Chemical Engineering Science 57 (2002) 2641 2652

www.elsevier.com/locate/ces

The e ect of a crystal size distribution on ZLC experiments

W. L. Duncan, K. P. Moller
Catalysis Research Unit, Department of Chemical Engineering, University of Cape Town, Private Bag, Rondebosch, 7701, South Africa Received 4 December 2000; accepted 15 January 2002

Abstract This paper reports the results of a theoretical investigation of the e ect of a sorbent size distribution on the zero length column (ZLC) method for measuring intracrystalline di usion. A new model incorporating a continuous log-normal distribution of crystal size is introduced. It is shown that such a distribution introduces tailing into the normally linear long time region of the desorption curve, although this curvature is not obvious when the initial concentration drop is fast. Analysing such a curve with the standard ZLC model causes the di usional time constant to be underpredicted, whereas the adsorption related parameter L is overpredicted. The error increases with increasing distribution width. It is shown that the curve may be well approximated by summing the desorption curves of the individual, discrete size fractions, weighted by their volume fraction. The systematic error introduced into the measured di usion coe cient when analysing data for sorbent samples with a size distribution with the standard model may be reduced if the mean particle size is dened as the ratio of the samples volume to its surface area. ? 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Zero length column; Di usion; Zeolites; Modelling; Size distribution; Numerical analysis

1. Introduction The zero length column (ZLC) method is a well established chromatographic technique for the measurement of di usion in sorbent powders, particularly zeolite crystals with gaseous hydrocarbon sorbates. While the experimental principle and procedure are simple and have been extensively documented (Eic & Ruthven, 1988; Hufton & Ruthven, 1993; Duncan & Moller, 2000a), many methods for interpreting the data have been postulated. In the simplest case, the sorbent is assumed to be isotropic and monosized, the carrier phase well mixed and the sorbate adsorption isotherm linear. These assumptions lead, for spherical particles, to the following model for the concentration of the sorbate in the column e uent (Eic & Ruthven, 1988): C = 2L C0

and n are given by the roots of n cot n + L 1 = 0: For long times and large L this simplies to C 2 = e C0 L
2

(4)

(5)

2 n=1 n

2 exp(n ) ; + L(L 1)

(1)

where D = 2 t; R FR2 L= 3Vs KD

(2) (3)

Corresponding author. Fax: +27-21-689-7579. E-mail address: km@chemeng.uct.ac.za (K. P. Moller).

This latter equation has led to the most widely used analysis approach, the long time method (Hufton & Ruthven, 1993), in which the di usion and adsorption parameters are extracted from the desorption curve by tting a straight line to the linear (on a semilogarithmic scale) tail. Various other short time and intermediate time (Hufton & Ruthven, 1993; Brandani & Ruthven, 1996a; Loos, Verheijin, & Moulijn, 2000) solution methods, as well as a moments-based method (Brandani & Ruthven, 1996b), have been proposed, but are not widely used. It has been shown that, generally, deviations from the standard model assumptions tend to a ect the accuracy of L more than that of D=R2 , i.e. the slope of the tail tends to give the correct di usional time constant (Eic & Ruthven, 1989; Brandani & Ruthven, 1996a; Brandani, 1998; Duncan & Moller, 2000b). There are, however, several examples of data in the literature in which the long time region of the desorption curve is not convincingly linear, but shows a tailing e ect (i.e. it is convex to the time axis) (Voogd, Bekkum, Shavit, &

0009-2509/02/$ - see front matter ? 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 0 0 9 - 2 5 0 9 ( 0 2 ) 0 0 1 6 1 - 6

