Liquid Activated

Dec 15, 2007 - A three-layer feed-forward neural network was constructed and tested to analyze the kinetic dye uptake of a batch activated carbon adso...
14 downloads 0 Views 356KB Size
Download PDF
486

Ind. Eng. Chem. Res. 2008, 47, 486-490

Neural Network Modeling and Simulation of the Solid/Liquid Activated Carbon Adsorption Process K. Vasanth Kumar,*,† K. Porkodi,‡ R. L. Avila Rondon,§ and F. Rocha† Departmento de Engenharia Quı´mica, Faculdade de Engenharia da UniVersidade do Porto, Rua Dr. Roberto Frias, 4200-465 Porto, Portugal; CIQ-UP, Department of Chemistry, Faculty of Science, UniVersidade do Porto, Rua do Campo Alegre 687, 4169-007 Porto, Portugal; and Centro de Estudios CAD/CAM, Facultad de Ingenieria, UniVersidad de Holguin, GaVeta Postal Nr. 57, Holguin 80100, Holguin, Cuba

A three-layer feed-forward neural network was constructed and tested to analyze the kinetic dye uptake of a batch activated carbon adsorption process. The operating variables studied are the contact time, initial dye concentration, agitation speed, temperature, initial solution pH, activated carbon mass, and volume of the dye solution treated. The studied operating variables were used as the input to the constructed neural network to predict the dye uptake at any time as the output or the target. The constructed network was found to be precise in modeling the rate of dye uptake for the operating conditions studied. The constructed neural network was found to be highly precise in predicting the dye uptake rate for the new input data, which are kept unaware of the trained neural network showing its applicability to determine the reaction rate for any operating conditions. 1. Introduction Activated carbon adsorption processes are found to be effective processes for the treatment of various pollutants from their aqueous solutions. The adsorption processes are mainly due to the two steps that include the external mass transfer of solute onto the surface of the adsorbent by film diffusion followed by the intraparticle diffusion.1 Multiple steps that include the intraparticle diffusion in macropores, mesopores, and micropores are reported for the sorption of basic dye by sphagnum peat moss.2 The nature of the sorption process will depend on physical or chemical characteristics of the adsorbent systems and also on the system conditions. The solid/liquid adsorption processes irrespective of the adsorbate/adsorbent are usually modeled using the mechanistic or empirical based kinetic expressions. Though the kinetic models, such as the external mass transfer model,3 intraparticle diffusion model,4 Lagergren first-order kinetics,5 and pseudo-second-order kinetics,6 are found to be excellent in representing the kinetics of the sorption process, they can represent well only for a particular operating condition studied. Recently, a pseudo-second-order kinetic expression of Ho7 was found to represent well the kinetics of any adsorption system. However, this model was good enough only in representing the kinetics of the solute uptake process for the experimental conditions in which there is a single operating variable. However, any attempt to make a relationship between the amount of solute uptake with all the operating variables, namely, initial dye concentration, adsorbent mass, initial solution pH, agitation speed, volume of solution treated, and contact time, using the kinetic or mechanistic based models is impossible. Currently, artificial neural networks (ANNs) are found to be excellent options for solving these types of complex issues. ANNs were found to be successfully applied in many fields, * To whom correspondence should be addressed. E-mail: [email protected]. Phone: 22 508 1678. Fax: +351 22 508 1449. † Faculdade de Engenharia da Universidade do Porto. ‡ Faculty of Science, Universidade do Porto. § Universidad de Holguin.

