703

yield is the proton exchange membrane. Some studies on MFCs have developed cells to omit the membrane [16–20]. Zhang et al. have replaced the membrane with a glass fiber separator [21]. Liu et al. have removed the membrane in a two chambers of MFC [22] and produced bioelectricity using acetate as the electron donors. Removal of the membrane in a membrane less-MFC (MLMFC) can cause high levels of oxygen intrusion from the aerobic cathode to the anaerobic anode. This phenomena decreases the ML-MFC columbic efficiency [2]. The main objective of this research was to omit membrane in dual chambers of MFC and generate bioelectricity in new configuration of ML-MFC. Also model the power generated by fabricated ML-MFC. For prediction of power artificial neural network was carried out. The used model had good ability for prediction of ML-MFC behavior. 2. Materials and methods 2.1. MFC fabrication and operation A ML-MFC was fabricated from Plexiglas and consisted of two separate chambers. The empty bed volume of each chamber was 1000 cm3 (10  10  10 cm). Each chamber has four baffles which facilitate mixing of the feed solution and increase the contact time between microorganism and electrode surface. The anodic and cathodic chambers were connected through a valve which could control flow between the chambers. Carbon paper electrodes coated with Pt were used as anode and cathode electrodes and were connected in an external circuit by copper wire. Experimental data were recorded using an online data logger (constructed by Biotechnology Group, Babol University of Technology, Iran). Oxygen was introduced to the cathode chamber by sparging with air. Fig. 1 shows a schematic diagram of the MFC with its auxiliary equipments. A characteristic of this design is that it effectively eliminates oxygen diffusion into anode chamber. The prepared substrate was transferred by a peristaltic pump (THOMAS, Germany) from the feed tank to the anode chamber and subsequently passed through the interconnecting tube and valve to the cathode chamber. ML-MFC was inoculated with anaerobic sludge collected from the dairy industry (Gella, Amol). The anodic media contained glucose, yeast extract, peptone and NH4Cl: 5, 3, 3 and 0.3 g/l, respectively. Experiments were conducted at ambient temperature, 26–30 °C. The pH of the media was initially adjusted to 6.5 using phosphate buffer solution and the inoculums were intro-

duced into the media at ambient temperature. The organisms were fully grown for the duration of 27 h. Substrate consumption was calculated on the basis of remaining sugars in the cultured media. Electron acceptor as an effective parameter on power generation at cathode compartment was analyzed. Optical density was measured using a spectrophotometer (UNICO, 2100 SERIES, USA) at wavelength of 540 nm and the calibration curve was prepared. 2.2. Chemical and analysis All chemicals and reagents used for the experiments were analytical grades and supplied by Merck (Darmstadt, Germany). The pH meter (Model HANA 211, Romania) was employed to measure pH values of the aqueous phase. The surface images of the graphite plate electrode were obtained using a Scanning Electronic Microscope (SEM) (Supra 55vp-Zeiss, Germany). The images of the surface of the graphite electrode were taken before and after the experimental run. The sample specimen size for SEM analysis was 1 cm  1 cm. SEM images were used to demonstrate the physical characteristics of the electrode surface and to examine the morphology of the bacteria on the anode surface. 2.3. Artificial neural network layers and validation Neural network constitutes a branch of artificial intelligence which has recently undergone rapid evolution and progress [23]. Artificial neural network (ANN) utilizes interconnected mathematical nodes or neurons to form a network that can model complex functional relationship [24]. Its development started in the 1940s to help cognitive scientists understand the complexity of the nervous system. Basically, ANN is numerical structures inspired by the learning process in the human brain [25] and also can be considered as potential alternative for predicting the performance of waste treatment system [26,27]. Neural networks consist of three basic layers such as input, hidden and output layers. Some of ANNs are more popular such as multi-layer perceptron and radial basis function [28]. It has been showed that more complex nonlinear function can be modeled by a multilayer feed forward neural network [29]. In this study, a multilayer feed forward neural network was used which consists of an input, hidden and output layers. These used layers are shown in Fig. 2. That is explained the general topology of a multilayer feed forward neural network. Neurons in each layer connect together by a weight coefficient that called Wij and Wjk [32]. There is a transfer function which change inputs

Fig. 1. Schematic diagram of ML-MFC with auxiliary equipment.

