5906
Langmuir 1999, 15, 5906-5912
Effects of Activated Carbon Cloth Surface on Organic Adsorption in Aqueous Solutions. Use of Statistical Methods To Describe Mechanisms† C. Brasquet and P. Le Cloirec* Ecole des Mines de Nantes, De´ partement Syste` mes Energe´ tiques et Environnement, 4 rue Alfred Kastler, BP 20722, 44307 Nantes Cedex 03, France Received August 27, 1998. In Final Form: November 30, 1998 The adsorption of polluted water is performed by activated carbon fibers. Three kinds of material are compared: microporous and mesoporous cloths and a microporous granular activated carbon. These porous solids are characterized by scanning electron microscopy and atomic force microscopy. BET surface areas and porous volumes are determined. Adsorption of a large number of organic compounds is carried out in water onto activated carbon cloths and granules. Kinetic and equilibrium data are expressed in terms of initial velocities and classical model parameters (Freundlich). These adsorbability parameters are then discussed according to solute molecular structures and activated carbon characteristics. The results obtained show that fluid-solid transfer is directly related to surface and porous structure (pore size distribution and pore connection with the external surface of adsorbents). The adsorption data of several aromatics and aliphatics onto a microporous activated carbon cloth are discussed statistically. A quantitative structureproperty relationship method is used. Multiple linear regression and neural networks enable the assessment of correlations between the Freundlich adsorption parameter (log K) and molecular structure defined by molecular connectivity indexes. The neural network architecture is optimized, and results determined by the two statistical methods are compared: the neural network approach seems to give better results than multiple linear regression to assess this kind of relationship between adsorption and adsorbate molecular structure, even though its predictive ability is low. From a variable analysis, the results are discussed in terms of adsorbate positions at the adsorbent surface and mechanisms of interaction between solutes and an activated carbon surface are proposed.
Introduction The Rio conference in 1992 has resulted in a sudden awareness of environmental problems and especially in a toughening of regulations in the areas of air and water treatment. In this context, interest in new treatment processes and materials have increased, and activated carbon fibers, in the form of felts or cloths, belong to the new technologies which have been developed. Granular activated carbon is well-known and currently used for water and air purification, but activated carbon fibers are new materials which have been little studied especially in terms of their characteristics and adsorption mechanisms. Some studies have shown their potential to remove organics,1-5 mineral compounds,6-8 and even microorganisms9 from the aqueous phase. In the case of organic compounds, the adsorption mechanism seems to be physisorption,3,10 and the adsorption rates are 2 to 20 times higher than those obtained with granular activated †
Presented at the Third International Symposium on Effects of Surface Heterogeneity in Adsorption and Catalysis on Solids, held in Poland, August 9-16, 1998. (1) Economy, J.; Lin, R. Y. Appl. Polym. Symp. 1976, 29, 199-211. (2) Petkovska, M.; Mitrovic, M. Chem. Biochem. Eng. Q. 1989, 3 (4), 153-159. (3) Kaneko, Y.; Abe, M.; Ogino, K. Colloids Surf. 1989, 37, 211-222. (4) Brasquet, C.; Le Cloirec, P. Carbon 1997, 35, 1307-1313. (5) Le Cloirec, P.; Brasquet, C.; Subrenat, E. Energy Fuels 1997, 11, 331-336. (6) Pimenov, A. V.; Lieberman, A. I.; Shmidt, J. L.; Cheh, H. Y. Sep. Sci. Technol. 1995, 30, 3183-3194. (7) Fu, R.; Lu, Y.; Zeng, H. Carbon 1998, 36, 19-23. (8) Camara, S.; Wang, Z.; Ozeki, S.; Kaneko, K. J. Colloid Interface Sci. 1994, 162, 520-522. (9) George, N.; Davies, J. T. J. Chem. Technol. Biotechnol. 1988, 88, 117-129. (10) Baudu, M.; Le Cloirec, P.; Martin, G. Water Res. 1993, 27 (1), 69-76.
