Competitive Adsorption of Organic Micropollutants in the Aqueous

determined for a large range of concentrations (from 10-5 to 0.5 mmol L-1) to ... Whereas the IAS approach is widely used to model binary adsorpti...
1 downloads 0 Views 81KB Size
Langmuir 2002, 18, 5163-5169

5163

Competitive Adsorption of Organic Micropollutants in the Aqueous Phase onto Activated Carbon Cloth: Comparison of the IAS Model and Neural Networks in Modeling Data P. Monneyron,† C. Faur-Brasquet,*,† A. Sakoda,‡ M. Suzuki,‡ and P. Le Cloirec† Ecole des Mines de Nantes, GEPEA, BP 20 722, 4 rue Alfred Kastler, 44 307 Nantes Cedex 3, France, and Institute of Industrial Sciences, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan Received January 7, 2002. In Final Form: April 2, 2002 The adsorption equilibria of organic micropollutants (phenol, benzoic acid, and para-chlorophenol) onto activated carbon cloths have been determined for a large range of concentrations (from 10-5 to 0.5 mmol L-1) to characterize adsorption mechanisms. Single-solute isotherms tend to confirm the decisive role of the adsorbent’s microporosity in the adsorption capacity of the activated carbons. Moreover, it was found that the hydrophobicities and solubilities of adsorbates have a significant influence on the adsorption energy and capacity, respectively. Binary solute isotherms confirm these conclusions, and experimental data were used to investigate two different approaches to competitive modeling: a deterministic model, the ideal adsorbed solution theory, and a statistical one, neural networks. Both models gave good agreement between experimental and predicted data in some cases, but the results also emphasized the difficulty in satisfying the respective hypotheses of these models.

Introduction Increased worldwide concern for the environment has promoted the concept of sustainable development, which can be defined as “an approach to using the earth’s resources in a way that does not compromise the ability of future generations to meet their needs”.1 Water resources are of particular concern, and in this context, interest in new processes to provide drinking water has increased. Activated carbon cloths belong to the new technologies that have been developed. Previous studies have demonstrated the ability of these adsorbents to remove various pollutants contained in water, namely, pesticides,2 metal ions,3-5 and organic micropollutants.6-11 With a view to making water drinkable, the removal of organic pollution is of prime importance. Previous works, carried out at high initial concentrations of organics (around 100 mg L-1), have demonstrated the ability of activated carbon cloths to remove organic pollutants (aromatics and pesticides) with an initial adsorption rate 2-20 times higher than that of granular activated carbon and with adsorption capacities * To whom correspondence should be addressed. † Ecole des Mines de Nantes. ‡ Institute of Industrial Sciences. (1) Thomas, G. P. In Proceedings of the IWA Specialized Conference on Critical Technologies for the World in 21th Century: Pollution Control and Reclamation in Process Industries; IWA Publishing: London, 2000; pp 1-8. (2) Pignon, H.; Brasquet, C.; Le Cloirec, P. Environ. Technol. 2000, 21, 1261-1270. (3) Pimenov, A. V.; Lieberman, A. I.; Shmidt, J. L.; Cheh, H. Y. Sep. Sci. Technol. 1995, 30, 3183-3194. (4) Fu, R.; Lu, Y.; Zeng, H. Carbon 1998, 36, 19-23. (5) Kadirvelu, K.; Faur-Brasquet, C.; Le Cloirec, P. Langmuir 2000, 16, 8404-8409. (6) Economy J.; Lin, R. Y. Appl. Polym. Symp. 1976, 29, 199-211. (7) Baudu, M.; Le Cloirec, P.; Martin, G. Water Sci. Technol. 1991, 23, 1659-1666. (8) Sakoda, A.; Kawazoe, K.; Suzuki, M. Water Res. 1987, 21, 717722. (9) Suzuki, M. Carbon 1994, 32 (4), 577-586. (10) Brasquet, C.; Subrenat, E.; Le Cloirec, P. Water Sci. Technol. 1999, 39, 201-205. (11) Brasquet, C.; Le Cloirec, P. Langmuir 1999, 15, 5906-5912.