which include character recognition, speech recognition, image processing, and stock performance prediction.8 In chemical engineering, ANNs were found to be successfully applied to predict adsorption equilibrium of solid/liquid systems,9 activity coefficients of aromatic organic compounds,8 and solubility of proteins.10 To our knowledge, no studies have been reported so far reporting the applicability of artificial neural networks in predicting the kinetics of solid/liquid adsorption systems considering all of the operating variables of the adsorption system. Artificial neural networks are used to correlate the complex relationship between the input and output of any process, irrespective of the physical meaning of the system. ANN consists of an input layer and an output layer connected by several nodes. In the present study, a feed-forward or backpropagation network with multiple layers was constructed. A Levenberg-Marquardt’s optimization was used to train the ANN. The constructed network was tested with the new data, which were kept unaware of the neural network in order to check the applicability of the network in predicting the dye-uptake rate for new experimental conditions. 2. Experimental Section The dye used in the present study, auramine O, was obtained from Central Drug House, Mumbai, India. Synthetic stock dye solutions were prepared by dissolving 1 g of dye powder in 1 L of distilled water. All working solutions of desired initial concentration were prepared from the stock solution by subsequent dilution. The structure of the dye auramine O is given by

The powdered activated carbon used in the present study was obtained from E-Merck Limited, Mumbai, India. The obtained activated carbon was directly used as adsorbent without any pretreatment. Sorption kinetics experiments were carried out using mechanically agitated overhead laboratory stirrers at different initial

10.1021/ie071134p CCC: $40.75 © 2008 American Chemical Society Published on Web 12/15/2007

Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 487 Table 2. Properties of Activated Carbon Used pore volumea,d surface areaa,d pore number fractiona,d bulk particle densitya apparent densitya total surface areaa total intruded volumea total interparticle porositya total intraparticle porositya true densityb theoretical porosityb BET surface areac microporous volume (BET)c nonmicroporous specific surface area (BET)c

7.030 × 10-1 cm3/g at diam ) 1.496 µm 5.023 × 10-1 cm3/(µm g) at diam ) 2.834 µm 4.872 × 10-1 cm3/g at diam ) 2.183 µm 1.143 m2/g at diam ) 1.496 µm 1.036 m2/(µm g) at diam ) 3.693 × 10-1 µm 1.301 m2/g at diam ) 1.306 µm 3.887 × 10-3 at diam ) 1.496 µm 7.892 × 101 at diam ) 2.928 × 10-1 µm 5.020 × 10-1 at diam ) 4.511 × 10-1 µm 0.5530 g/cm3 0.5530 g/cm3 2.6029 m2/g 0.9741 cm3/g 2.29% 53.87% 1.7470 g/cm3 68.35% 1000.1 m2/g 0.38 cm3/g 130 m2/g

a Determined by mercury porosimetric analysis. b Determined by helium pycnometer. c N2 adsorption using t-method. d Mean, mode, and median, respectively.

3. Characterization of Adsorbent

Figure 1. (a) SEM image of activated carbon (200×); (b) SEM image of activated carbon (1000×). Table 1. Range of Operating Variables Used to Train the Network operating variable

range

initial dye conc, mg/L adsorbent mass, g/1.5 L initial solution pH agitation speed, rpm irradiation time, min

85, 100, 140, 170, 180, 200 0.3, 0.6, 1.2, 1.8 3, 4, 5, 6, 7, 8 800, 700, 600, 500, 400 0, 1, 2, 3, 4, 5, 10, 20, 25, 30, 45, 50, 60, 90, 120 305, 313, 323, 333

temperature, K

dye concentrations. The effect of dye concentration on the adsorption rate was estimated by agitating 1.5 L of dye solution of known initial dye concentration with 0.3 g of activated carbon in 2 L beakers at room temperature (32 °C) at the desired solution pH. Unless specified, all the experiments were carried out at a solution pH of 8 and at a constant agitation speed of 800 rpm. Samples of 2.5 mL were pipetted out using a 10 mL syringe filter at different time intervals. The collected samples were then centrifuged, and the concentration in the supernatant solution was analyzed using UV spectrophotometer. All the kinetic experiments were predesigned and operated for a fixed operating line of 1.5 L/0.3 g. The range of operating variables studied in the present study to train and test the neural network is given in Table 1.