carbon11 due to the large external surface area of the fibers and to the connection of micropores to this external surface area, which involves a decrease in mass transfer resistance.12 However, the mechanisms involved in the adsorption processes are not well-known. The main objective of this work is to study the dependence of adsorption on adsorbent and adsorbate characteristics. Microporous and mesoporous activated carbon cloths (ACCs) are compared with a granular activated carbon, in terms of adsorption rates and capacities toward organics. The results are discussed in terms of the influence of solute molecular structure and activated carbon characteristics on adsorption. To quantify these relations, a quantitative structure-property relationship (QSPR) bringing molecular connectivity indexes into play is set up by two statistical methods: multiple linear regression and neural network. Statistical tools are compared, and adsorption mechanisms are confirmed. Experimental Section 1. Activated Carbons. The activated carbons used in this study are commercial products from the Actitex Co. and Pica Co. (Levallois, France). Their main characteristics are given in Table 1. Specific surface areas and porous volumes were determined by nitrogen adsorption isotherms at 77 K carried out with a COULTER SA 3100 apparatus (IPSN, Fontenay aux Roses, France), and using BET and R-plot equations. These porosity parameters were used to confirm microscopic observations carried out with a JEOL JSM 6400-F scanning electron microscope (IMN, Nantes, France) and a PARK atomic force microscope (CRMD, Orle´ans, France). (11) Baudu, M.; Le Cloirec, P.; Martin, G. Water Sci. Technol. 1991, 23, 1659-1666. (12) Brasquet, Ph.D. Thesis, Universite´ de Pau et des Pays de l’Adour, 1998.
10.1021/la9811160 CCC: $18.00 © 1999 American Chemical Society Published on Web 03/19/1999
Organic Adsorption on Activated Carbon
Langmuir, Vol. 15, No. 18, 1999 5907
Table 1. Main Characteristics of Activated Carbon Materials CS 1501 RS 1301 company presentation raw material activation, oxidation gas temperature (°C) BET surface area (m2 g-1) porous volume (cm g-1) microporous volume (%) median micropore diameter (Å) a
Actitex cloth rayon CO2 1200 1689 0.665 96.4 6.9
Actitex cloth rayon H2O 900 1460 0.506 68.2 7.3
NC 60 Pica granules coconut CO2-H2O 900 1173 0.471 94.5 7.3
Levallois, France.
Figure 3. AFM observation of CS 1501 fiber surface.
Figure 1. SEM observation of CS 1501 fiber cross section.
Figure 4. AFM observation of RS 1301 fiber surface.
Figure 2. SEM observation of RS 1301 fiber cross section. Scanning electron micrographs in Figures 1 and 2 compare the activated carbon fiber cross sections of CS 1501 and RS 1301 samples, respectively. Both fibers have diameters around 10 µm, with a ribbed structure arising from the rayon precursor.13 Whereas the CS 1501 fiber surface is quite smooth with numerous ribs along the fiber axis, the RS 1301 fiber surface contains numerous holes and only three larger ribs. The atomic force micrograph (Figure 3) of the CS 1501 sample shows features which did not appear in scanning electron microscopy (SEM) observations: some linear ridges are obvious at the fiber surface, which could indicate a linear microporosity developing in the fiber bulk and connecting the bulk to the external surface area. They would confirm the microporous character of this sample, more than 96% of whose volume is microporous. The atomic force micrograph (Figure 4) of the RS 1301 sample shows aligned mesopores with diameters ranging from 100 to 800 Å. Differences in porosity characteristics arise from the activation conditions. Carbon dioxide, which has a higher diffusion coefficient than steam, leads to a tight microporosity in the fiber bulk, whereas steam induces a larger pore size distribution at the fiber surface.14,15 (13) Lord, E. J. Text. Inst. 1955, 46, T191-T213.