ranging from 50 to 500 mg g-1.12,13 However, to use these adsorbents for the treatment of drinking water, it would be interesting to learn how activated carbon cloths behave at low concentrations of organic pollutants (1 mg L-1). Furthermore, investigations are also needed to study adsorption competition between organics. The primary focus of this paper is to assess the adsorption performance of activated carbon cloths for a large range of initial concentrations from low (0.008 mM) to high (1 mM) values. Three organic compounds are studied: phenol, benzoic acid, and para-chlorophenol. The results enable the adsorption mechanisms of the three compounds on activated carbon cloths to be explained. A second objective consists of modeling experimental data obtained in the first part using both a deterministic model, the ideal adsorbed solution (IAS) model, and a statistical one, neural networks (NNs). Whereas the IAS approach is widely used to model binary adsorption,14-16 the modeling of adsorption using NNs is particularly useful for the prediction of monocomponent adsorption, for example, for modeling adsorbent breakthrough in a dynamic reactor17,18 or for assessing quantitative structure-activity relationships relating adsorption parameters to adsorbate characteristics.19,20 However, few studies have been performed on the modeling of competitive adsorption by NNs, and they were all carried out using (12) Le Cloirec, P.; Brasquet, C.; Subrenat, E. Energy Fuels 1997, 11, 331-336. (13) Brasquet, C.; Rousseau, B.; Estrade-Szwarckopf, H.; Le Cloirec, P. Carbon 2000, 38, 407-422. (14) Baudu, M.; Le Cloirec, P.; Martin, G. Chem. Eng. J. 1989, 41, 81-89. (15) Crittenden, J. C.; Luft, P.; Hand, D. W.; Oravitz, J. L.; Lopers, S.; Ari, M. Environ. Sci. Technol. 1985, 19, 1037-1043. (16) Annesini, M. C.; Gironi, F.; Ruzzi, M.; Tomei, C. Water Res. 1987, 21, 567-571. (17) Basheer, I. A.; Najjar, Y. M.; Hajmeer, M. N. Environ. Technol. 1996, 17, 795-806. (18) Bulsari, A. B.; Palosaari, S. Neural Comput. Appl. 1993, 1, 160165. (19) Brasquet, C.; Le Cloirec, P. Water Res. 1999, 33, 3603-3608. (20) Brasquet, C.; Bourges, B.; Le Cloirec, P. Environ. Sci. Technol. 1999, 33, 4226-4231.

10.1021/la020023m CCC: $22.00 © 2002 American Chemical Society Published on Web 05/29/2002

5164

Langmuir, Vol. 18, No. 13, 2002

Monneyron et al.

Table 1. Main Characteristics of Activated Carbons commercial name raw material activation gas activation temperature (°C) BET surface area (m2 g-1) pore volume (cm3 g-1) micropore volume (%) pHPZC

CS-1501 rayon CO2 1200 1689 0.665 96.4 7.60

RS-1301 rayon H2O 900 1460 0.506 68.2 9.50

NC-60 coconut H2O + CO2 900 1173 0.471 94.5 9.90

Table 2. Main Properties of Adsorbates property

phenol

benzoic acid

parachlorophenol

symbol formula molecular weight (g mol-1) pKa solubility 25 °C (g L-1) Kow Bondi volume (cm3 mol-1)