2642

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Kouwenhoven, 1991; Hufton & Ruthven, 1993; Brandani, Jama, & Ruthven, 2000; Loos et al., 2000; Duncan & Moller, 2000a). The only theoretical work to attempt to describe this behaviour is that incorporating the e ect of a distribution of discrete sorbent particle sizes (Loos et al., 2000), although the e ect is not readily apparent as the authors of that work chose a logarithmic time scale when plotting their desorption curves. Analogous behaviour has been shown for the jet loop reactor (Moller & OConnor, 1996). This paper quanties the e ect of a crystal size distribution on the ZLC desorption curve and the deviation expected when modelling such curves using the standard long time approach. This is done by comparing Eq. (1) to a model incorporating a continuous, log-normal size distribution. A second model, for possible use when experimental size distribution data are available, is also developed. Finally, a mean particle size is dened in such a way as to reduce the error in calculating di usivity when using the standard model to analyse a sorbent sample showing a size distribution. 2. Theoretical 2.1. Continuous size distribution model Ficks Second Law in spherical geometry is used to describe transport within the adsorbed phase: @q =D @r @ q 2 @q + @r 2 r @r
2

The uid phase mass balance, neglecting interstitial sorbate accumulation (Brandani & Ruthven, 1995), takes the form 4 Vs D
2 R (@q=@r)|r=R p(R) dR 0 3 4 R p(R) dR 3 0

= FC:

(11)

These equations were solved in dimensionless form in the Laplace domain, making use of the following substitutions: R=e X=
m + 2

(12) (13) (14) (15)

r ; Rm D t; R2 m

Rm = e :

It should be noted that Rm = R, i.e. the former is not the number mean particle radius, which is instead given by R= =e
0

R p(R) dR
2

+0:5

= Rm e0:5 :

(16)

(6)

The solution for the desorption curve in the Laplace Domain is ; (17)

2 [(e 2 s) coth(e 2 s) 1]e 2 d C 1 = C0 s Lm e4:5 2 + [(e 2 s) coth(e 2 s) 1]e 2

where with the standard initial and boundary conditions q(r; t = 0) = KC0 ; q(r = R; t) = KC; @q (r = 0; t) = 0: @r (7) (8) (9) Lm = FR2 m : 3Vs KD (18)

This may be compared to the solution of the standard model (i.e. = 0) in the Laplace domain (Brandani & Ruthven, 1996a): C s coth s 1 1 = : (19) C0 s L + s coth s 1 Eq. (17) may be solved numerically using the Fast Fourier Transform algorithm, with the indicated integral being evaluated by GaussHermite quadrature. Since a dimensionless time scale is used, the model has two independent parameters, namely Lm , which is similar to L in the standard model, and , which denes the width of the particle size distribution. The usual third experimental parameter, D=R2 , denes the real time length of the experiment and has no e ect on the shape of the desorption curve. Size distributions are shown in Fig. 1 for various values of at constant (i.e. constant Rm ). It may be seen that

The sorbent particles are assumed to follow a log-normal size distribution, which is common for zeolite crystals (Hsu & Haynes, 1981): p(ln R) = Rp(R) = 1 2 exp (ln R )2 2 2 ; (10)

where represents the number mean of the normal distribution of the logarithm of particle size and its standard deviation. Literature values for the distribution width include = 0:90 (Fu, Ramesh, and Haynes, 1986) and = 0:60 (calculated from data given in Haq & Ruthven (1986)).

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

2643

Fig. 1. Particle size distribution plot in arbitrary units for

= 0 and values of

as indicated.

increasing not only increases the width of the distribution, but also skews it toward the larger particles. Fig. 2 shows desorption curves generated using Eq. (17). For = 0 the curve corresponds to that of the standard model (Eq. (1)), as would be expected. When the width of the size distribution is increased, the normally linear tail becomes increasingly curved. The slope of the tail is, of course, dependent on the di usional time constant D=R2 ; since multiple values of R exist, as described by the size distribution, the slope cannot be constant. The large increase in experimental duration is due to the chosen distribution model being skewed towards the larger particles for large . 2.2. Discrete size distribution model One may write the uid phase mass balance in the form FC = Vs dq ; dt (20)

with Li = FR2 i ; 3Vs KD D i = 2t Ri Vsi ; Vs (23) (24)

and vi = (25)

where Ri is the representative radius of size fraction i and C=C0 ( i ; Li ) is calculated by Eq. (1). Note that the denition of Li here is an approximation only, in that it has been assumed that the solution for the desorption curve of each fraction is linear with respect to L. This is only valid for the case of large L and long times, where the analogue of Eq. (5) may be written as C = C0
N

vi
i=1

where q is the space-averaged adsorbed phase concentration. Considering N discrete sorbent size fractions, this may be rewritten as
N