2. Adsorbates. All adsorbates are commercially available with a high degree of purity (>98%). They are all organic compounds, and the majority are aromatics (Sigma Aldrich). To determine the influence of adsorbate structure on adsorption, the substitution pattern was varied in terms of the content, size, and position of heteroatoms. Finally, the database used for the adsorption study contained 55 molecules. 3. Adsorption Procedure. 3.1. Kinetics. Kinetics were carried out by stirring, at 25 ( 1 °C, 500 mg of activated carbon in the form of cloth or granules in 1 L of aqueous solution containing an organic compound with an initial concentration of 100 mg L-1. Samples of 5 mL were withdrawn at regular times until a steady state was obtained, and concentration was plotted as a function of time. The initial adsorption rates were calculated from the initial stage of the adsorption reaction by the following equation
γ)
V dC C0m dt
( )
tf0
(I)
where γ is the initial adsorption coefficient (L mg-1 min-1), V the solution volume (L), C0 the initial solute concentration (mg L-1), m the activated carbon weight (mg), C the solute concentration at time t (mg L-1), and t the time (min). 3.2. Isotherms. Isotherms were carried out at 25 ( 1 °C, by stirring various weights (from 25 to 500 mg) of activated carbon (14) Alcan˜iz-Monge, J.; Cazorla-Amoro´s, D.; Linares-Solano, D.; Yoshida, S.; Oya A. Carbon 1994, 32, 1277-1283. (15) Molina-Sabio, M.; Gonzales, M. T.; Rodriguez-Reinoso, F.; Sepulveda-Escribano, A. Carbon 1996, 34, 505-509.
5908 Langmuir, Vol. 15, No. 18, 1999
Brasquet and Le Cloirec
Table 2. Initial Adsorption Coefficients and Freundlich Parameters for Adsorption of Organic Compounds onto CS 1501, RS 1301, and NC 60 organic compound benzoic acid 3,5-dimethoxybenzoic acid 4-butylbenzoic acid aniline benzaldehyde p-chlorophenol diethyl phthalate 1-ethoxy-2-butoxyethane p-nitrophenol phenol phenylalanine 2-butyl-4-methylphenol toluene DL-tyrosine
vaniline
activated carbon
105γ
K
1/n
r
CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 CS 1501 RS 1301 NC 60 CS 1501 RS 1301 NC 60
4.62 5.54 1.24 3.83 4.24 0.45 3.73 3.96 0.26 6.75 5.24 1.84 12.5 3.90 2.19 7.65 9.37 2.29 6.50 4.10 3.56 4.67 1.94 13.2 8.81 2.09 5.44 4.25 1.96 3.85 2.62 0.40 5.24 6.76 0.71 8.40 8.12 5.37 5.26 0.64 8.17 7.10 1.68
82.0 88.5 53.7 41.9 30.3 45.3 42.7 21.3 12.8 47.5 43.8 43.3 114.1 75.6 82.4 131.6 141.2 122.7 222.3 215.1 110.0 97.2 46.0 187.5 196.2 174.7 56.0 43.9 50.3 75.2 83.9 8.3 301.0 207.7 74.5 270.4 206.4 78.9 46.3 35.2 162.4 188.7 131.6
0.384 0.245 0.391 0.231 0.235 0.241 0.309 0.655 0.381 0.330 0.373 0.332 0.276 0.255 0.278 0.206 0.271 0.220 0.226 0.121 0.175 0.225 0.280 0.110 0.153 0.150 0.354 0.300 0.289 0.205 0.124 0.348 0.117 0.167 0.328 0.182 0.204 0.272 0.346 0.258 0.184 0.095 0.168
0.99 0.99 0.97 0.99 0.98 0.99 0.95 0.98 0.95 0.98 0.99 0.98 0.98 0.97 0.99 0.99 0.99 0.98 0.95 0.95 0.96 0.95 0.98 0.97 0.97 0.99 0.96 0.99 0.99 0.96 0.94 0.98 0.99 0.98 0.99 0.92 0.99 0.99 0.98 0.95 0.93 0.98 0.99
in the form of cloth or granules in 250 mL of an aqueous solution containing the organic pollutant at a concentration of 100 mg L-1. Equilibrium is reached after 24 h of stirring for fibers and 48 h for granules. Isotherms are modeled on the Freundlich classical model
qe ) K Ce1/n
(II)
where qe is the equilibrium adsorption capacity (mg g-1), Ce the equilibrium solute concentration (mg L-1), K (mg1-1/n L1/n g-1), and 1/n the Freundlich parameter.