P C6H5OH 94 9.9 66.7 1.50 53.5

BA C6H5COOH 122 4.2 2.9 1.87 64.0

PCP C6H5Cl 129 9.4 28.0 2.35 61.8

activated carbon in the form of granules.21,22 This work will investigate the ability of NNs to model binary adsorption in the aqueous phase onto activated carbon cloths and will compare this performance with that of the IAS model. Experimental Section 1. Adsorbents. Two activated carbon cloths were used in this work, one microporous (CS-1501) and the other mesoporous (RS-1301), both from Actitex Co. (Levallois, France). They were compared with a common granular activated carbon (NC-60) from Pica Co. (Levallois, France). The main characteristics of these activated carbons are presented in Table 1. Pore characteristics were determined by nitrogen adsorption at 77.7 K with a Coulter SA 3100 apparatus. The point of zero charge (PZC), i.e., the pH above which the total surface of the carbon is negatively charged, was measured by the so-called pH drift method. 2. Adsorbates. Three organic compounds were studied: phenol (P), benzoic acid (BA), and para-chlorophenol (PCP). They are all based on the same chemical structure (aromatic) to enable a discussion of the dependence of the adsorption results on the adsorbate characteristics. The basic molecule, phenol, was selected because of its simple structure. The other two molecules are also aromatic compounds, whose maximum adsorption capacities are close to that of phenol20 but whose substitutents, -COOH and -Cl, have a different polarity and size. The main properties of the three adsorbates are given in Table 2, where Kow is the partition coefficient for octanol-water. All solutes were analyzed using a Waters 600 high-performance liquid chromatograph, provided with a UV Waters 486 detector and a Novapack C18 nonpolar column. Multisolute analyses were performed using gradient mode. In the case of low concentrations (less than 0.08 mM), a solid-phase extraction (SPE) preceded the HPLC analysis. SPE was carried out with a Waters Oasis HLB cartridge, and recovery ranged between 89 and 94%. In all cases, solutions were filtered before analysis with a Lida Manufacturing Corp. filter of diameter 0.45 µm. 3. Isotherms. Monocomponent isotherms were measured at 25 ( 1 °C by stirring, at a rate of 300 rpm, an activated carbon weight of 80 mg in 250 mL of solution with an initial concentration of phenol, benzoic acid, or para-chlorophenol ranging between 0.008 and 1 mM. The stirring times necessary to reach equilibrium were 24 h for activated carbon cloths and 48 h for granular activated carbon.11 For each isotherm, a reference solution with an intermediate concentration was stirred without activated carbon. The initial pH was 5.6 for phenol, 3.8 for benzoic acid, and 5.3 for para-chlorophenol. Solutions were not buffered to avoid adsorption competition between organics and buffer. (21) Yang, M.; Hubble, J.; Lockett, A. D.; Rathbone, R. R. Sep. Sci. Technol. 1996, 31, 1259-1265. (22) Carsky, M.; Do, D. D. Adsorption 1999, 5, 183-192.

Figure 1. Particular shape of isotherm curve of BA adsorption onto CS-1501. The procedure for binary isotherms was the same, with the initial solutions being composed of equimolar concentrations of the two compounds. The experimental error was equal to 3% for equilibrium concentration and 3.5% for equilibrium adsorption capacity.

Results and Discussion 1. Monocomponent Isotherms. 1.1. Experimental Results. Before examining the binary adsorption onto activated carbon cloths, monocomponent isotherms were measured with the three organic compounds and the three activated carbons. Figure S-1 (see Supporting Information) presents the adsorption isotherms of the three organics on the activated carbon cloth CS-1501. The same shape of isotherm, of type I according to the IUPAC classification,23 was obtained for the three activated carbons for P and for PCP, indicating a physisorption process onto a microporous adsorbent. In the case of BA, the isotherm shape in the low concentration range (given in Figure 1 for adsorption onto CS-1501) might illustrate an adsorption mechanism that combines physisorption and chemisorption. The experimental isotherm curve would be the sum of two curves: the first (linear isotherm 1) being close to Henry’s law in the concentration range considered and the second (curvilinear isotherm 2) being due to electrostatic interactions between activated carbon surface groups and benzoic acid. Indeed, equilibrium pH is close to 5.5 for the low concentration range. At this pH, the surface of activated carbon is, on average, positively charged (pH < PZC), and 90% of benzoic acid is in benzoate form, C6H5COO- (pH > pKa). This explanation is in accordance with the previous observations: (1) For high concentrations, chemisorption can be disregarded with respect to physisorption, and the isotherm’s shape is again of type I. Isotherms of P and PCP do not present this particular shape; indeed, hydroxyl groups of both compounds are in basic form, and electrostatic interactions are thus repulsive. A modification of the activated carbon surface by rinsing with HCl would allow this hypothesis to be confirmed. 1.2. Modeling of Isotherm Curves. The different experimental isotherms were modeled according to three equations: the Freundlich, Langmuir, and Dubinin models. (23) IUPAC Physical Chemistry Division. Pure Appl. Chem. 1985, 57, 603-619.

