You are on page 1of 18

TAYLOR DISPERSION THEORY

3. TAYLOR DISPERSION THEORY


In chapter 1 the important role of diffusion in liquid-liquid extraction has been pointed out. For a quantitative description of a ternary extraction, we need the diffusion coefficients of the system. Measuring diffusion coefficients used to be, and still is, a rather time-consuming occupation. There are two methods in common use: the diaphragm cell method [13] and the Taylor dispersion method. This chapter deals with the latter, which is named after G.I.Taylor, who, in the 1950s, published a number of articles on the dispersion in fluid media flowing through a tube [15, 16]. Although it has taken a while before the principles he discussed were actually applied, the method is now recognised as a fast and reliable way to measure diffusivities in homogeneous fluids. This chapter consists of two parts. To start with, we look at the theory of the dispersive process, but not in too great a detail, because this has been done by others. Instead, we will pay special attention to the design of the dispersion tube. In the second part, we will see how to obtain multicomponent diffusivities from dispersion measurements. 3.1 THE TAYLOR DISPERSION EXPERIMENT To decide what equipment exactly is needed, and how it must be designed, we must be acquainted with the theory behind the processes on which the measurements are based. Thus we will briefly discuss this theory, and then some design rules. 3.1.1 Theory of the dispersive process When a fluid flows through a tube, a velocity profile develops. This profile is not uniform over the cross-section of the tube: at some radial positions, the fluid flows faster than at others. The laminar profile for Newtonian fluids has a parabolic velocity distribution. Matter present in the fluid moves along with it and is therefore dispersed along the tube axis. Apart from this purely convective dispersion, diffusion can also be of influence. Suppose that somewhere in the tube we have a small plug of fluid with a composition different from an otherwise similar bulk of fluid. The flow of the fluid causes this plug to disperse, but it also induces radial composition gradients. This leads to diffusive fluxes at the front and back sides of the plug. These fluxes become important if the radial diffusive fluxes are roughly of the same order of magnitude as the convective axial fluxes. This is the case when either the axial velocity is very low, or when the radial distances are very small. When either condition is satisfied, diffusion tends to keep the plug together, in contrast to what one might intuitively expect. So, through the combined action of convection and diffusion, the plug will leave the tube as a broadened, but still more or less compact plug. To achieve this situation, Taylor dispersion experiments are usually carried out in tubes with small diameters, also called capillaries.

25

Chapter 3 Taylor considers the dispersion of a Dirac-pulse of a tracer fluid in a laminar flow. The tube is circular with a diameter d, and the bulk fluid has a superficial velocity usup. The distribution after t seconds is given by c1 z , t =

b g

2 n1E d2

F dz u t i I exp G GH 4 t JJK t
2 sup

(1)

Here, c1 is the concentration of component 1 (the tracer) minus its concentration in the eluent, n1E is the excess number of tracer moles in the pulse, is the dispersion coefficient, and z is the down-stream distance from the initial location of the Dirac-pulse. The approximate expression for given by Taylor [15] is =
2 usup d2

192 D

(2)

where D is the molecular Fick diffusivity of the tracer with respect to the eluent. Aris [2] showed that the exact solution is =
2 usup d2

192 D

+D

(3)

Later, Price [12] derived expressions for the concentration distributions for three components in a way analogous to Taylor, and which are therefore not exact.

F dz u t i I 2 K c bz, t g = exp G GH 4 t JJK + d d t F dz u t i I 2 K 2K exp G c bz, t g = GH 4 t JJK + d d t


1

2 K1

sup 1

sup 1

F dz u t i I exp G GH 4 t JJK t F dz u t i I exp G GH 4 t JJK t


2 sup 2 2 sup 2 i +1

(4)

with i = 0 i* K1 = where 0 =
2 usup d2

192 D

, i* =

dD

2,2

E E * 2 n1 D1, 2 n2

disc D
1,1 E E * 2 n2 D2 ,1 n1

, K2

etr D + b1g dD i n D =
1 2 2,2 * 1 E 1

disc D n2E

j
(5)

1, 2

disc D

K3

dD =

disc D

, K4

dD =

1,1

* 1 n2E D2 ,1 n1E

disc D

Strictly speaking, components 1 and 2 are not interchangeable except when the total number of excess moles in the pulse is zero, that is, when the partial molar volumes are equal and constant. But this is a prerequisite for the derivation of eq. (1) anyway.