Results 1. Adsorption Kinetics and Isotherms. Initial adsorption coefficients and Freundlich parameters obtained for the adsorption of various organic compounds onto CS 1501, RS 1301, and NC 60 are given in Table 2. They show that adsorption is higher with activated carbon cloth than with granular activated carbon, in terms of velocity and capacity. Initial adsorption rates seem to be dependent on the external surface area of activated carbon. The low fiber diameter (10 µm) compared with that of granules (1 mm) allows faster adsorption for the same activated carbon weight. Generally, the adsorption rates are higher with CS 1501 than with RS 1301, due to the ribbed form of CS 1501 fibers which may increase their external surface area.
Furthermore, the microporous character of this ACC, associated with the fact that micropores are directly connected to the external surface area, induces a decrease in the mass transfer resistance. However, in the case of larger compounds (for instance chlorophenol or 2-butyl4-methylphenol), the mesoporous character of RS 1301 seems to favor the adsorption over the CS 1501 microporous character. This tendency is confirmed by adsorption isotherms. Indeed, larger molecular weight compounds such as chlorophenol or vaniline are better adsorbed onto the RS 1301, which contains some mesopores. The majority of organic micropollutants in Table 2 are, however, better removed by the more microporous ACC, the CS 1501 sample, thanks to its high specific surface area (1689 m2 g-1) and to the direct connection of micropores to its external surface area. This woven adsorbent was then chosen to study the adsorption of a large data set of organic molecules and to set up a quantitative structure-property relationship. 2. Quantitative Structure-Property Relationship. 2.1. Choice of the QSPR Method. QSPRs are used to correlate some physical or chemical properties of compounds to their molecular characteristics. Originally, they were conceived to develop drugs,16 but recent studies have extended their use to environmental chemistry and toxicology. Several methods may be used to assess QSPR: the linear solvation energy relationship,17 the solvophobic theory,18 the linear free-energy relationship,19 Molecular connecivity indexes, etc. This last tool, developed by Kier and Hall20,21 from the topological indexes of Randic,22 is related to the structural characteristics of organic compounds.21 Furthermore, they are empirical indexes, easely computed with MolconnZ software23 and available for all compounds of the database. Moreover, some studies have compared several QSPR methods and designated molecular connectivity indexes as the most useful method for adsorption correlation.24,25 2.2. Choice of the Statistical Tools. Several statistical tools are available but two methods were preferred in this study, multiple linear regression and neural network. Their choice is based on the great number of variables and the impossibility of anticipating the kind of function involved in the relation. Indeed, neural networks enable nonlinear relationships between adsorption parameters and molecular connectivity indexes to be assessed. In a neural network, a number of nodes, called neurons, are interconnected into a netlike structure and can perform parallel computation for data processing and knowledge representation. The network consists in general of three different layers, each containing a number of neurons. The input layer contains the input variables (molecular connectivity indexes), which affect the model output, while (16) Dearden, J. C. Chemom. Intell. Lab. Syst. 1994, 24 (2), 77-87. (17) Taft, R. W.; Abboud, J.-L. M.; Kamlet, M. J.; Abraham, M. H. J. Solution Chem. 1985, 14 (3), 153. (18) Belfort, G.; Altshuler, G. L.; Thallam, K. K.; Feerick, C. P., Jr.; Woodfield, K. L. AIChE J. 1984, 30 (2), 197. (19) Abe, I.; Hayashi, K.; Kitagawa, M.; Hirashima, T. Bull. Chem. Soc. Jpn. 1983, 56, 1002. (20) Kier, L. B.; Di Paolo, T.; Hall, L. H. J. Theor. Biol. 1977, 67, 585-595. (21) Kier, L. B.; Hall, L. H. Molecular connectivity in structureactivity analysis; Reas. Studies Press (J. Wiley & Sons): Letchworth, England, 1986. (22) Randic, M. J. Am. Chem. Soc. 1975, 97, 6609-6615. (23) Hall, L. H. MolconnZ software, Dept Chem., Eastern Nazarene College, Quincy, MA, 1988. (24) Blum, D. J. W.; Suffet, I. H.; Duguet, J. P. Water Res. 1994, 28, 687-699. (25) Nirmalakhandan, N.; Speece, R. E. Book of Abstracts; National Meeting of the American Chemical Society; American Chemical Society: Washington, DC, 1989; pp 478-485.