2 e Li

(26)

3. Results and discussion 3.1. Comparison of size distribution and standard models In order to quantify the deviation from the standard ZLC model caused by non-monosized sorbent particles, desorption curves were generated with Eq. (17) and Eq. (1) was then t to them. The comparison is based on the apparent values of L and D=R2 obtained from the regression compared to the values used to generate the original curves.

FC =
i=1

dq Vsi i ; dt

(21)

where Vsi is the volume of sorbent in fraction i. It can then be seen that the overall desorption curve is given by the sum of the desorption curves for the discrete size fractions, i.e. C = C0 vi
i

C ( i ; Li ) C0

(22)

2644

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Fig. 2. Desorption curves generated using the size distribution ZLC model (Eq. (17)) for L = 20 and various values of model (Eq. (1)) is shown for verication.

as shown. The standard ZLC

All points in the range 0:001 6 C=C0 6 1 were considered. The results are shown in Figs. 3 and 4. Examples of the model ts obtained by this procedure are shown for various L and in Figs. 5 and 6. The curvature in the desorption curve tail is most evident for the lower L value. At high L most of the tail is below the observed concentration range, so it is possible to obtain a good model t in this regime although, as may be seen from Figs. 3 and 4, the parameters one obtains from the t may be subject to a substantial error. This problem would be practically impossible to identify in an experiment. One should note that, although the errors introduced in this manner are relatively small (generally less than a factor of three), they are systematic and additive to the random error inherent in any experimental work. It is evident from Figs. 5 and 6 that, for a series of experiments at constant temperature and varying ow rate, a plot of L versus ow rate would tend to rise sharply, then atten o somewhat into a linear region. D=R2 would appear to increase with ow rate, the dependency becoming decreasingly marked at increasing ow rate. This may go some way toward explaining similar behaviour previously published in the literature (Duncan & Moller, 2000a). 3.2. Description of experimental data with the crystal size distribution model With the above observation in mind, it is tempting to t Eq. (17) to experimental data. It is, however, important to note the quality of the information this is likely to produce. It

is inevitable that the model will t the experimental desorption curves better: it has an extra parameter that can provide curvature. This will be true regardless of the physical meaning of the third parameter. It is important to check the consistency of the parameters over several experiments, rather than assuming that the model is more accurate because it provides a better t. Fig. 7 shows an example how Eq. (17) is able to t an experimental desorption curve far better than the standard model (Eq. (1)). The Arrhenius plot in Fig. 8 conrms that the former tends to give larger values for the di usional time constant, as was indicated in Fig. 3, but it is also apparent that some scatter is introduced into the data. The activation energy given by the standard model is 52:8 kJ=mol, while the size distribution model indicates a value of 65:8 kJ=mol, a signicant increase. Most signicant, however, is the wide range of obtained for these data, as shown in Fig. 9: a constant value of should have been obtained for all experiments if Eq. (17) were applicable to the sample. As would be expected if the size distribution was not the same shape as assumed, or if the curvature in the desorption curve was caused by a di erent e ect, decreases with increasing L as less of the tail is visible. The scatter is not likely to be caused by model insensitivity to the adjustable parameters; several of the regression runs were repeated with di erent initial guesses, all resulting in similar solutions. On the other hand, increasing the number of adjustable parameters in the curve t (to three from the usual two) would tend to increase the variance in the estimates.

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

2645

Fig. 3. Apparent deviation in di usional time constant when tting the standard ZLC model to that with a crystal size distribution, for various distribution widths.

Fig. 4. Apparent deviation in L when tting the standard ZLC model to that with a crystal size distribution, for various distribution widths.