26

TAYLOR DISPERSION THEORY So, by monitoring the concentration distribution we are able to determine the dispersion coefficient(s). 3.1.2 The capillary: coiled or straight?

To obtain a concentration distribution at the end of the capillary which has spread enough to allow us to extract the value of the diffusivity with a reasonable accuracy, the dispersive process must have had enough time to do its work. For several reasons outlined in this section, we cannot decrease the fluid flow at will, and therefore the capillary must have a definite length, which turns out to be about 20 m. This makes it desirable to coil the capillary: this configuration saves space, makes the capillary less vulnerable, shortens data communication lines, and enables better temperature control. But can we do it without any repercussions? Arguments for a straight capillary are of a theoretical nature: the solutions of the mathematical equations governing the Taylor dispersion dynamics are simpler and require fewer simplifications in the case of a straight capillary. These solutions may therefore be assumed to be more accurate, and also fitting these solutions to measured RTD curves favours simple equations. Usually, the best of two worlds is taken: the capillary is coiled but at the same time it is assumed to be mathematically straight. Under what circumstances are we allowed to do this? Taylor dispersion is usually carried out in a coiled capillary, the coil diameter being of the order of several decimetres, about 4 dm say. Although the measured diffusivities are probably accurate enough, literature does not provide a firm theoretical basis for neglecting the curvature of the capillary. At least, it is impossible to satisfy all criteria for a Taylor dispersion capillary for all possible system parameters without falling out of the range of reasonable dimensions and experimental conditions. These criteria are: 1. The flow within the capillary should be laminar, so Re = 4 V < 2000 d (6)

2. The volume of the injected sample Vinj must be very small compared with the capillary volume Vcap [13]. This gives the sample enough volume to disperse in, and allows us to ignore the initial shape of the pulse. Vcap Vinj > 100 (7)

3. This criterion originates from the theory of Taylor dispersion [1, 2, 15, 16]. The cross-section averaged second moment of the concentration distribution in the capillary, m2 , is given by

Rutten [13] used a capillary with an inner diameter of 0.53 mm and a coil diameter of 0.40 m, for Snijder's [14] capillary these measures were 0.56 mm and 0.1 m, and v.d. Ven [21] uses a capillary of similar dimensions.

27

Chapter 3

m2 = 2 D +

F GH

2 usup d2

I t 8 u d 192 D J K D
2 sup 2

i =1

8 i

LM1 expF 4 D t I OP MN GH d JK PQ
2 i 2

(8)

In this expression, i represents the ith zero of the derivative of J0, the zeroth-order Bessel function of the first kind. Although not strictly necessary, it is convenient to simplify the first term on the right-hand side: in most cases the diffusivity D will be much smaller than usup2d 2/ 192 D, and when it is negligibly small, it can be omitted. However, under what conditions do we say that one term is negligibly small compared to another term? Let us say that a positive number a is negligible compared to another positive number b if a < b, being some small number between 0 and 1, 0.01 for instance. With the definition of the radial Pclet number Pe (= usup d / D) and with usup = 4 V / d 2, we find D+
2 usup d2

192 D

2 usup d2

192 D

if

Pe 2 > 1 192

(9)

4. This condition also follows from equation (8). Note that the summation function is quite impractical for curve-fitting, and it would be convenient if it could be eliminated. First the exponential terms get in our way, but it turns out that all of them are usually much smaller than 1. Of all i, 1 has the smallest value: 1 3.83, 2 7.01 etc. Hence, we only have to consider whether the term containing this first root is negligible, because if it is, so are all the other terms. This means according to our negligibility definition: ln Dt > 2 d2 4 1 Now, equation (8) reduces to m2 =
2 usup d2 t

(10)

96 D

2 usup d4

D2

(11)

where A represents the sum i8, which has a value of 2.1701...105. The second term of this expression can be ignored if
2 usup d2 t

96 D

>

2 8 usup d4

D2

A or

D t 768 A 1.67 10 2 > d2

(12)