Organic Adsorption on Activated Carbon
Langmuir, Vol. 15, No. 18, 1999 5909
Table 3. Database Used To Assess the QSPR organic compound
log K
3χ v p
5χ v p
6χ v p
3χ
toluene benzaldehyde benzoic acid chlorobenzene anisole nitrobenzene fluorobenzene acetophenone dimethyl acetalbenzaldehyde bromobenzene 4-chlorophenol 4-nitrophenol diethyl phthalate 4-methoxyphenol 4-cresol salycilic acid 4-aminophenol 2-carboxybenzaldehyde 4-iodoaniline 3-nitrobenzoic acid 3-aminobenzoic acid catechol 4-tert-butylphenol 4-ethylphenol vaniline 2-tert-butyl-4-methylphenol atrazine 3-chloro-4-fluoroaniline chlorohydroquinone 2,3-dichloroaniline 2-chloro-5-nitrobenzoic acid 2,4-dichlorobenzoic acid 2,5-dimethylphenol 4-amino-2,6-dichlorophenol 2,4,6-trimethylaniline 4,6-dinitro-o-cresol 2,3,5-trimethylphenol DL-tryptophane amino-4-benzenesulfonamide 1-ethoxy-2-tertbutoxyethane ethylbenzene phenol aniline 4-butylbenzoic acid hydroquinone phthalic acid 4-chlorobenzoic acid 3,5-dimethoxybenzoic acid 3-methylsalycilic acid DL-tyrosine DL-phenylalanine phenyl salycilate bromophenol p-anisidine methylethyldioxalane
2.179 2.027 1.914 2.478 2.158 2.308 2.142 2.271 2.248 2.680 2.119 2.264 2.347 2.192 2.101 1.955 2.355 2.009 2.608 2.105 1.925 1.953 2.475 2.277 2.211 2.479 2.270 2.292 2.326 2.605 1.974 2.067 2.273 2.353 2.392 1.906 2.396 2.304 1.968 1.040 2.58 1.808 1.785 1.603 1.684 1.754 1.260 1.622 1.286 1.897 1.876 1.137 1.684 1.435 1.316
0.941 0.936 1.021 0.985 0.979 0.886 0.733 1.180 1.500 1.262 1.079 0.979 1.918 1.072 1.034 1.135 0.893 1.318 1.589 1.224 1.134 0.882 1.733 1.345 1.340 1.960 1.530 1.233 1.234 1.815 1.641 1.724 1.440 1.616 1.738 1.528 1.904 2.432 2.138 0.967 1.251 0.756 0.800 1.876 0.850 1.377 1.343 1.598 1.465 1.709 1.616 2.078 1.355 1.116 1.539
0.304 0.302 0.330 0.318 0.316 0.298 0.234 0.384 0.505 0.411 0.403 0.335 0.740 0.348 0.381 0.342 0.308 0.423 0.710 0.425 0.383 0.259 0.545 0.448 0.462 0.849 0.564 0.362 0.355 0.407 0.558 0.727 0.367 0.418 0.484 0.531 0.547 1.044 0.742 0.213 0.407 0.242 0.257 0.591 0.286 0.464 0.490 0.619 0.394 0.646 0.597 0.784 0.545 0.369 0.250
0.064 0.063 0.080 0.073 0.072 0.061 0.024 0.110 0.180 0.126 0.088 0.091 0.347 0.117 0.080 0.097 0.057 0.129 0.173 0.149 0.115 0.050 0.309 0.185 0.165 0.249 0.389 0.128 0.119 0.168 0.221 0.195 0.166 0.238 0.263 0.218 0.204 0.555 0.429 0.118 0.124 0.029 0.037 0.364 0.050 0.142 0.213 0.265 0.156 0.265 0.234 0.284 0.134 0.133 0.083
0.289 0.204 0.500 0.289 0.204 0.204 0.289 0.500 0.333 0.289 0.577 0.493 0.744 0.493 0.577 0.705 0.577 0.636 0.577 0.704 0.789 0.471 1.799 0.493 0.607 2.008 1.105 0.760 0.760 0.664 0.909 0.994 0.664 0.953 0.953 0.805 0.856 1.039 1.799 1.561 0.204 0.289 0.289 2.010 0.577 0.939 0.789 0.908 0.916 1.062 0.773 0.812 0.577 0.493 0.402