Some of the specifications of the activated carbon used in the present study as supplied by the manufacturer are given by the following: substances soluble in water e 1%; substances soluble in HCl e 3%; Cl e 0.2%; and SO42- e 0.2%; heavy metals as lead (Pb) e 0.005%, as iron (Fe) e 0.1%, and incomplete carbonization ) passes test; methylene blue adsorption e 180 mg/g; loss on drying e 10%; and residue on ignition e 5%. The surface morphology of the carbon particles is characterized by scanning electron microscopy (SEM) analysis. The SEM images at different magnifications are given in parts a and b of Figure 1, which show the surface texture and porosity of the activated carbon. From parts a and b of Figure 1, it can be observed that the shape of the activated carbon is irregular and porous. The physical characteristics of the commercial activated carbon used in this study are given in Table 2. Table 2 lists the pore-size distribution of activated carbon used obtained from the mercury intrusion porosimetry analysis. The true density and percentage of theoretical porosity of activated carbon as determined with helium pycnometry were 1.747 g/cm3 and 68.35%, respectively. The surface tension of the mercury was assumed to be 480 mN/m, and the contact angle of mercury with activated carbon was assumed to be 140. Table 2 confirms that the activated carbon used in the present study is highly porous. The Brunauer-Emmett-Teller (BET) surface area for the commercial activated carbon used in the present study was determined by nitrogen adsorption method and was found to be 1000.1 m2/g. 4. Neural Network Modeling In the present study, a multiple-layer feedforward network with two layers was constructed. Multiple-layer networks can approximate any function very well.11 Feedforward ANN allows signals to flow only in one direction, i.e., from input to output. The feedforward ANN adjusts the transfer function that is associated with the inputs and the outputs. In the present study, initially a network with five hidden layers was constructed and was trained to simulate the dye uptake process for various operating conditions. The detailed structure of the network and

488

Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

Figure 2. Structure of the constructed two-layer network and flow of information within the network.

Figure 3. Training strategy of the constructed feed-forward artificial neural network.

the training strategy of the constructed neural network are shown in Figure 2 and Figure 3, respectively. Figure 2 shows the feedforward network with one hidden layer. P1 is the input vector to the hidden layer, and W1 and b1 represent the weight and bias of the hidden layer, respectively. The information from the hidden layer is transferred to the output layer as shown in Figure 2. The term P2 represents the output vector and can be determined from the weight W2 and bias b2 of the output layer. In the present study, a tansig function and a purelin function were used as the propagation functions in the hidden layer and in the output layer, respectively. The training strategy of the network is shown in Figure 3. As shown in Figure 3, the input vectors and the corresponding output vectors are used to train the network until it approximates the propagation function. Thus, the bias and the weights can be obtained from the training procedure, which is based on the experimental data. In the present study, the contact time, initial dye concentration, agitation speed, temperature, initial solution pH, activated carbon mass, and volume of the dye solution were treated as the input vector and the corresponding dye uptake was defined as the output vector to train the neural network. The neural network toolbox Version 4 of MATLAB, Mathworks, Inc., was used for simulation. The experimental conditions and the corresponding experimentally determined dye uptake rate were set as the input and the target vectors, respectively. The input vectors (experimental conditions) and the target vector were normalized before the training process such that they fall in the interval of 0-1, so that their standard deviation and mean will be below the value of 1. The neural network was trained in a batch mode. The training was made using the Levenberg-Marquardt’s training strategy. The incorporation of Marquardt’s algorithm into the back-propagation algorithm to train feedforward networks was explained elsewhere.12 The training of the neural networks by the LevenbergMarquardt algorithm is sensitive to the number of neurons in the hidden layer. The more the number of neurons, the better is the performance of the neural network in fitting the data. However, too many neurons in the hidden layer may result in overfitting. During the training process, the number of neurons in the hidden layer was changed while optimizing the transfer function for the given input and output vectors. In order to avoid the problems due to the