699

A. Tardast et al. / Fuel 117 (2014) 697–703

Fig. 2. Artificial neural network structure.

to output. Before using an artificial neural network, it is necessary to train network. Training is a method for calculation of weight and bias. During training, network works with iterative method until produces a new output. There are different kinds of training methods, in which back propagation is a common method. At the beginning of training process, initial weights are randomly given to connections. Inputs are inserted into input layer and then move forward through the hidden layer of neurons to the output layer. At the end, outputs would be compared with real outputs. Changes of weight coefficient can decrease required time and calculate errors. After that the neural network was prepared, the network is ready for prediction. Before using any method for training an ANN has to normalize input and output. So input and output data were normalized between 0 and 1 by the following equation:

Normalize value ¼

ðActualÞvalue  ðminimumÞActual value ðmaximumÞActual value  ðminimumActual value Þ

g is learning rate, l is momentum coefficient, X l1 is the input from i L  1th interlayer and b is bias. dlj ¼ ðX Lj  Y j ÞX Lj ð1  X Lj Þ

ð8Þ

For the neuron in hidden layer lowing equation:

dlj ¼ X lj ð1  X lj Þ

dij

was calculated from the fol-

r X dl1  W lþ1 k kj

ð9Þ

k¼1

The above equations are based on sigmoid function.X lj ð1  X lj Þ and ðX lj  yi ÞX lj would be different for other transfer functions [30–32]. The simulation and curve fitting was carried out by MATLAB 7.7.0 (MathWorks, US) software for programming, training, validation and testing of the network. 3. Results and discussion

ð1Þ Training is an iterative process that optimizes weights and biases. Inputs data were divided into two parts which 70% of data was used for training and 30% data was applied for validation. Back propagation was used to train the ANN. As mentioned above, consideration of a multilayer perceptron (MLP) which consisted of L interlayers for each interlayer like l, Ni node and Ni  Ni1 connections with weight W 2 RNiNi1. Ni and Ni1 are number of nodes in interlayer l and l  1. Wji defines the connection between node j of layer l to node i of layer l  1. In each interlayer l and neuron j, weighted input values are integrated as per following equation: N l1 X W lji  X l1 I

ð2Þ

i¼1

In the next step, the transfer function would be used to generate X lj .

X lj ¼ hðulj Þ ¼ h

(N ) l1 X W lji  X l1 i

6

ð3Þ

1

i¼1

0.9

1 hðuÞ ¼ 1 þ eu

ð4Þ

In each interlayer, a weight W lji at iteration (t) will be changed from its previous value (t  1) according to following equation:

W lji ¼ W lji ðt  1Þ þ DW lji ðtÞ

DW lji

dlj

¼g 

X 1l j

þl

ðpreviousÞ DW lji

ð5Þ

¼g

dlj

þl

lðpreviousÞ Dbji

0.8 0.7

4

0.6 3

0.5 0.4 Glucose consumption Growth curve

2

0.3 0.2

1

0.1

ð6Þ

where DW iji represents the weight change which can be calculated by delta rule defined as follows: l Dbji

5

Glucose consumption (g/l)

One of the most common functions used in back propagation training method is a differentiable sigmoid function such as

ð7Þ

0

0 0

5

10

15

20

25

30

Time (h) Fig. 3. Growth of mixed culture in anaerobic condition.

