Environ. Sci. Technol. 2008, 42, 7970–7975
Modeling and Optimization of Heterogeneous Photo-Fenton Process with Response Surface Methodology and Artificial Neural Networks M. B. KASIRI, H. ALEBOYEH, AND A. ALEBOYEH* ´ ´ es ´ Traitement des Effluents, Laboratoire de Genie des Proced ´ Ecole Nationale Superieure de Chimie de Mulhouse, Universite´ de Haute Alsace, 3 rue Alfred Werner, 68093 Mulhouse, France
Received May 18, 2008. Revised manuscript received June 24, 2008. Accepted August 11, 2008.
In this study, estimation capacities of response surface methodology (RSM) and artificial neural network (ANN) in a heterogeneous photo-Fenton process were investigated. The zeolite Fe-ZSM5 was used as heterogeneous catalyst of the process for degradation of C.I. Acid Red 14 azo dye. The efficiency of the process was studied as a function of four independent variables, concentration of the catalyst, molar ratio of initial concentration of H2O2 to that of the dye (H value), initial concentration of the dye and initial pH of the solution. First, a central composite design (CCD) and response surface methodology were used to evaluate simple and combined effects of these parameters and to optimize process efficiency. Satisfactory prediction second-order regression was derived by RSM. Then, the independent parameters were fed as inputs to an artificial neural network while the output of the network was the degradation efficiency of the process. The multilayer feed-forward networks were trained by the sets of input-output patterns using a backpropagation algorithm. Comparable results were achieved for data fitting by using ANN and RSM. In both methods, the dye mineralization process was mainly influenced by pH and the initial concentration of the dye, whereas the other factors showed lower effects.
1. Introduction Advanced oxidation processes (AOPs) are efficient methods for treatment of wide range of organic pollutants in wastewater. Hydroxyl radicals (•OH), highly reactive species generated in sufficient quantities by these systems, have the ability to oxidize the majority of the organic pollutants in the industrial effluents (1). Among AOPs, photo-Fenton processes, as a powerful source of •OH from H2O2 in the presence of iron cations and mild reaction conditions, have been used to treat industrial effluents (2-4). But homogeneous photoFenton has some disadvantages such as (i) the tight range of pH in which the reaction proceeds, (ii) the need for recovering the precipitated catalyst after the treatment, and (iii) deactivation by some ion-complexing agents like phosphate anions. An alternative method could be the use of * Corresponding author phone: (+33) (389336800); fax: (+33) (389336805); e-mail:
[email protected]. 7970
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 21, 2008
heterogeneous solid Fenton catalysts, such as transition metals containing zeolites, clays, bentonites (5-7). The use of synthetic zeolites is very promising due to their unique properties such as microporous structure, high BET surface area, and ion exchange capacity, which could give them an advantage over other carriers (8-10). Conventional and classical methods of studying a process by maintaining other factors involved at an unspecified constant level does not depict the combined effect of all the factors involved. This method is also time-consuming and requires a number of experiments to determine optimum levels, which are unreliable (11). A statistical-based technique commonly named response surface methodology (RSM) is a powerful experimental design tool and has been used to optimize and understand the performance of complex systems (12-14). The wastewater treatment applying AOPs is, in general, quite complex. This is caused by the complexity of solving the equations that involve the radiant energy balance, the spatial distribution of the absorbed radiation, mass transfer, and the mechanisms of a photochemical or photocatalytic degradation involving radical species. Since the process depends on several factors, modeling of these multivariate systems is relatively complex. It is evident that the complexity of modeling cannot be solved by simple linear multivariate correlation. Artificial neural networks (ANNs) are now commonly used in many areas of science and engineering and they represent a set of methods that may be useful in solving such problems (15-17). Their ability to recognize and reproduce cause-effect relationships through training, for multiple input-output systems, makes them efficient to represent even the most complex systems (17). The application of ANN analysis to solve the environmental engineering problems (e.g., the difficulties associated with different processes of control, remediation, and destruction of environmental pollutants) has been the subject of numerous of review articles. Recently, ANNs were used for modeling dye removal by advanced oxidation processes. These results have confirmed that ANNs modeling could effectively reproduce experimental data and predict the behavior of the process (18-21). In this work, two methods of modeling, RSM and ANNs, were used to recognize the relationship between inputs, including concentration of the catalyst, molar ratio of initial concentration of H2O2 to that of the dye (H value), initial concentration of the dye, and initial pH of the solution and output variable. In the first step, central composite design using response surface methodology was employed to evaluate the influence of the process key variables on the process efficiency. In the second step, these parameters were fed as inputs to an ANN while the output of the network was degradation efficiency of the process. The importance of each input variable on the variation of output response was determined and compared with the results obtained by RSM.