Table 3. Details of the Trained Neural Network Used to Predict the Dye Uptake Kinetics of Solid/Liquid Adsorption Process type layer 1 layer 2 layer 3 number of data used for training function in hidden layer function of output layer

value/comment 6 neurons 7 neurons 1 neuron 272 tansigmoid linear

overfitting, a Bayesian regularization technique in combination with the Levenberg-Marquardt’s algorithm was used during the ANN training process. The Bayesian algorithm works best when the network’s input and output are scaled within the range of -1 to +1.13 After several trials, the neural network with 10 neurons in the hidden layer was found to be excellent in representing the dye uptake process. In the hidden layer, initially three types of transfer functions, namely, the exponential sigmoid, tangent sigmoid, and linear functions, were tested while training the neural network. The linear function was used at the output layer. A tansigmoid function in the hidden layer and the linear function in the output layer are found to be excellent in predicting the dye uptake rate, irrespective of the initial operating variable conditions. The training program was terminated when the neural network had truly converged, and the network was set ready for the prediction. The network is converged if the sum of the squared errors and the sum of the squared weights are nearly constant over several iterations. The ANN predicted normalized targets were converted back to their original target values. The details of the completely trained neural network used in the present study to design the auramine O uptake process are given in Table 3. Figure 4 shows the plot of predicted dye uptake rate qann by ANN and the qexperimental calculated from the dye adsorption experiments. From Figure 4, it can be observed that the developed neural network was found to be excellent in predicting the dye uptake rate of auramine O by activated carbon for various operating conditions. The accuracy of the newly constructed neural network was verified using the error function defined as the correlation coefficient. The correlation coefficient between the experimentally determined qexperiment and the qann determined by the neural network was found to be 0.99 with a slope value ) 1. The high correlation coefficient confirms that the newly constructed ANN

Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008 489

Figure 4. Parity plot between the amount of dye adsorbed experimentally and the amount of dye adsorbed predicted by artificial neural networks during the training process.

Figure 5. Experimental and ANN predicted adsorption kinetics of auramine O by activated carbon for various operating conditions (([): Co ) 85 mg/ L, pH ) 8, V ) 800 rpm, M ) 0.3 g, V ) 1.5 L, temp ) 305 K; (0): Co ) 180 mg/L, pH ) 8, V ) 400 rpm, M ) 0.3 g, V ) 1.5 L, temp ) 305 K; (2): Co ) 200 mg/L, pH ) 8, V ) 700 rpm; M ) 0.3 g, V ) 1.5 L, temp ) 305 K; (b): Co ) 200 mg/L, pH ) 8, V ) 500 rpm, M ) 1.8 g, V ) 1.5 L, temp ) 305 K; (O): Co ) 200 mg/L, pH ) 6, V ) 800 rpm, M ) 0.3 g, V ) 1.5 L, temp ) 305 K; (s) ANN predicted kinetics).

was highly precise in predicting the uptake of auramine O from its aqueous solutions. The ANN predicted kinetics was represented by the plot of amount of dye adsorbed, q (mg/g), versus time, t, for different operating conditions, as shown in Figure 5. Figure 5 shows the experimental data and the kinetics predicted by ANN (continuous solid lines) during the training process. From Figure 5, it can be observed that the constructed ANN represents very well the experimental kinetics of aruamine O by activated carbon. Thus, the ANN can be used successfully to model the activated carbon adsorption processes for various ranges of operating conditions. From the design point of view, it would be helpful to know the amount of dye uptake at any time, irrespective of the operating conditions. Thus, the constructed network was used to simulate the adsorption system for new operating conditions that are kept unaware of the trained neural network. For simulation, new inputs that are not used while training were

Figure 6. Parity plot between the amount of dye adsorbed experimentally and the amount of dye adsorbed predicted by the trained artificial neural networks.