Both conditions (10) and (12) concern the same quantity D t /d 2, called the Fourier number Fo. It turns out that whenever (12) is satisfied, so is (10), which is therefore a redundant criterion. What is the meaning of t in eq. (12)? Strictly speaking, t is the time after injection of the sample. A representative value of t is the mean residence time of the sample: = Ld2 4 V (13)

28

TAYLOR DISPERSION THEORY If this expression is inserted in (12) for t, then this equivalent to the statement that this criterion holds at t = , and it will be assumed that it still holds when the last traces of the sample leave the capillary. The resulting inequality is D L 3072 A > V (14)

5. This condition is a rather prosaic one, for it is imposed by the limitations of the HPLC-pump used for pumping the fluid through the capillary. The maximum pressure drop Pmax such pumps usually can generate is about 400 bar, and according to Poiseuille's law, the parameters must be chosen to satisfy the inequality P= 128 V L < Pmax d4 (15)

6. The criteria discussed so far concern straight capillaries, and, apart from numbers 3 and 4, are not likely to cause too many problems. The next criterion should take care that coiling does not lead to significant deviations of the measured diffusivities. Intuitively, one might argue that coiling the capillary leads to a secondary flow perpendicular to the tube axis. The extra radial mixing diminishes radial concentration gradients. Radial transfer in straight capillaries is a diffusive process only, and therefore we would expect convection due to coiling to increase the radial transport. However, it will be seen that coiling also influences the axial profile, and can even reduce the dispersion coefficient. Approximate solutions of the equations describing simultaneous convection and diffusion in curved pipes have been found by Erdogan and Chatwin [8], and Nunge, Lin and Gill [10]. In both papers power series expansions in the coil-capillary diameter ratio are used. In the first paper the authors find the following expression for the dispersion coefficient , which is based on the velocity profiles derived by Dean [6, 7] EC = Pe 2 D a1 + a 2 2 Re 4 a 3 Sc 2 + a 4

(16)

Here the constants a1 to a4 have the following values a1 = 1 1 2569 1109 ; a2 = ; a3 = ; a4 = 192 15840 43200 576 2 40 (17)

In the straight capillary case, that is, as , EC D Pe2 a1, which is the approximation found by Taylor [15]. This indicates that the approximation is not accurate, since the limit should yield the Aris solution for straight tubes. In the article of Nunge, Lin and Gill, equation (16) was extended to NLG = Pe 2 D 20 + 2 22 with

(18)

29

Chapter 3

20 =

1 + a1 Per2
4

22 = a 2 Re a 3 Sc + a 4 + a5 Re a 6 Sc a 7 + a8 + a 9 20
2 2

(19)

The constants a1,..., a4 have the same values as before, and a5 = 1 31 8499 419 1 ; a6 = ; a7 = ; a8 = ; a9 = 41472 60 4480 11520 4 NLG A 2 Pe 2 22 = A Pe 2 a1 + 1 (20)

Figure 1 shows a plot of the deviation of NLG relative to the Aris solution, NLG = (21)

We see that The straight-tube limit of equation (18) is the Aris solution, as expected. For high Re-numbers both dispersion coefficients EC and NLG tend to the same limit. For normal liquids (Sc 103), an increase of the dispersion coefficient due to coiling is found only at very low Re-numbers. For higher Sc-values very large deviations may be found, even at moderate Re-numbers. To keep the relative error smaller than 1% may require an extremely high . Although both equations predict that the dispersion coefficient increases only for Re-, Sc- and values within definite ranges, there are notable differences, and equation (18) is probably the better approximation. The criterion for coiling then becomes 2 20 1 22
10 2 NLG (-)

(22)

Sc (-) 10 100 1000

-10 0.0

2.5

5.0 Re (-)

7.5

10.0

Figure 1. The deviation of the dispersion coefficient according to Nunge, Lin and Gill for several Scnumbers.