Optical density at 540 nm

u¼

Mixed culture of microorganisms was used as active biocatalyst at anode surface. Growth curve and glucose consumption for the microorganisms in an anaerobic condition was studied. Maximum absorbance was obtained at 27 h after incubation. With pure culture of microorganisms, the maximum absorbance was observed after 16 h of incubation time [33]. In the anaerobic MFC chamber, mixed culture was applied for fully glucose consumption and bioelectricity production. The obtained results indicated that microorganisms had great ability to grow under anaerobic condition (Fig. 3). Fig. 4 shows the open circuit voltage of fabricated membraneless MFC for a long operation and duration period of 300 h. Anode chamber was fed with medium then microorganisms in anodic chamber were consumed medium and produced electrons and protons. The generated protons with filling of anodic chamber moved to cathode compartment. Maximum generated voltage

700

A. Tardast et al. / Fuel 117 (2014) 697–703

was about 750 mV and it was rarely stable in the period of operation time. 3.1. Effect of electron acceptoron power generation Polarization curve and power density versus current density when oxygen was final electron acceptor in cathodic chamber are presented in Fig. 5. Maximum obtained power and current density with the final electron accepter were 13.9 mW/m2 and 42.2 mA/ m2, respectively. The produced bioelectricity was too low. In order to achieve higher power, hexacyanoferrate, potassium dichromate and potassium permanganate were selected as final electron acceptor in cathode compartment of MFC. Those chemicals can considerably improve the cathode performance. These materials can increase reduction reaction on the surface of cathode electrode. Cathode losses can also be minimized using this set of electron acceptor. In aqueous solution the standard oxidation–reduction potential (E0) for the half-cell potassium permanganate is 1.7 V. This potential is higher than the redox potential for hexavalent chromium (E0 = 1.33 V) and ferricyanide (E0 = 0.43 V). Table 1 shows oxidation–reduction potential and reduction reaction for permanganate, hexavalent chromium and hexacyanoferrate. Also several concentrations of different electron acceptors were randomly selected and the generated power data are presented in Table 2. 3.2. Prediction of power and current density with artificial neural network (ANN) Modeling of MFC by ANN is a new approach. Different parameters affect on MFC power such as temperature, pH and electron

900

Voltage (mv)

800

PN  MSE ¼

i¼1

yexp  ycal N

2 ð10Þ

Table 1 Three electron acceptor with their reaction and reduction potential. Oxidant

Reaction

Potential

Kmno4 K2Cr2O7

+  MnO 4 + 4H + 3e ? MnO2 + 2H2O

E0 = 1.7 E0 = 1.33

K3 [Fe(CN)6]

4  FeðCNÞ3 6 + e ? FeðCNÞ6

+  3+ Cr2 O2 + 7H2O 7 + 14H + 6e ? 2Cr

E0 = 0.43

700 600 Table 2 Maximum power, current density and voltage obtained with presence of electron acceptor.

500 400 300 0

50

100

150

200

250

300

Time (h)

700

16

600

14 12

500

10 400 8 300

6 voltage v.s current density

200

4

100

power density v.s current density

Power density (mW/m2)

Fig. 4. Produced open circuit voltage by ML-MFC.

Voltage (mv)

acceptor concentration so these parameters were selected as input variable for network and power and current density was selected as output for network. Among different transfer function available in MATLAB log sigmoid function was selected for all neurons due to its more desired prediction performance than other transfer functions. The log sigmoid function is bounded between 0 and 1, so the input and output data should be normalized to the same range. In other words the logarithm sigmoid transfer function gives scaled output in this range (0–1). Table 3 shows amounts of electron acceptor in different pH and temperature. Accuracy of ANN depends on numbers of experiments; so variations of pH, temperature and electron acceptor were selected for prediction. A high degree of correlation between actual and predicted power density was obtained. The first four sets of data presented in Table 3 are shown in Fig. 6(A–D). R2 values for these data were 0.99, 0.99, 0.98 and 0.99, respectively. With very high accuracy, the predicted and actual powers are close together. The similar results for the next six experiments were obtained (The related data are not shown). Mean square error (MSE), R and R2 are three essential factors for comparison between actual experimental data and predicted results, so MSE, R and R2 are calculated by:

2 0

0 0

10

20

30

40

50

60

70

Current density (mA/m2) Fig. 5. Generated voltage and power density versus current density before use of electron acceptor.