Competitive Adsorption of Organic Micropollutants

Langmuir, Vol. 18, No. 13, 2002 5165

The Freundlich empirical model24 is represented by

qe ) KfCe1/n

(I)

where Ce is the equilibrium concentration (mol L-1), qe is the amount adsorbed at equilibrium (mol g-1), and Kf (mol1-1/n L1/n g-1) and 1/n are Freundlich constants depending on the temperature and the given adsorbentadsorbate couple. n is related to the adsorption energy distribution, and Kf indicates the adsorption capacity. Langmuir’s model25 does not take into account the variation in adsorption energy, but it is the simplest description of the adsorption process. It is based on the physical hypothesis that the maximum adsorption capacity consists of a monolayer adsorption, that there are no interactions between adsorbed molecules, and that the adsorption energy is distributed homogeneously over the entire coverage surface. Langmuir’s equation is

qe )

bqmCe 1 + bCe

(II)

where b is the equilibrium adsorption coefficient (L mol-1) and qm is the maximum adsorption capacity (mol g-1). Dubinin’s approach is based on the earlier potential theory of Polanyi and on the concept of micropore filling.26 The Dubinin-Radushkevich equation represents the isotherm in terms of fractional filling (qe/qDR) of the micropore volume in which the maximum adsorption capacity is qDR

qe ) exp(-B2) qDR

(III)

where qDR is the maximum adsorption capacity in the micropore volume (mol g-1),  is the Polanyi adsorption potential, and B is the characteristic parameter related to adsorption energy for the given system (mol2 kJ-2). The Polanyi adsorption potential  is an adsorption affinity defined in the liquid phase by

 ) RT ln

() Cs Ce

(IV)

where T is the temperature (K) and Cs is the solubility of the solute (mol L-1). Compared with Langmuir’s theory, the Dubinin approach assumes that there is a surface area where the adsorption energy is homogeneous. The characteristic adsorption energy E0 (kJ mol-1) is then given by

E0 ) 1/xB

(V)

The three models were applied to experimental data, and the results of Freundlich, Langmuir, and Dubinin modeling are given in Tables S1, S2, and S3, respectively (all in the Supporting Information), where R2 is the determination coefficient of linear forms of eqs I-III. To use Freundlich’s model in an optimal way as suggested by Tien,27 the isotherm was divided into various concentration ranges, and the Freundlich equation was applied to each (24) Freundlich, H. Colloid and Capillary Chemistry; Methuen and Co Ltd.: London, 1926. (25) Langmuir, L. J. Am. Chem. Soc. 1918, 40, 1361-1403. (26) Dubinin, M. M.; Radushkevich, L. V. Proc. Acad. Sci. USSR 1947, 55, 331. (27) Tien, C. Adsorption Calculations and Modelling; ButterworthHeinemann: Woburn, MA, 1994.

of them. In contrast, Langmuir’s equation was applied only to the high concentration range because of the highly heterogeneous character of activated carbon, which prevents the model from describing the entire range of concentrations. 1.3. Qualitative Approach to the Adsorption Mechanism. First, modeling results allow for a comparison of the three adsorbents. For the three micropollutants, the CS-1501 sample presents the highest adsorption capacity whereas the adsorption capacities of RS-1301 and NC-60 are similar. This observation is based on a comparison of the qm and qDR parameters, which can be related to the micropore volumes of the adsorbents. Indeed, CS-1501 has a high pore volume (0.665 cm3 g-1) compared to RS1301 (0.506 cm3 g-1) and NC-60 (0.471 cm3 g-1), whose pore volumes are quite similar. This result confirms previous research on the adsorption of phenolic compounds onto various activated carbons,28 where the Langmuir maximum adsorption capacity qm (mmol g-1) could be related to the micropore volume of activated carbons for phenol (R2 ) 0.998) and para-chlorophenol (R2 ) 0.978). In our case, linear regressions between qm and micropore volume Vm give determination coefficients R2 equal to 0.560 for P, 0.999 for PCP and 0.874 for BA, respectively. Regarding the regression between qDR and Vm, higher values of R2 are obtained, i.e., 0.980 for P, 0.980 for PCP, and 0.941 for BA, which confirms that the pore size distribution has a great influence on adsorption capacity. Now, modeling results can be compared for the three adsorbates in terms of adsorption capacity and adsorption energy. When the isotherms were divided into several intervals of concentration, the highest concentration range was considered for comparison. The adsorption capacity parameters (characterized by Kf, qm, and qDR coefficients) order as follows:

BA > PCP > P for qm and Kf PCP > P > BA for qDR This classification orders Kf and qm in a reciprocal way to the solubilities given in Table 2, as has previously been reported by various studies.29,30 For the qDR parameter, where the solubility being taken into account is given by eq IV, the order is different. Concerning energy parameters (related to 1/n, b, and E0), the classification is PCP > BA > P or PCP > P > BA. In this case, the polarity of the adsorbate seems to influence the adsorption energy greatly. Octanol-water partition coefficients, Kow, given in Table 2, show that PCP is more hydrophobic than BA and P. It has clearly been demonstrated that adsorption of phenolic compounds onto activated carbon induces the formation of π-π bonds, where AC acts as an electron donor and the solute benzene ring has an electron-withdrawing character.31 The mesomeric and/or inductive character of the substitutent of the aromatic compound influences this formation and thus the molecule’s adsorption energy. This interpretation is supported by experimental results: the withdrawing inductive character of the chloride substitutent decreases the electron density of the PCP benzene ring compared with that of the P benzene ring. The adsorption energy (28) Hu, Z.; Srinivasan, M. P.; Ni, Y. In Proceedings of the 2nd Pacific Basin Conference on Adsorption Science and Technology; Do, D. D., Ed.; Singapore, 1999; pp 274-278. (29) Moreno-Castilla, C.; Rivera-Utrilla, J.; Lopez-Ramon, M. V.; Carrasco-Marin, F. Carbon 1995, 33, 845-851. (30) Liu, J. C.; Huang, C. P. J. Colloid Interface Sci. 1992, 153, 16. (31) Mattson, J. S.; Mark, H. B., Jr.; Malbon, M. D.; Weber, W. J., Jr.; Crittenden, J. C. J. Colloid Interface Sci. 1969, 31, 116-130.

5166

Langmuir, Vol. 18, No. 13, 2002

of PCP is then higher than that of P. For P and BA, the adsorption energies are similar because of the hydrophilic groups -OH and -COOH, which have similar inductive effects on the benzene ring. 2. Binary Isotherms. 2.1. Experimental Results. To confirm the influence of solubility and hydrophobicity observed in the monocomponent study, binary adsorption was performed using equimolar initial concentrations. Isotherm curves are presented in the same figure, but they do not really feature competitive adsorption, which would require equal equilibrium concentrations of the two given compounds for all points of the graph. This condition is difficult to obtain in a static reactor. Three kinds of binary systems were studied on each activated carbon, and similar results were obtained for all three activated carbons. Figure 2 presents the isotherm curves obtained on the CS-1501 sample for high initial concentrations. Binary adsorption confirms the conclusions of the monocomponent isotherm analysis. For the P-BA system, the adsorption capacity of BA is much higher than that of P, and the discrepancy increases with the concentrations of the two compounds. This observation might be related to the respective solubilities of the compounds, that of phenol being 30 times greater than that of BA. The higher the initial concentrations of the adsorbates are, the better the adsorbate with the lower solubility (BA) adsorbs. This influence of solubility on adsorption capacity is included in the Brunauer-Emmett-Teller model.32 As initial concentrations decrease, solubility is no longer such a deciding factor, so the adsorption capacities become similar for P and BA. For the P-PCP system, the same isotherm shape is obtained, but the discrepancy might be due instead to a difference in adsorption energy, related to the hydrophobicity of the Cl substitutent. The behavior of the BA-PCP system is different from that of the other two: adsorption capacities are similar with values higher for PCP at low concentrations. This would confirm the major influence of hydrophobicity for concentration ranges distant from the saturation concentration of one compound. Binary adsorption was also performed in the low concentration range, C0 ) 0.008-0.8 mM. To improve the presentation of the data, coadsorption isotherms of the BA-P system onto CS-1501 are provided in Figure 3 on a logarithmic scale. The same kind of results were obtained with the other activated carbons. To check the agreement between the two experimental data sets, the first data obtained at high concentrations are also presented (encircled points). This figure reveals the importance of the initial concentration on the competition between the two adsorbates. To identify the value of the concentration where a reversal of competition might occur, the ratio of adsorption capacities of the two compounds [qe(1)/qe(2)] was plotted as a function of initial equimolar concentration for the three couples P-BA, P-PCP, and BA-PCP. The results are presented in Figure 4. In the case of (P-PCP), the adsorption of PCP is larger than that of P, resulting in a decrease in the ratio qe(P)/qe(PCP) as C0 increases. Competition is significant from C0 ) 0.045 mM. In the case of (P-BA), the influence of solubility appears from C0 ) 0.020 mM. For C0 slightly less then 20 mM, the adsorption capacity of phenol is larger than that of BA. For very low C0, the electrostatic attraction of BA, demonstrated by monocomponent adsorption, induces an increase in BA removal. Finally, in the case of the BA(32) Brunauer, S.; Emmett, P. H.; Teller, E. J. Am. Chem. Soc. 1938, 60, 309-319.