30

TAYLOR DISPERSION THEORY 7. A last criterion is that the sampling time must be long enough for accurate measurements. We will choose the minimum sampling time (tm) as 120 s. After having injected a Dirac-pulse in the fluid stream at the beginning of our capillary at t = 0, we stroll over to the other end of it and wait for the pulse to come out. So, our position is fixed (z = L), and the remaining variable is the time t. The response of the system is given by equation (1), which is a Gauss function with respect to t. The standard deviation of this function equals (2 t). Since the peak width interval is chosen arbitrarily, there is no need for excessively scrutinous consideration of which should be used, and here it is taken to be the Aris approximation. The peak width of a response curve is usually defined as 4 [3], and the seventh criterion then becomes tm = 4 2 t usup > 120 (23)

As stated above, the criteria 1, 2, and 5, which concern straight tubes and are not related to , are usually easily satisfied. The additional peak width criterion may not be met in all practical cases, but this is a problem we can live with. The most severe troubles can be caused by the criteria 3, 4 and 6. For a given configuration these criteria can be written in the form > f (V). In figure 2 these three functions are plotted for a capillary with dimensions L = 20 m, d = 0.53 mm, and = 0.80 m. The relevant physical fluid parameters were taken as D = 510-9 m2 s1, = 1.5103 kg m3 and = 2104 Pa s. It is clear that the highest accuracy is obtained at a flow rate of about 1.5109 m3 s1.
1.0 10 (-) 0.8 criterion # 0.6 0.4 0.2 0.0 0.0 3 4 6
3

best accuracy

0.5
9

1.0 10 V (m s )
3 -1

1.5

2.0

Figure 2. The accuracy derived from the criteria 3, 4 and 6 as a function of the volume flow. The
equipment parameters are: L = 20 m, d = 0.53 mm, and = 0.80 m, and the physical fluid properties are D = 5109 m2 s1, = 103 kg m3 and = 2104 Pa s.

In table 1 the best obtainable accuracy of the Taylor dispersion method is listed for several sets of physical properties of the fluid. The best accuracy is obtained at a specific flow rate, which may, however, be too low for practical applications. Therefore, the accuracy is also computed for a more convenient flow rate of 2.5109 m3 s1 (= 0.15 ml min1). 31

Chapter 3 Table 1. The best possible accuracy with the Taylor dispersion method for several sets of system

parameters, the corresponding flow rate, and the criteria by which this -minimum is determined. Capillary dimensions: L = 20 m, d = 0.53 mm, and = 0.80 m.

physical fluid properties 103 (kg m3) 1.5 0.6 1.5 0.6 1.5 0.6 1.5 0.6 103 (Pa s) 200 200 0.2 0.2 200 200 0.2 0.2 109 D (m2 s1) 5 5 5 5 0.5 0.5 0.5 0.5 104 () 3.3 3.3 3.7 3.3 3.3 3.3 3.3 3.3

best accuracy 109 V (m3s1) 1.6 1.6 1.5 1.6 0.16 0.16 0.16 0.16 criteria 3, 4 3, 4 3, 6 3, 4 3, 4 3, 4 3, 4 3, 4

V = 2.5109 m3 s1 104 () 5.3 5.3 30 5.3 53 53 3.0103 4.8102 criterion 4 4 6 4 4 4 6 6

We see that, although the minimum accuracy is quite good in all cases, it is not always obtained at flow rates which allow quick measurements: is about 8 hours at a flow rate of 0.16109 m3 s1. But it appears that the accuracy deteriorates dramatically in some cases if we choose a more convenient flow. Thus, for mixtures with extreme properties, it is necessary to have a look at the criteria before any measurements are carried out. Extreme conditions as those of the last two rows of table 1 are not likely to occur, because low viscosity is usually accompanied by high diffusivities, and only in the case of mixtures of large entities such as macromolecules and low-viscous fluids, are problems to be expected. With the above set of rules, it is no longer necessary to work with rules-of-thumb, as most experimenters seem to do. Rutten [13] does not explicitly derive his coiling criteria ( > 100, Re < 50), although he claims them to be based on work done by Tijssen [17, 18, 20], Nunge et al. [10], and Erdogan and Chatwin [8]. Snijder [14] uses the results of Janssen [9] to justify his choice of capillary dimensions. Janssen found on the basis of numerical simulations that for values of De2Sc < 100 there will be no significant difference with the straight tube, whenever > 20. What is meant exactly with significant is not completely clear, but it probably refers to relative errors of up to about 3%, which is much more than the minimum allowable deviation of 0.1% used here. If the more restrictive condition De2Sc < 50 is used, the relative deviation shrinks to less than 0.5 %, which is in better accordance with the criterion given by Alizadeh et al. [1]: De2Sc < 20 and > 100. Evidence supporting the notion that coiling effects cannot always be ignored was found by Chen and Blanchard [4, 5], who measured diffusivities of macromolecular substances (PS) in an organic solvent (THF), and who found significant deviations from Taylor dispersion theory. Tijssen [19] showed that these deviations can probably be ascribed to the small curvature of the tube used in the experiments.