Electron acceptor

Concentration (g/l)

Pmax (mW/m2)

Imax (mA/m2)

KMnO4 KMnO4 KMnO4 K3 [Fe(CN)6] K3 [Fe(CN)6] K3 [Fe(CN)6] K2Cr2O7 K2Cr2O7 K2Cr2O7 K2Cr2O7

0.01 0.03 0.05 0.02 0.03 0.04 0.01 0.03 0.05 0.07

64.095 79.18 80.129 32.95 36.401 38.43 14.48 14.71 14.90 14.88

135.0477 149.663 150.557 72.407 76.095 97.11 53.66 56.44 54.43 54.32

Table 3 Experimental conditions for artificial neural network. Experiment

Electron acceptor

Temperature (°C)

pH

Concentration (g/l)

1 2 3 4 5 6 7 8 9 10

KMnO4 KMnO4 KMnO4 K3 [Fe(CN)6] K3 [Fe(CN)6] K3 [Fe(CN)6] K2Cr2O7 K2Cr2O7 K2Cr2O7 K2Cr2O7

30 28 27 25 26 28 28 27 28 28

7 6.8 6.7 6.5 6 6.7 6.5 6.7 6.5 6.5

0.01 0.03 0.05 0.02 0.03 0.04 0.01 0.03 0.05 0.07

701

A. Tardast et al. / Fuel 117 (2014) 697–703

A

B

70

Actual power density

Actual power density

60 50 40 30 20

y = 1.0058x - 0.271 R² = 0.9999

10

70 60 50 40 30 y = 0.9929x + 0.2995 R² = 0.9998

20 0

0 -10

20

40

60

0

80

Predicted power density

D

90

Actual power density

60 50 40 30 y = 1.0318x -6.402 R² = 0.9893

20 10 0 - 20

40

60

80

100

45 40

70

- 10 0

20

Predicted power density

80

Actual power density

80

10

0

C

90

20

40

60

80

35 30 25 20 15

y = 1.0272x -1.8912 R² = 0.9949

10 5 0 -5 0

100

10

20

30

40

50

Predicted power density

Predicted power density

Fig. 6. Actual power versus predicted power for experiment 1, 2, 3 and 4 (neuron 3).

P Þðyexp  yexp Þ ðy  y R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2ffi P  exp P  Þ2 y  yexp ðy  y

ð11Þ

 Pn    predict yexp  y exp 2 ypredict  y R2 ¼ Pn i¼1   2 Pn     i¼1 ypredict  ypredict i¼1 yexp  yexp

ð12Þ

The network is tested with several numbers of neurons to find the optimal number of neurons at hidden layer by observing the mean square error. For the first time, 2 neurons were selected and results of training, validation and testing obtained by neural network with other neurons were compared. Such selection was based on MSE and R square and convergence time for training and testing in low epoch, so the first 2 neuron and 3, 4 and 5 neuron were used. By comparison between these neurons showed that maximum R-square was obtained for 3 neurons. Maximum R-square (9.8868  101) for testing was obtained in 3 neuron. Results showed that 3-neuron is optimal neuron for network. Table 4 shows mean square error and R2 for all experimental and predicted data, respectively. One of the more important MFC features is the curve of power versus current density. ANN could predict power and current density with a high accuracy in fitting experimental data. Fig. 7(A–D) shows power densities versus current density for the first four

experiments in Table 3. Prefect fitting between obtained data and predicted results were also obtained from the others experiments (data was not presented). The surface morphology of used electrode after experiment was evaluated by SEM. A piece of the electrode (1  1 cm) was analyzed by SEM before and after using as electrode at anode compartment and the obtained images are presented in Fig. 8. The obtained images demonstrated that microorganisms were grown on the carbon surface as attached biofilm. Some clusters of microorganism were detected on the anode surface.