Monneyron et al.

Figure 2. Binary adsorption isotherms of (a) P-PCP, (b) BAPCP, and (c) P-BA onto CS-1501. V ) 250 mL, T ) 25 °C, stirring rate ) 300 rpm, stirring time ) 12 h, AC weight ) 80 mg, equimolar C0 ) 0.08-1 mM.

Competitive Adsorption of Organic Micropollutants

Langmuir, Vol. 18, No. 13, 2002 5167

phase onto activated carbon cloth and to compare its performance with that of a thermodynamic model, the ideal adsorbed solution (IAS) model. The IAS theory was first established for mixed-gas adsorption by Myers and Prausnitz33 and then extended to multisolute adsorption from dilute liquid solution by Radke and Prausnitz.34 The model is based on the fundamental hypothesis that a multicomponent solution has the same spreading pressure π as an ideal single-component solution of the ith component, the spreading pressure being the difference between the interfacial tension of the pure solvent and that of the solution containing the solute. This hypothesis is described by the Gibbs equation

πA ) RT

Figure 3. Binary adsorption of phenol (P) and benzoic acid (BA) in the low concentration range on CS-1501.

∫0

Ci0

qi0

dCi0

0

Ci

(VI)

where A is the adsorbed surface area per unit mass of adsorbent (m2 g-1), π is the spreading pressure, T is the temperature (K), Ci0 is the equilibrium concentration of the pure-component adsorption of the ith compound (mol L-1), and qi0 is the equilibrium adsorption capacity of the pure-component adsorption of the ith compound (mol g-1). The following equations complete the system description:

q i ) q t xi

(VII)

Ci ) Ci0(π,T)xi

(VIII)

(∑ ) n

qt )

xj

0

-1

(IX)

j)1 q j

where Ci is the equilibrium concentration of the multicomponent adsorption of the ith compound (mol L-1), qi is the equilibrium adsorption capacity of the multicomponent adsorption of the ith compound (mol g-1), qt is the total adsorption capacity of the multicomponent system (mol g-1), xi is the mole fraction of the ith adsorbate in the adsorbed phase, and n is the number of components. The evaluation of the IAS model requires a precise determination of eq VII subject to the constraint Figure 4. Influence of initial concentration C0 on adsorption competition of binary systems on CS-1501.