32

TAYLOR DISPERSION THEORY 3.2 RETRIEVAL OF DIFFUSIVITIES FROM EXPERIMENTAL DATA We now know how to conduct an experiment, and what to expect from the data it produces. The final step is to extract diffusivities from the data. To this aim, eqs. (1) or (4) are fitted to the experimental data. The following section treats the problems connected with this retrieval of the diffusivities from the experimental data. The extraction of the diffusivities from set of measured composition profiles is not always as easily done as it is said. Problems can arise due to unruliness of the model functions as well as of the detection method. First, the functions. We wish to fit the functions (1) and (4) to signals measured at the end of the capillary, that is, at z = L. We will use the functions in the mole fraction form: the mole fraction is simpler to determine than the concentration, especially at the moment that the solution is prepared. The equations then become y1 t = or

bg

b t

exp k

F b1 t g I GH t JK
2 2

with b =

2 n1E c 3 2 d 2

and

k=

c Lh
1 2

(24)

F b1 t g I GH t JK + t F b1 t g I + b y bt g = exp G k t H t JK
y1 t =

bg

b1

exp k

F b1 t g I GH t JK t F b1 t g I b exp G k t H t JK
b2
2

exp k

(25)

2 K1 2 K2 2 K3 2 K4 , b2 = , b3 = , b4 = b1 = 32 2 32 2 32 2 c 3 2 d 2 c d c d c d It will be shown below that there is little wrong with the first equation from a fitting point of view, but the similarity of the two component profiles in the second one will cause trouble. The righthand members of these eqs. require the total concentration c, but because n E / c = y E Vinj everything can be defined in terms of mole fractions. Vinj is a quantity that must be determined anyway for the evaluation of the injected excess amounts of the independent components. The detection of the concentration-time functions may give problems as well. A good measuring device is linear, and does not deform or delay the output signal. That is, if f ( t ) is the output of the apparatus, we have f t = S y t + s+r t

bg

bg

bg

(26)

Ideally, the matrix [S ] of the detector is independent of y, and in even more ideal equipment, it is also diagonal. In this section we will assume [S ] is constant for the composition ranges considered. The constant s is the offset of the detector, and it would be nice if it equals o but we shall see below that it is no problem if it is not. The function r (t) is the (white) noise function of the measuring apparatus. The number of independent outputs, the dimension of f, is bound to a maximum of n 1, for we cannot obtain more information from the n 1 independent mole fractions yi. 33

Chapter 3 Measuring equipment can exhibit baseline drift, and eq. (26) should also include a drift function. If the drift is not too serious, it does not render an experiment useless. Any moderate, well-behaved function can be accurately described with a polynomial, and so, assuming the drift function is a polynomial, we get f t = S y t + s + r t + a j t j
j=0

bg

bg

bg

(27)

We can, of course, try to fit the baseline drift simultaneously with the other parameters. Because of the increased number of parameters, this slows down the fitting process, but above all it may shrink the domain of convergence. This is the set of all combinations of initial parameter estimates for which the fitting routine converges to a solution. Luckily, there is a simple and often effective alternative, namely manual pre-fit baseline correction. Given the functions (24) and (25), we can see that y (t) 0 as | t | , so f t s + r t + a j t j
j=0

bg

bg

as

(28)

We now make a selection of data points at both sides of the peak that clearly belong to the baseline of the curve. They represent measurements of the eluent composition. To these selected data we fit a polynomial of a certain degree. (A degree of three will do perfectly in most cases.) This polynomial is then subtracted from all points in the data set. For moderate forms of baseline drift, this method gives us a neatly zeroed baseline. Apart from this, it also frees us of the constants s (offset) and [S ] ye (the eluent signal). These are incorporated in the constants a0, but this also happens when the baseline drift is fitted along with the signal, and that is why the offset of the detector is not important. The resulting signal f ( t ), which can be used to fit the equations (24) and (25), is f t = S y t +r t

bg

bg bg

(29)