4. Conclusion The use of microorganisms as active biocatalyst is an interesting point in MFCs. New configuration of membrane-less MFC was fabricated and used for electricity generation. Several electron acceptors was selected and used to increase generated power. Maximum power (80.129 mW/m2) and current densities (150.55 mA/m2) were obtained when permanganate was used as final electron accepters in cathode chamber. Also, ANN was implemented for perdition of generated power in the fabricated membrane-less MFC. Prediction of power and current density by several inputs (pH, temperature, concentration of final electron acceptor) was

Table 4 Results of artificial neural network prediction. Neuron

2 3 4 5

R2

MSE Train

Validation

Testing

Train

Validation

Testing

3.93045  104 3.61619  105 1.5434  103 7.8808  104

2.4829  104 3.8079  102 2.1523  103 2.0249  103

4.89353  104 2.387  103 9.7002  103 5.25936  103

9.9829  101 9.998  101 9.9236  101 9.9619  101

9.992  101 9.8302  101 9.89905  101 9.9047  101

9.857  101 9.8868  101 9.5563  101 9.776  101

702

A. Tardast et al. / Fuel 117 (2014) 697–703

A

B

70

80

50 Actual power v.s Actual current density

40 30 20

Predicted power v.s Predicted current density

10

Power (mW/m2)

60

Power(mW/m2)

90

70 60 50

Actual power v.s Actual current density

40 30

Predicted power v.s Predicted current density

20 10 0

0 0

50

100

150

0

50

100

Current (mA/m2)

D

90 80 70 60 50

Actual power v.s Actual current density Predicted power v.s Predicted current density

40 30 20 10 0

Power density (mW/m2)

Power density (mW/m2)

C

150

200

Current (mA/m2) 45 40 35 30 25

Actual power v.s Actual current density Predicted power v.s Predicted current density

20 15 10 5 0

0

50

100

150

200

0

Current density (mA/m2)

50

100

150

Current density (mA/m2)

Fig. 7. Power density versus current density for experiments 1, 2, 3 and 4.

Fig. 8. SEM images of anode surface. (A) Graphite electrode before use as anode electrode. (B) Attached microorganisms on electrode surface at the end of experiment.

tested. The obtained data indicated that ANN could predict power density versus current density at high accuracy. Maximum R2 value (9.8868  101) and MSE (2.387  103) were obtained for 3 neurons. Acknowledgement The authors wish to acknowledge Biotechnology Research Center, Noshirvani university of Technology, Babol, Iran for the facilities provided to accomplish the present research.

References [1] Rabaey K, Verstraete W. Microbial fuel cells: novel biotechnology for energy generation. Trends Biotechnol 2005;23:291–8. [2] Logan BE, Hamelers B, Rozendal R, Schröder U, Keller J, Freguia S, et al. Microbial fuel cells: methodology and technology. Environ Sci Technol 2006;40:5181–92. [3] Larrosa-Guerrero A, Scott K, Head I, Mateo F, Ginesta A, Godinez C. Effect of temperature on the performance of microbial fuel cells. Fuel 2010;89:3985–94. [4] Mokhtarian N, Rahimnejad M, Najafpour G, Daud WRW, Ghoreyshi A. Effect of different substrate on performance of microbial fuel cell. Afr J Biotechnol 2012;11:3363–9.