PCP system, three phenomena can be noted that explain the curvilinear shape of the curve qe(BA)/qe(PCP) ) f(ln C0): (1) Binary adsorption seems to be governed overall by the stronger affinity of activated carbon for PCP, inducing larger adsorption capacities. (2) As the initial concentration increases, especially from C0 ) 0.8 mM, the influence of the BA solubility is more pronounced, and the ratio qe(BA)/qe(PCP) increases again. (3) However, for C0 less than 0.1 mM, the electrostatic attraction of BA for activated carbon surface functional groups induces an increase in the ratio qe(BA)/qe(PCP). This experimental study of binary adsorption onto activated carbons has shown that various parameters have a joint influence on the adsorption capacities, namely, adsorbate solubility and adsorption energy in the range of high concentrations and chemical interactions between adsorbate and activated carbon surface in the low concentration range. 2.2. Modeling by the IAS Model and Neural Networks. Experimental data were used to test the ability of neural networks (NNs) to model binary adsorption in the aqueous

n

xi ) 1 ∑ i)1

(X)

which enables the calculation to be carried out using an iterative estimation of the spreading pressure. The determination of adsorption capacities in multicomponent solutions requires an appropriate description of the singlesolute isotherm. Freundlich’s model was selected here, with the division of isotherms into linear concentration ranges introduced by Tien.27 Calculations were performed using a program developed in Microsoft Excel 97. The quality of modeling is not an obvious function of the activated carbon but is highly dependent on the binary system considered and on the accuracy of single isotherms because small errors in the low concentration range generate large deviations in the prediction of adsorption selectivity. For the system P-PCP, experimental and modeled values are in agreement, as can be seen for the binary adsorption onto RS-1301 in Figure S-2 (see Supporting Information). However, for high initial con(33) Myers, A. L.; Prausnitz, J. M. AIChE J. 1965, 11, 121-127. (34) Radke, C. J.; Prausnitz, J. M. AIChE J. 1972, 18, 761-768.

5168

Langmuir, Vol. 18, No. 13, 2002

Figure 5. IAS modeling results of P-BA binary adsorption on RS-1301.

centrations, the increase in the ionic strength of the solution induces a small decrease in modeling performance. This confirms a previous study, which demonstrated that the performance of the IAS model was lower for high concentrations.14 In the case of both binary systems including BA, the quality of IAS modeling is now unsatisfactory, with an underestimation of BA adsorption capacities at high concentrations and an overestimation at low concentrations, as illustrated in Figure 5 for the P-BA system. IAS modeling of the coadsorbate, P or PCP depending on the system considered, presents a reversal tendency. This low estimation of BA adsorption would confirm the influence of its low solubility, which is not taken into account in the IAS model. Furthermore, as C0 increases, the multicomponent solution characteristics (for example, in terms of pH) will be dictated by BA. This would explain why PCP adsorption modeling is quite stable at high concentrations, whereas that of P is more and more overestimated, because the characteristics of PCP are closer to those of BA, especially in terms of solubility. Finally, it seems that the IAS model can be directly used for multicomponent systems where solutes have similar chemical characteristics. To extend the use of this model, some researchers have established modified IAS models.35 In this paper, rather than modifying the IAS method, a statistical approach was carried out that consists of using neural networks (NNs). Neural networks are easily used to model nonlinear industrial processes such as the control of hydraulic servo systems,36 the dynamic ultrafiltration of proteins,37 or an anaerobic biological wastewater treatment system.38 The modeling of adsorption has also been performed with neural networks in a dynamic reactor17,18,39 to predict the breakthrough time of an adsorbent for given operating conditions. In this work, we plan to model binary adsorp(35) Seidel, A.; Gelbin, D. Chem. Eng. Sci. 1988, 43 (1), 79-89. (36) Burton, R. T.; Ukrainetz, P. R.; Nikiforuk, P. N.; Schoenau, G. J. Proc. Inst. Mech. Eng. 1999, 213 (1), 349-358. (37) Bowen, W. R.; Meirion, G. J.; Yousef, H. N. S. J. Membr. Sci. 1998, 146, 225-235. (38) Tay, J.-H.; Zhang, X. J. Environ. Eng. 1999, 125 (12), 11491159. (39) Yang, M.; Hubble, J.; Fang, M.; Lockett, A. D.; Rathbone, R. R. Biotechnol. Technol. 1993, 7 (2), 155-158.