This baseline correction method has one possible disadvantage. The drift function is only fitted to a limited selection of the data set. Any irregularities in the baseline in the part of the signal that does not belong to this selection are not seen and remain present in the corrected data set. This is why it would be better to fit the baseline along with the other parameters, had it not been for the drawbacks mentioned above. Further, manual baseline correction requires that the data contain enough baseline points, otherwise it is not possible to determine the correction polynomial with sufficient accuracy. Do we need to know the linearity matrix [S ] to be able to fit eqs. (24) and (25)? We do not, although it may make things easier if we do. In the binary case, substitution of eq. (24) into (26) yields

The Marquardt-Levenberg method [11] was used for all parameter optimisations. As was mentioned at the beginning of this chapter, c = c ce. Since c is assumed constant, this is equivalent to y = y ye. Hence: [S ] y(t) = [S ] y(t) + [S ] ye. A good fit always requires a well-defined baseline, so normally this constraint should be fulfilled anyway.

34

TAYLOR DISPERSION THEORY

f t =

bg

S1,1 b t

exp k

F b1 t g I GH t JK
2

(30)

So, in principle S1,1 b, D and can be determined from a single output data set. In this case practice follows theory, and usually, the lumped parameter S1,1 b and the other two parameters can be determined from just one set of data with reasonable accuracy. Only if S1,1 is known, it is possible to evaluate n1E, but the injected excess amount of a substance is a less relevant parameter anyway. Besides, it can be determined beforehand while preparing the samples. If one really wants to fit n1E, two experiments at different sample compositions are required, and eq. (24) must be fitted simultaneously to the two resulting data sets. In the ternary case things are more complicated, because of the similarity of y1 ( t ) and y2 ( t ). With three components, f can have two independent elements at most, but many measurement devices have only one output (e.g. refractive index at one wave length, or conductivity). The constants S2,1 and S2,2 then are zero, and the output becomes f1 t =

bg

S1,1 b1 + S1,2 b3 1 t

F b1 t g I + S exp G k H t JK
2 1

1,1

b2 + S1,2 b4 2 t

F b1 t g I exp G k H t JK
2 2

(31)

The fit of a single data set gives values for (S1,1 b1 + S1,2 b3), (S1,1 b2 + S1,2 b4), 1, 2 and . Only under certain conditions is one data set sufficient to fit all parameters. In most cases however, we will need more than one set of experimental data. How many more? This depends on 1. whether we know the excess number of moles n E for each experiment, 2. whether the linearity [S ] is known 3. the number of outputs of the detector. Suppose that we have a number of m experiments at constant eluent composition. The parameters [D ] and [S ] are the same for all m data sets, while may differ, and n E certainly will. The diffusivities are unknown by definition, whereas the mean residence time is a fit parameter for the sake of accuracy. Table 2 contains an inventory of the number of unknowns for all possible situations that can be encountered. Table 2. The number of unknowns and equations for all possible sets of experimental conditions. The
numbers in parentheses are the numbers of experiments needed for the determination of the parameters. The numbers in this table are valid for ternary dispersion only.

detector calibrated Yes Yes No No


*

# outputs 1 2 1 2
E

# unknowns n unknown 4 + 3 m () 4 + 3 m (1) 6 + 3 m () 8 + 3 m (3) n known 4 + m (1) 4 + m ()* 6 + m (2) 8 + m (1)*


E

# equations 2+3m 2+5m 2+3m 2+5m

In the case of two outputs, half an experiment corresponds to one peak.

Note that a varying implies that either L is not correct or usup deviates. Since L can be measured accurately, it must be usup that potentially contains a significant error. Therefore, we should use usup = L / in eqs. (3) and (4), which makes a function of both D and .