A. Tardast et al. / Fuel 117 (2014) 697–703 [5] Ieropoulos IA, Greenman J, Melhuish C, Hart J. Comparative study of three types of microbial fuel cell. Enzyme Microb Technol 2005;37:238–45. [6] Rahimnejad M, Najafpour G, Ghoreyshi A, Shakeri M, Zare H. Methylene blue as electron promoters in microbial fuel cell. Int J Hydrogen Energy 2011;36:13335–41. [7] Mohan Y, Manoj Muthu Kumar S, Das D. Electricity generation using microbial fuel cells. Int J Hydrogen Energy 2008;33:423–6. [8] Rabaey K, Boon N, Höfte M, Verstraete W. Microbial phenazine production enhances electron transfer in biofuel cells. Environ Sci Technol 2005;39:3401–8. [9] Katz E, Shipway AN, Willner I. Biochemical fuel cells. In: Handbook of fuel cells-fundamentals, technology, and application. New York: John Wiley & Sons, Ltd; 2003. [10] Tardast A, Najafpour GD, Rahimnejad M, Amiri A. Bioelectrical power generation in a membrane less microbial fuel cell. World Appl Sci J 2012;16:179–82. [11] Ren Z, Ward TE, Regan JM. Electricity production from cellulose in a microbial fuel cell using a defined binary culture. Environ Sci Technol 2007;41:4781–6. [12] Najafpour G, Rahimnejad M, Ghoreshi A. The enhancement of a microbial fuel cell for electrical output using mediators and oxidizing agents. Energy Source A 2011;33:2239–48. [13] Kim JR, Cheng S, Oh SE, Logan BE. Power generation using different cation, anion, and ultrafiltration membranes in microbial fuel cells. Environ Sci Technol 2007;41:1004–9. [14] Kargi F, Eker S. Electricity generation with simultaneous wastewater treatment by a microbial fuel cell (MFC) with Cu and Cu–Au electrodes. J Chem Technol Biotechnol 2007;82:658–62. [15] Logan BE. Exoelectrogenic bacteria that power microbial fuel cells. Nat Rev Microbiol 2009;7:375–81. [16] Jang JK. Construction and operation of a novel mediator-and membrane-less microbial fuel cell. Process Biochem 2004;39:1007–12. [17] Liu H, Logan BE. Electricity generation using an air-cathode single chamber microbial fuel cell in the presence and absence of a proton exchange membrane. Environ Sci Technol 2004;38:4040–6. [18] Lovley DR. Microbial fuel cells: novel microbial physiologies and engineering approaches. Curr Opin Biotechnol 2006;17:327–32. [19] You S, Zhao Q, Zhang J, Jiang J, Wan C, Du M, et al. A graphite-granule membrane-less tubular air-cathode microbial fuel cell for power generation under continuously operational conditions. J Power Sources 2007;173:172–7.

703

[20] Zhan YL, Zhang P, Yan G, Guo SH. Constructing and operating of mediatorand membrane-less microbial fuel cell. J Chem Eng Chin Univ 2008;22: 177–81. [21] Zhang X, Cheng S, Huang X, Logan BE. The use of nylon and glass fiber filter separators with different pore sizes in air-cathode single-chamber microbial fuel cells. Energy Environ Sci 2010;3:659–64. [22] Liu H, Cheng S, Huang L, Logan BE. Scale-up of membrane-free single-chamber microbial fuel cells. J Power Sources 2008;179:274–9. [23] Eldin NN, Senouci AB. Condition rating of rigid pavements by neural networks. Can J Civ Eng 1995;22:861–70. [24] Sha W, Edwards K. The use of artificial neural networks in materials science based research. Mater Des 2007;28:1747–52. [25] Mjalli FS, Al-Asheh S, Alfadala H. Use of artificial neural network black-box modeling for the prediction of wastewater treatment plants performance. J Environ Manage 2007;83:329–38. [26] Chairez I, García-Peña I, Cabrera A. Dynamic numerical reconstruction of a fungal biofiltration system using differential neural network. J Process Control 2009;19:1103–10. [27] Rene ER, Veiga MC, Kennes C. Experimental and neural model analysis of styrene removal from polluted air in a biofilter. J Chem Technol Biotechnol 2009;84:941–8. [28] Movagharnejad K, Nikzad M. Modeling of tomato drying using artificial neural network. Comput Electron Agric 2007;59:78–85. [29] Cybenko G. Approximation by superpositions of a sigmoidal function. Math Control Signals Syst 1989;2:309–14. [30] Mohanty S. Estimation of vapour liquid equilibria of binary systems, carbon dioxide-ethyl caproate, ethyl caprylate and ethyl caprate using artificial neural networks. Fluid Phase Equilib 2005;235:92–8. [31] Tardast A, Rahimnejad M, Najafpour G, Pirzade K, Mokhtarian N. Prediction of bioelectricity production by neural network. J Biotechnol Pharm Res 2012;3:62–8. [32] Mehdizadeh B, Movagharnejad K. A comparison between neural network method and semi empirical equations to predict the solubility of different compounds in supercritical carbon dioxide. Fluid Phase Equilib 2011;303:40–4. [33] Rahimnejad M, Mokhtarian N, Najafpour G, Daud W, Ghoreyshi A. Low voltage power generation in abiofuel cell using anaerobic cultures. World Appl Sci J 2009;6:1585–8.