Quite clearly, Eq. (17) does not adequately describe the experimental data, probably simply due to the shape of the crystal size distribution. The data did not show any signicant trends in with temperature, ow rate or initial sorbate concentration.

3.3. Approximate models to account for a particle size distribution An approximate method of evaluating the desorption curve for a non-monosized sorbent has been previously

2646

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Fig. 5. Model comparison for the standard model tted to the crystal size distribution model when

= 0:3 for L as shown.

Fig. 6. Model comparison for the standard model tted to the crystal size distribution model when

= 0:5 for L as shown.

suggested in the literature (Loos et al., 2000): one simply adds the normalised responses of each (discrete) size fraction. The fractions are weighted according to the fraction ai of the total surface area of the sample that each one contains. Each size fraction is represented by its largest dimension, which denes its di usional time constant and value of L.

The desorption curve is then given by C C = ai ( i ; Li ) C0 C0 i

(27)

analogously to Eq. (22), with ai being the fraction of the total surface area contained in size fraction i.

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

2647

Fig. 7. Example of model t for standard and size distribution models to an experimental desorption curve (T = 200 C, F = 41:6 ml=min, ms = 2 mg).

Fig. 8. Temperature dependence of the di usional time constant as determined by both the standard and size distribution models.

To compare this approximate approach and the volume fraction based method developed above to the full solution (Eq. (17)), ten size fractions were dened as shown in Table 1, each having an approximately equal surface area or volume (ai = vi = 0:1). The resulting desorption curves are shown in Figs. 10 and 11.

It can be seen that the volume based approximation generally corresponds to the full model better than does the area based method. This is unsurprising since it has been shown above to be exact for large L and long times. Curves at large L show very good agreement between the exact and approximate models for both indicated distribution widths, as does

2648

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Fig. 9. Crystal size distribution width as a function of L for all points shown in Fig. 8.

Fig. 10. Desorption curves plotted using the full and approximate size distribution models for

= 0:3 and L as indicated.

the example for small L and wider distribution. The exception is the curve for L = 20 and = 0:3; this is likely the result of L not being quite large enough that the approximations in Eq. (23) hold. The obvious advantage to this sort of approach is that it is not tied to a specic shape of size distribution. This makes

it a superior choice in practical terms over the analytical sort of approach leading to Eq. (17). In fact, it is possible that one may introduce more scope for error by trying to describe the real distribution with a simple analytical one and using an exact model than simply using this semi-empirical approach.

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

2649

Fig. 11. Desorption curves plotted using the full and approximate size distribution models for

= 0:5 and L as indicated.

Table 1 Division of a continuous size distribution into discrete fraction such that all fractions have equal surface area or volume; = 0 ai = 0:1 Size range 0.3 0.20 0.81 0.81 0.93 0.931.02 1.021.11 1.111.20 1.20 1.29 1.29 1.40 1.40 1.54 1.54 1.76 1.76 4.00 4 R2 = 15:04 0.10 0.87 0.871.08 1.081.27 1.271.45 1.45 1.65 1.65 1.87 1.872.14 2.14 2.51 2.513.12 3.1212.00 4 R2 = 20:72 Area in fraction 1.45 1.56 1.46 1.56 1.54 1.43 1.52 1.51 1.52 1.50 =15:05 2.08 2.04 2.12 2.03 2.11 2.05 2.06 2.09 2.06 2.08 =20:72 vi = 0:1 Size range 0.20 0.89 0.89 1.02 1.021.12 1.121.21 1.211.31 1.311.41 1.411.53 1.531.68 1.681.91 1.91 6.00 4 R3 = 6:280 3 0.10 1.10 1.10 1.38 1.381.62 1.621.86 1.86 2.11 2.112.40 2.40 2.75 2.75 3.22 3.22 4.01 4.0112.00 4 R3 = 12:90 3 Volume in fraction 0.620 0.649 0.619 0.596 0.656 0.608 0.633 0.621 0.622 0.654 =6:278 1.23 1.30 1.29 1.31 1.28 1.31 1.30 1.29 1.29 1.30 =12:90

distribution data. It is also useless if the sorbent does not have the type of size distribution that has been assumed here. Inspection of Eq. (17) suggests that a starting point for a practically more useful denition of mean particle size would be Rmean = e
+2:25
2