2. Materials and Methods 2.1. Reagent. C.I. Acid Red 14 was obtained from SigmaAldrich as commercially available dye (50%) and used without further purification. H2O2 (30%w/w) was purchased from Prolabo. Catalyst Fe-ZSM5 was synthesized at 90 °C according to the general procedure described by Patarin et al. (22) and then calcined at 550 °C. 2.2. Photoreactor and Procedures. For UV/Fe-ZSM5/ H2O2 process, irradiations were performed in a batch photoreactor of 0.8 L in volume equipped with an immersion 9 W low-pressure mercury lamp protected with a quartz tube 10.1021/es801372q CCC: $40.75
2008 American Chemical Society
Published on Web 09/23/2008
FIGURE 1. Main effect plot of the parameters on the process efficiency. (Philips, emission: 253.7 nm). The distance between the lamp and the solution was 3 mm, whereas the thickness of the solution surrounding the lamp was 36 mm. Energy of the lamp on the surface of the quartz tubesmeasured by a LUTRON Ultra-Violet radiometerswas 2.5 × 10-3 W cm-2. Solution with the desired concentration of the dye and the load of Fe-ZSM5 in ultra pure water was fed into the reactor. The aqueous solution was magnetically stirred. The pH of the solution was adjusted using dilute hydrochloric acid or aqueous sodium hydroxide solutions. After 30 min of premixing, defined quantity of H2O2 was added to the mixture, and then the lamp was switched on to initiate the reaction. In all experiments, the sample was taken after 90 min and filtered with a 0.45 µm filtration paper. The experiments were done at room temperature. The extent of mineralization of the dye was determined based on total organic carbon (TOC) measurements using a Shimadzu TOC-VCSN analyzer.
3. Experimental Design 3.1. Response Surface Methodology. Response surface methodology was used for the experimental design and optimization. The most popular class of second-order designs called central composite design (CCD) was used for RSM in the experimental design. The CCD is well suited for fitting a quadratic surface, which usually works well for the process optimization (23). The most important parameters, which affect the efficiency of a heterogeneous photo-Fenton process, as described in our previous paper (24), are concentration of the catalyst, molar ratio of concentration of H2O2 to that of the dye (H value), initial concentration of the dye and initial pH of the solution. The main effect of each parameter on the process efficiency was given in Figure 1. Therefore, full factorial central composite design with five replicates at center point was applied as experimental design. Above mentioned parameters were chosen as independent variables and degradation efficiency as output variable or response. Independent variables and their experimental ranges that have been determined by the preliminary experiments are given in Table 1. For statistical calculations, the variables Xi were coded as xi according to the following relationship: X i - X0 xi ) δX
(1)
Where X0 is value of the Xi at the center point and δX presents the step change. Experimental data were analyzed using the response surface regression procedure of a statistical analysis system (NemrodW version 2000 LPRAI, Marseille, France) and fitted to a second-order polynomial model. 3.2. Artificial Neural Network. A computational neural network consists of simple processing units called neurons. Each network consists of artificial neurons grouped into layers
TABLE 1. Experimental Range and Levels of the Process Independent Variables factor independent variable [dye]0 (mg l-1) [catalyst] (g l-1) H value initial pH
xi
range and level -2
-1
0
1
2
X1
10
20
30
40
50
X2 X3 X4
0.25 250 1.50
0.50 375 2.85
0.75 500 4.20
1.00 625 5.55
1.25 750 6.90
and put in relation to each other by parallel connections. The strength of these interconnections is determined by the weight associated with them. For every ANN, the first layer constitutes the input layer (independent variables) and last one forms the output layer (dependent variables). Between them one or more neurons layers called hidden layers can be located. The number of input and output neurons represents effectively the number of variables used in the prediction and the number of variables to be predicted, respectively. The hidden layers act like feature detectors and in theory, there can be more than one hidden layer. Universal approximation theory, however, suggests that a network with a single hidden layer with a sufficiently large number of neurons can interpret any input-output structure (16, 17, 25). The number of neurons in the hidden layer is determined by the desired accuracy in the neural predictions. Hence, it may be considered as a parameter for the neural net design. Most widely used transfer function for the input and hidden layers is the sigmoid transfer function and given by f(x) )
1 1 + e-x
(2)
The linear activation function (eq 3) is used as the output layer activation function (26). f(x) ) x
(3)
In this work, multilayer feedforward ANN with one hidden layer was used. For all data sets sigmoidal activation function in the hidden layer and a linear transfer function in the output node was used. The ANNs was trained using the backpropagation algorithm. All calculations carried out in Matlab mathematical software (version 7) with ANN toolbox.