the output layer contains the desired output variables (adsorbability parameters). The intermediate layer is called the hidden layer, the number of hidden neurons being determined using an iterative process. Each neuron in the network is influenced by those neurons to which it is connected, the degree of influence being dictated by the values of the links called connection weights. The overall behavior of the system can be modified by adjusting the values of the weights, through the repeated application of a learning algorithm, in general the back propagation algorithm.26 2.3. The Database. The database was composed of 55 adsorption isotherms of compounds onto CS 1501. For each compound, 18 molecular connectivity indexes were computed with MolconnZ software,23 using the well(26) Basheer, I. A.; Najjar, Y. M.; Hajmeer, M. N. Environ. Technol. 1996, 17, 795-806.
c
4χ
v pc
0.193 0.150 0.220 0.218 0.147 0.109 0.073 0.331 0.342 0.378 0.304 0.195 0.505 0.233 0.279 0.333 0.197 0.408 0.599 0.318 0.317 0.207 1.298 0.340 0.393 1.421 0.422 0.467 0.478 0.952 0.645 0.745 0.617 0.761 0.850 0.575 1.008 0.821 1.083 0.433 0.254 0.086 0.111 1.153 0.172 0.454 0.438 0.480 0.591 0.518 0.432 0.46 0.464 0.258 0.548
defined method of Kier and Hall.21 The adsorbability parameter was chosen as log K because, using the Freundlich equation, we obtain
K ) lim
( ) qe
Ce1/n
(III)
Ce)1mgL-1
where K is the Freundlich parameter (mg1-1/n L1/n g-1) defined by eq II, Ce the solute equilibrium concentration (mg L-1), and qe the equilibrium adsorption capacity (mg g-1). Indeed, several descriptors may be used to characterize the adsorption, but they are generally similar to log(qe/Ce)min. The final database, comprising the molecular connectivity indexes used for the MLR and NN, and the adsorbability parameters log K of the 55 organic compounds, is given in Table 3.
5910 Langmuir, Vol. 15, No. 18, 1999
Brasquet and Le Cloirec
Figure 5. Architecture of the three-layer neural network used for the QSAR study.
3. Numerical Results. 3.1. Multiple Linear Regression (MLR). By use of a stepwise procedure, the following regression was obtained:
log K ) 3.40 - 2.03 3χv + 1.42 5χv + 3.47 6χv 1.39 3χc + 2.55 4χpcv (IV) n ) 55; raj2 ) 0.450; sy ) 0.285 The statistical quality of this relation was low, but the Student parameters t associated with each independent variable (>2) showed that there was a relationship between molecular connectivity indexes and adsorbability.27 The statistical quality of the MLR was improved by removing six compounds from the database, because of their very particular topology (methylethyldioxalane, 2-tertbutyl-4-methylphenol) or a poor modeling by the Freundlich equation (anisidine, 4-chlorobenzoic acid, bromophenol, 3-methylsalycilic acid). The equation was then:
log K ) 3.33 - 1.55 3χv + 0.58 5χv + 3.52 6χv 1.42 3χc + 2.29 4χpcv (V) n ) 49; raj2 ) 0.648; sy ) 0.199 The Student parameters t are all higher than 2, representing significant variables. All molecular connectivity indexes are related to physicochemical properties.21 Some of them carry redundant information, but also some specificities. The importance of the electronic content of the molecules is explained by the lack of simple (27) Stein, R. AI Expert 1993, 8 (2), 42-47.