Monneyron et al.

tion in a batch reactor and compare the results with those obtained with the IAS model. Neural networks (NNs) are algorithmic systems derived from a simplified concept of the brain. In a NN, a number of nodes, called neurons, are interconnected into a netlike structure and can perform parallel computation for data processing and nonlinear phenomenon representation thanks to a sigmoidal function.40 The degree of influence between neurons is dictated by the values of links called connection weights. The kind of network used in this work is a three-layer perceptron consisting of one input layer containing independent variables, one output layer containing the dependent parameter, and one hidden layer in which the number of neurons is optimized using an iterative process. The NN training is achieved by adjusting the values of connection weights, through the repeated application of the back-propagation algorithm.41 A training set pattern is presented at the input units, with bond weights starting from random values. Bond weight values are then changed by iteratively minimizing the difference between the desired and the calculated output. The neural network training is performed with a socalled training data set. Its architecture, in terms of transfer function and number of hidden neurons, is optimized using a validation data set. The generalization ability of a NN is assessed with a test data set that has not been used for training or validation. In this study, the input layer consisted of 4 neurons: Cem,1 and Cem,2, which are the experimental equilibrium concentrations of the two compounds in a multicomponent solution, and qe1 and qe2 the equilibrium adsorption capacities in monocomponent solutions computed from Freundlich constants of the different concentration ranges for monocomponent equilibrium concentrations Ce1 and Ce2, taken as equal to Cem,1 and Cem,2 (Figure S-1, Supporting Information). The output neuron that has to be predicted is qem,i (i ) 1 or 2), the adsorption capacity of solute 1 or 2 in a multicomponent solution. An optimization process enabled the assessment of the number of neurons in the hidden layer to be equal to 2. The final architecture of the NN is given in Figure S-3 in the Supporting Information. Because of the low number of compounds, the generalization ability of the NN was tested not on binary compounds but rather on randomly selected data to have a good representation of the test data type in the training subset. Indeed, previous studies have demonstrated that the training data set must be very representative of the process that has to be predicted.21 Training and validation of the NN were thus performed with 70% of randomly selected experimental data. The generalization ability of the trained NN was tested using the 30% remaining data. Generalization results for both solutes are given in Figure 6 in the form of a linear regression between experimental and NNpredicted adsorption capacities in the binary system. For both solutes, predicted data are in agreement with observed data with determination coefficient R2 equal to 0.943 and 0.975 for solutes 1 and 2, respectively. Rootmean-squared errors are lower for the NN than for the IAS model: whereas RMSE ≈ 0.09 for the NN prediction, for the IAS model, a value of 0.199 was obtained for the binary system P-BA and 0.143 for P-PCP. (40) Bishop, C. Neural Networks for Pattern Recognition; Clarendon Press: Oxford, U.K., 1995 (41) Rumelhart, D. E.; Hinton, G. E.; Williams, R. J. Nature 1986, 323, 533-536.

Competitive Adsorption of Organic Micropollutants

Langmuir, Vol. 18, No. 13, 2002 5169

nent adsorption demonstrated the influence of adsorbent microporosity. Moreover, adsorbate solubility and hydrophobicity have a large influence on adsorption capacity and energy, respectively. These influences were confirmed by binary adsorption. The modeling of multicomponent experimental results by a thermodynamic model, the IAS model, and a statistical model, neural networks, demonstrated the difficulty in satisfying the hypotheses relative to each model. In the case of the IAS model, the ideality of the solution is a restrictive assumption for the use of the model. In the case of NNs, when the training data set is representative of the future data to be predicted, good generalization results are obtained. Acknowledgment. The experimental part of this work was carried out in France (DSEE, Ecole des Mines de Nantes) and in Japan (Institute of Industrial Science). The authors thank the Pica and Actitex companies for providing activated carbons. Figure 6. Generalization performance of the NN with 2 test databases.

Conclusion This work has studied the ability of activated carbon cloths to remove microorganic compounds in the aqueous phase over a large range of concentrations. Monocompo-

Supporting Information Available: Figures showing monocomponent adsorption isotherms of pollutants, IAS modeling results of P-PCP binary adsorption, and optimized architecture of the neural network and tables giving Freundlich, Langmuir, and Dubinin equation parameters. This material is available free of charge via the Internet at http://pubs.acs.org. LA020023M