4. Results and Discussion 4.1. Response Surface Optimization. Full factorial central composite design (CCD) was employed to determine the simple and combined effects of four operating variables on the process efficiency. The following response equation was used to correlate the dependent and independent variables. VOL. 42, NO. 21, 2008 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
7971
Y ) bo + b1x1 + b2x2 + b3x3 + b4x4 + b11x21 + b22x22 + b33x23 + b44x24 + b12x1x2 + b13x1x3 + b14x1x4 + b23x2x3 + b24x2x4 + b34x3x4 (4) where, Y is response variable or degradation efficiency, b0 is constant, b1, b2, b3, and b4 are regression coefficients for linear effects, b11, b22, b33, and b44 are quadratic coefficients and b12, b13, b14, b23, b24, and b34 are interaction coefficients. From experimental observations, it was assumed that the higher order interactions were small relative to the low order interactions, because a system with several process variables is conducted primarily by some of the main effects and low order interactions (27). Therefore, the present work considers only the two-way interactions. The four factors and five levels CCD experimental results for C.I. AR14 dye degradation efficiencies are presented in Table 2. The regression coefficient values, standard deviation, texp., and significance level are given in Table 3. It could be seen that, b1, b2, b3, and b4 as well as the interaction coefficients b13 and b14 are significant at a level less than 5%. The quadratic coefficients b33 and b44 are also significant at the same level. Therefore, the linear effect of all variables (coefficients b1, b2, b3, and b4), the interaction of initial concentration of the dye with concentration of H2O2 and initial pH of the solution (coefficient b13 and b14) and the quadratic effect of concentration of H2O2 (coefficient b33) and initial pH of the solution (coefficient b44) are the most influential parameters. The significance of these quadratic and interaction effects between the variables would have been lost if the experiments were carried out by conventional methods. On the contrary to what is observed for initial concentration of the dye, variations of two other factors, i.e. H value (concentration of H2O2) and initial pH of the solution, have
TABLE 2. A 24 full factorial CCD with five replicates of the centre point [dye]0 [catalyst] h initial run mg L-1 g L-1 value ph -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -2 2 0 0 0 0 0 0 0 0 0 0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
7972
9
-1 -1 1 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1 1 0 0 -2 2 0 0 0 0 0 0 0 0 0
-1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 1 1 1 1 0 0 0 0 -2 2 0 0 0 0 0 0 0
-1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 -2 2 0 0 0 0 0
degradation efficiency (%) exp.
pred.
91.77 81.92 90.07 80.57 90.87 77.50 89.41 76.42 94.71 87.01 92.80 86.01 93.76 81.54 91.12 81.96 95.96 77.58 87.94 86.58 85.71 82.34 71.36 76.21 89.10 90.27 88.71 88.06 89.14
90.511 80.569 88.596 79.829 89.933 76.516 88.317 76.076 92.816 85.334 91.016 84.709 91.733 80.776 90.233 80.451 99.135 79.412 90.883 88.643 88.947 84.110 72.948 79.628 89.056 89.056 89.056 89.056 89.056
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 21, 2008
TABLE 3. Estimated Regression Coefficients and Corresponding texp. and Significance Level coefficient
value
standard deviation
texp.
significance level (%)
b0 b1 b2 b3 b4 b11 b22 b33 b44 b12 b13 b23 b14 b24 b34
89.056 -4.931 -0.560 -1.209 1.670 0.054 0.177 -0.632 -3.192 0.294 -0.869 0.075 0.615 0.029 -0.126
1.00 1.00 1.00 1.00 1.00 1.08 1.08 1.08 1.08 1.00 1.00 1.00 1.00 1.00 1.00
247.26 -29.99 -3.41 -7.36 10.16 0.34 1.12 -4.00 -20.19 1.46 -4.31 0.37 3.05 0.14 -0.63