indexes. Their volume (3χv), branching number (3χc), substitutent length (6χv) and heteroatoms (5χv), and the global substitution pattern (4χpcv) will also be influential. Although the general quality of the regression was improved by removing these six compounds, it was still low indicating that the relation between log K and molecular structure would not be linear. A neural network approach was then tested, using a nonlinear function as the transfer function. 3.2. Neural Network. In our application of a neural network (NN) to adsorption and molecular structure parameters, a classical and simple network was developed. The architecture was optimized by varying the transfer function and the number of hidden neurons. The agreement between calculated log K and observed log K values was tested using statistical parameters such as the determination coefficient r2 and the standard deviation (Statistica). The final architecture consists of a hyperbolic tangent function between input and output data. Only one hidden layer was used, containing three hidden neurons. The input layer contains the same molecular connectivity indexes as those used in the MLR. The schematic representation of this neural network is given in Figure 5. The training step enabled log K calculated values to be plotted as a function of log K observed values, and Figure 6 compares the results obtained with MLR and NN. It is obvious that NN seems to be more effective for describing adsorbability data than MLR. This descriptive ability was used to carry out a variable analysis (molecular connectivity indexes) in order to explain the adsorption mechanisms of organic pollutants onto CS 1501. 3.3. Comparison of Both Statistical Tools. Figure 6 shows that a neural network gives a better representa-
Organic Adsorption on Activated Carbon
Langmuir, Vol. 15, No. 18, 1999 5911 Table 5. Relative Importance of Variables on the Adsorbability Parameter log Ka relative importance (%) influence on log K a
Figure 6. Comparison of the descriptive abilities of MLR and NN. Table 4. Comparison of the Predictive Abilities of MLR and NN MBEa RMSEb a
MLR
NN
0.082 0.231
0.179 0.462
MBE, mean bias error. b RMSE, root mean square error.
tion than MLR. However, both statistical tools have different weight numbers, which may explain the difference in their descriptive ability. To really assess both tools, their prediction abilities were compared using a database of 10 compounds whose molecular structures were well represented in the training database but were not included in this database. The database was then divided into a training database of 45 compounds and a prediction database of 10 compounds whose structures were well described in the training database (toluene, acetophenone, bromobenzene, salycilic acid, 4-aminophenol, vaniline, 2,4dichlorobenzoic acid, 2,4,6-trimethylaniline, hydroquinone, 3,5-dimethoxybenzoic acid). Results are compared in Table 4 using the mean bias error and the root-meansquare error:
mean bias error (MBE) )
∑res
root mean square error (RMSE) )
(VI)
n
∑res2
x
n
(VII)
where residual (res) ) observed log K - calculated log K. These results show that from a predictive point of view, MLR is better than a neural network. Finally, this study seems to confirm the good descriptive ability of the neural network for data such as solubility28,29 or boiling point30 but shows its poor predictive ability, which has been rarely studied in the literature apart from Livingstone’s study.31 3.4. Variable Analysis. The descriptive ability of NN was, nevertheless, used to assess the relative importance of molecular connectivity indexes to the adsorbability parameter in order to explain adsorption mechanisms. The Garson weight partitioning method was used32 to (28) Bodor, N.; Harget, A.; Huang, M. J. Am. Chem. Soc. 1997, 113, 9480-9483. (29) Tetteh, J.; Metcalfe, E.; Howells, S. L. Chemom. Int. Lab. Syst. 1996, 32, 177-191. (30) Cherquaoui, D.; Villemin, D. J. Chem. Soc., Faraday Trans. 1994, 90, 97-102. (31) Livingstone, D. Data analysis for chemists; Oxford Science Publishers: 1995; 239 pages. (32) Garson, G. D. AI Expert 1991, 4, 47-51.