(28)

By analogy with Eq. (16), the volume and area based mean radii may be found from the relations R3 = e 3 R2 = e 2
+4:5
2

(29) (30)

+2

and it is easy to see that R3 R2 =e

+2:5
2

0.5

+2:25

(31)

3.4. Denition of mean particle size While Eq. (15) is the mathematically most elegant way to dene the mean particle radius, it is certainly not the most intuitive method to use when dealing with experimental size

Thus the common engineering approximation, taking the representative dimension as essentially the ratio of volume to area, would be reasonable in this case, and probably for other distributions as well. This denition of mean size is useful for the case when insu cient particle size data are available to use, for example, Eq. (22) makes it feasible to use the standard model (Eq. (1)). The e ect of this choice may be observed by comparing Figs. 12 and 13 to Figs. 3 and 4, respectively: the apparent model parameters become less severely over- or underpredicted, although the ow rate dependence is exaggerated.

2650

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Fig. 12. Apparent deviation in di usion time constant when tting the standard ZLC model to that with a crystal size distribution, for various distribution widths, including the correction factor indicated by Eq. (31).

Fig. 13. Apparent deviation in L when tting the standard ZLC model to that with a crystal size distribution, for various distribution widths, including the correction factor indicated by Eq. (31).

3.5. Some practical considerations When packing the amounts of sorbent generally used in ZLC work, i.e. a couple of milligrams, it is very di cult to obtain a representative sample. Thus, even when size distribution data are available, they are not neces-

sarily useful. This may be simply demonstrated by the cyclohexane=ZSM-5 experiments shown in Fig. 14; D=R2 was obtained by tting the standard model. For the lower two sets of data, the sorbent was allowed to settle for a day on a vibrating surface and then the material at the top of the container packed in an attempt to get a narrow distribution

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

2651

Fig. 14. Arrhenius plot showing series of experiments where the sorbent was allowed to settle or was agitated before packing.

of the largest material. For the upper two sets, the container was shaken before packing to get the widest distribution possible. The measured di usional time constants for the similarly packed samples di er by approximately a factor two, agreeing with the factor two accuracy previously proposed for the ZLC technique (Brandani, 1998). There was no notable di erence, visually, in the quality of the model t between the settled and agitated samples. As can be seen by the data provided in the gure, the di usional activation energies are in good agreement. A further consideration is the source of the size distribution, which will probably be well dened and relatively easy to measure for spherical particles (provided they may be considered isotropic) and well-formed single crystals (e.g. perfect ZSM-5 crystals). Twinned zeolite crystals, however, would be more di cult to describe from this perspective, as indeed would any anisotropic sorbent particles, where the denition of the characteristic dimension is unclear (Duncan & Moller, 2000a), although an attempt has been made to elucidate e ective geometry through the consideration of a shape factor (Loos et al., 2000). 4. Conclusions Using an analytical model, the e ect of a sorbent particle size distribution on the ZLC desorption curve has been demonstrated. A non-monosized sorbent introduces curvature into the normally linear long time asymptote (when plotted on a semilogarithmic scale). This is expected, as the

slope of this linear region is given by the di usional time constant, D=R2 . A size distribution introduces a range of values of R, and thus a varying slope. When the standard model is used to interpret data obtained from a non-monosized sorbent sample, D=R2 tends to be underpredicted while L is overpredicted. This systematic error increases with the distribution width and decreases with increasing L. The model incorporating the size distribution is unsuited to modelling experimental data as it assumes a log-normal distribution. It is more advisable to use a sum of the standard model for each size fraction, weighted by volume fraction. If size distribution data are not available, the systematic error in D when using the standard model may be decreased by dening the average particle radius as the ratio of the volume of the sample to its surface area.