fed to the trained neural network, and the corresponding amounts of dye adsorbed were determined from the neural network. Figure 6 shows the plot of predicted qann using the trained network and the qexperiment calculated experimentally for various initial operating conditions. From Figure 6, it can be observed that the newly constructed neural network was precise in predicting the dye uptake rate with a high correlation coefficient of 0.98. This shows that the developed neural network model can be precise in predicting the kinetics of auramine O uptake by activated carbon for the range of experimental conditions studied. Another advantage of the newly constructed neural network model over the theoretical models is its accuracy to predict the dye uptake as a function of studied parameters within the ranges studied. The present investigation and some of our previous works suggested the strong influence of the operating variables on the dye uptake rate. It is a highly complicated process to propose a generalized expression correlating the operating variables involved in the sorption system with the dye uptake rate. The constructed network was trained considering all the operating variables of the system, making the neural network precise enough to predict the dye uptake rate of the auramine O by activated carbon particles for any operating conditions within the ranges studied. Though the lab-scale study is limited to the range of operating variables studied, it is always possible to introduce new inputs to train the network whenever new experimental data are available. The future work is aimed to extend this idea for other solid/liquid sorption systems and also for adsorption isotherms. Acknowledgment We give sincere acknowledgments to Luis Carlos and Filomena Gonc¸ alves of FEUP for their kind help in characterizing the adsorbent. Thanks are extended to Prof. Manuel Azenha, FCUP, for the SEM analysis. Notations ANN ) artificial neural network Co ) initial dye concentration, mg/L M ) adsorbent mass, g q ) amount of dye adsorbed, mg/g rpm ) revolutions per minute t ) contact time, min

490

Ind. Eng. Chem. Res., Vol. 47, No. 2, 2008

temp ) temperature, K V ) agitation speed, rev‚min-1 V ) volume of dye solution, L Literature Cited (1) Kumar, K. V.; Ramamurthi, V.; Sivanesan, S. Modeling the mechanism involved during the sorption of methylene blue onto fly ash. J. Colloid Interface Sci. 2005, 284 (1), 14-21. (2) Allen, S. J.; McKay, G.; Khader, K. Y. H. Intraparticle diffusion of a basic dye during adsorption onto sphagnum peat. EnViron. Pollut. 1989, 56, 39-50. (3) Furusawa, T.; Smith, J. M. Fluid-Particle and lntraparticle Mass Transport Rates in Slurries. Ind. Eng. Chem. Fundam. 1973, 12 (2), 197203. (4) Weber, W. J., Jr.; Morris, J. C. Kinetics of adsorption on carbon from solution. J. Sanit. Eng. DiV., Am. Soc. CiV. Eng. 1963, 89, 31-60. (5) Lagergren, S. Zur theorie der sogenannten adsorption geloster stoffe. K. SVen Vetenskapsakad. Handl. 1898, 24 (4), 1-39. (6) Blanchard, G.; Maunaye, M.; Martin, G. Removal of heavy metals from waters by means of natural zeolites. Water Res. 1984, 18, 15011507. (7) Ho, Y. S. Adsorption of heavy metals from waste streams by peat. Ph.D. Thesis, The University of Birmingham, Birmingham, U.K., 1995.

(8) Chow, H.; Chen, H.; Ng, T.; Myrdal, P.; Yalkowsky, S. H. Using back propogation networks for the estimation of aqeous activity coefficients of aromatic compounds. J. Chem. Inf. Comput. Sci. 1995, 35, 723. (9) Aber, S.; Daneshvar, N.; Soroureddin, S. M.; Chabok, A.; Zeynali, K. S. Study of acid orange 7 removal from aqueous solutions by powdered activated carbon and modeling of experimental results by artificial neural network. Desalination 2007, 211, 87-95. (10) Naik, A. D.; Bhagwat, S. S. Optimization of an artifical neural network for modelling protein solubility. J. Chem. Eng. Data 2005, 50, 460-467. (11) Yang, M.; Wei, H. Application of neural network for the prediction of crystallization kinetics. Ind. Eng. Chem. Res. 2006, 45, 70-75. (12) Hagan, M. T.; Menhaj, M. B. Training feedforward networks with the Marquardt algorithm. IEEE Trans. Neural Networks 1994, 5 (6), 989993. (13) Neural Network Toolbox, User Guide; The MathWorks, Inc.: Natick, MA, 2000.

ReceiVed for reView August 19, 2007 ReVised manuscript receiVed November 6, 2007 Accepted November 21, 2007 IE071134P