3χv
5χv
6χv
3χ c
19.6 V
16.9 v
36.0 v
11.4 V
4χ
v pc
16.1 v
Key: V, negative influence; v, positive influence.
calculate these relative importances. Furthermore, the signs of the MLR coefficients determined wether the influence of each index was positive or negative. The results are given in Table 5. 4. Discussion. To sum up the results, the adsorbability parameter log K was related to five molecular connectivity indexes 3χv, 5χv, 6χv, 3χc, and 4χvpc. The most influential index is 6χv; the influences of the three indexes 3χv, 5χv, and 4χvpc are quite similar and higher than that of 3χc. Although the influence of order 3 indexes is negative, the influence of other indexes is positive. From the positive or negative influence of each index and their physical meaning (section 3), we may assume that the adsorption of bulky molecules is not favored. This could be due to an adsorption of molecules between the basic planes constituting the activated carbon fibers. This result confirms a QSPR study carried out with the LSER method,33 which introduced a flatness parameter. Once located between these planes, molecules would be involved in electronic interactions with the surface of activated carbon fibers. It seems that electron-withdrawing substitutents such as halogens (Br, Cl, I, etc., inducing high 5χpv values) increase the adsorption capacity for aromatic compounds. Thus the benzene ring may accept electrons from the active carbon surface. Furthermore, the result that aromatic compounds with large substitutents (4χvpc, 6χpv) seem to be well removed by CS 1501 may be due to the fact that a decrease in their affinity with the aqueous phase induces an increase in their affinity with the adsorbent. All these assumptions could have been made qualitatively with the large adsorption database, but here they are confirmed and even quantified by the QSPR method. Furthermore, another fact is shown: the better adsorption of flat molecules. Conclusion The objective of this work was to study the dependence of adsorption on adsorbent and adsorbate characteristics. The experimental study carried out in a batch reactor showed the higher adsorption kinetics and capacities of activated carbon cloths compared with granular activated carbon, due to their large external surface area and their porous specificities (high specific surface areas and a direct connection of their micropores with their external surface area). Furthermore, the comparison of a microporous ACC (CS 1501) and a mesoporous one (RS 1301) showed the better performances of the microporous adsorbent for microorganics. On the other hand, for removal of larger organic compounds, the mesoporous ACC is more effective. This influence of adsorbate characteristics was quantified with QSPR using molecular connectivity indexes and assessed by two statistical tools: multiple linear regression and neural network. The statistical quality of the relation set up by MLR was low, but the neural network gave a good description of adsorption data. However, the predictive ability of the NN was low. A variable analysis allowed some assumptions to be made about the adsorption (33) Luehrs, D. C.; Hickey, J. P.; Nilsen, P. E.; Godbole, K. A.; Rogers, T. N. Environ. Sci. Technol. 1996, 30, 143-152.
5912 Langmuir, Vol. 15, No. 18, 1999
mechanisms of organics onto ACC: it seems that the adsorption is located between the basic planes of the activated carbon, favoring a better adsorption of flat molecules. The adsorption seems to occur by an electron donor-acceptor reaction between the activated carbon surface and the aromatic ring of the solute. Therefore, electron-withdrawing substitutents favor the adsorption process.
Brasquet and Le Cloirec
Acknowledgment. The authors thank E. Subrenat, Actitex, and G. Dagois, Pica Co. (Levallois, France), for financial support of this work and B. Rousseau and H. Estrade-Sczwarckopf, Centre de Recherche sur la Matie`re Divise´e (Orle´ans, France), for their atomic force microscopy observations of activated carbon fibers. LA9811160