Notation C D F K p q r R Rm s t uid phase sorbate concentration intracrystalline di usion coe cient volumetric ow rate of purge uid dimensionless Henry adsorption constant probability density function adsorbed phase sorbate concentration radial position within sorbent particle sorbent particle radius mean radius; dened by Eq. (15) Laplace parameter corresponding to m or time

2652

W. L. Duncan, K. P. Moller / Chemical Engineering Science 57 (2002) 26412652

Vs X

volume of sorbent dimensionless radial coordinate; dened by Eq. (13)

Greek letters n eigenvalues of standard ZLC model; dened by Eq. (4) number mean of normal distribution of ln(R) standard deviation of normal distribution of ln(R) dimensionless time for =0; dened by Eq. (2) dimensionless time; dened by Eq. (14) dummy variable

References
Brandani, S. (1998). E ects of nonlinear equilibrium on zero length column experiments. Chemical Engineering Science, 53(15), 2791 2798. Brandani, S., Jama, M., & Ruthven, D. (2000). Di usion, self-di usion and counter-di usion of benzene and p-xylene in silicalite. Microporous and Mesoporous Materials, 35 and 36, 283300. Brandani, S., & Ruthven, D. M. (1995). Analysis of ZLC desorption curves for liquid systems. Chemical Engineering Science, 50(13), 20552059. Brandani, S., & Ruthven, D. M. (1996a). Analysis of ZLC desorption curves for gaseous systems. Adsorption, 2, 133143. Brandani, S., & Ruthven, D. M. (1996b). Moments analysis of the zero length column method. Industrial and Engineering Chemistry Research, 35, 315319. Duncan, W. L., & Moller, K. P. (2000a). On the di usion of cyclohexane in ZSM-5 measured by zero-length-column chromatography. Industrial and Engineering Chemistry Research, 39(6), 21052113.

Duncan, W. L., & Moller, K. P. (2000b). A zero length criterion for ZLC chromatography. Chemical Engineering Science, 55(22), 54155420. Eic, M., & Ruthven, D. M. (1988). A new experimental technique for measurement of intracrystalline di usivity. Zeolites, 8, 4045. Eic, M., & Ruthven, D. M. (1989). Intracrystalline di usion of linear para ns and benzene in silicalite studied by the ZLC method. In Jacobs, P.A., & van Santen, R.A. (Eds.), Zeolites: Facts, gures, future (pp. 897905). Amsterdam: Elsevier Science Publishers, B.V. Fu, C. C., Ramesh, M. S. P., & Haynes, H. W. (1986). Analysis of gas chromatography pulse dispersion data for the system n-butane=zeolite NaY. AIChE Journal, 26(16), 2406. Haq, N., & Ruthven, D. M. (1986). Chromatographic study of sorption and di usion in 4A zeolite. Journal of Colloid and Interface Science, 112(1), 154163. Hsu, L. K. P., & Haynes, H. W. (1981). E ective di usivity by the gas chromatography technique: Analysis and application to measurement of di usion of various hydrocarbons in zeolite NaY AIChE Journal, 27(1), 8191. Hufton, J. R., & Ruthven, D. M. (1993). Di usion of light alkanes in silicalite studied by the zero length column method. Industrial and Engineering Chemistry Research, 32(10), 23792386. Loos, J.-B. W. P., Verheijin, P. J. T., & Moulijn, J. A. (2000). Improved estimation of zeolite di usion coe cients from zero-length column experiments. Chemical Engineering Science, 55, 5165. Moller, K. P., & OConnor, C. T. (1996). The measurement of di usion in porous catalysts using a CSTR. Chemical Engineering Science, 51(13), 34033408. Voogd, P., Bekkum, H. V., Shavit, D., & Kouwenhoven, H. W. (1991). E ect of zeolite structure and morphology on intracrystalline n-hexane di usion in pentasil zeolites studied by the zero-length column method. Journal of the Chemical Society, Faraday Transactions, 87(21), 35753580.