35

Chapter 3 The minimum number of unknowns is 4 + m: the four Ds and m s. For an uncalibrated detector this number must be increased by the number of unknown linearity factors, namely 2 (single output) or 4 (dual output). The n E make for 2 m extra variables if they are not known. The simultaneous fit of the data set from a single-output detector yields a number of 2 + 3 m equations: namely values for 1 and 2, which are the same for all sets, and 3 m values for and the composite parameters (S1,1 b1 + S1,2 b3) and (S1,1 b2 + S1,2 b4). Dual-output detectors supply us with two peaks per experiment and therefore we have 5 m + 2 equations. The table shows that with a single-output measurement apparatus it is impossible to find the values of all parameters unless we know n E. This is because every data set introduces as many new unknowns as extra equations. With a dual-output device it is possible to fit all possible parameters, which is why two outputs are to be preferred to one output. Thus, one peak suffices to determine all 5 parameters with such a measuring device if the n E are known and if it is calibrated. If it is not calibrated, we need three peaks. This may seem a bit strange, but this situation is not essentially different from the single-output case with known n E, and we need two peaks from, say, the first output to determine all parameters except S2,1 and S2,2. For the determination of these two parameters, we need a third data set from the second output. If, as has been said above, it is easy to determine the n E, then why would we like to fit them? This is because possible experimental errors in n E or in [S ] if we choose to calibrate our detector can then move into the fitted parameters, of which the Ds are the ones that really matter. This is also the reason why the mean residence time of a sample is taken as a fit parameter, even though it can be determined experimentally with good accuracy. The fact that this introduces extra parameters does not weigh heavily, since usually can be fitted very accurately, because it is almost independent of other parameters. Moreover, it is possible to make good initial guesses, either from the flow rate and the capillary volume, or from the output data. It must be realised that the numbers of experiments listed in table 2 are an absolute minimum. More data sets usually give better fits, and tend to increase the stability and enhance the domain of convergence of the fitting routine. As an example consider the case of a calibrated detector with one output. Suppose that we know n E, then according to the table only one experiment should do. To verify this, we use constructed data sets instead of measured ones so as to be sure about what the exact values of all parameters are. Let L = 19.0 m, d = 5.3104 m, usup = 1.0102 m s1, c = 1.33104 mol m3, D1,1 = 3.0109, D1,2 = 5.01010, D2,1 = 1.3109, D2,2 = 5.0109 m2 s1, S1,1 = 2.0, S1,2 = 1.0 V and s = o V. Artificial peaks were constructed with eq. (31) for 250 points on the time interval [1.6, 2.2] ks (i.e., every 2.4 s). Noise was added according to

r t = r cos 2 x t

bg

c b gh

2 ln x t

bg

(32)

which is said to give normally distributed random numbers with a variance of r2 if x(t) is a uniformly distributed random number on [0, 1. The responses to the following samples were calculated

is located approximately at the (middle) extremum of the signal, but not exactly: if t0 is the position of 2 this extremum, then to a good approximation we have (for binary dispersion): t0 / u sup.

36

TAYLOR DISPERSION THEORY I. (n E)T = (5.0 9.0)107 mol, and II. (n E)T = (8.0 2.0)107 mol. The noiseless composition-time distributions of these two samples in the absence of baseline drift are shown in the figure 3.
6e-5 f (V) II 6e-4

0e+0 I

0e+0

-6e-5 1.6

1.8 t (ks)

2.0

-6e-4 2.2

Figure 3. The simulated ideal responses of the two samples. Notice that the response of pulse II has a
maximum amplitude about 13 times larger than that of pulse I.

Although these distributions contain the same information about [D ], they are different when it comes to parameter fitting. This is best illustrated by the actual fit results. Table 3. One-peak fit results for both peaks I and II for varying r. Calibrated detector, sample
compositions known. The diffusivities Dij are given in 109 m2 s1, the estimated standard errors (ESEs) Dij are given as a percentage of the fitted value of Dij . The s, which are not shown, fitted very well under all circumstances not only with respect to their magnitude, but also to their ESE , a fact that has already been mentioned above. The enormous ESEs are no mistakes!

Sample # r (V) I 10 106 105 1010 106 105


10

D1,1 D1,1 2.91 4% 2.83 6104% 2.62 3105% 2.78 2101% 2.77 3104% 2.77 1106%

D1,2 D1,2 0.494 3% 0.593 2104% 1.45 2105% 0.153 7102% 0.208 1107% 0.485 3107%

D2,1 D2,1 0.932 610 % 0.835 8105% 1.39 2106% 1.09 3101% 1.28 3104% 7.80 4104%
1

D2,2 D2,2 5.02 2% 5.26 3104% 6.50 1105% 5.22 9% 6.15 1104% 31.5 1105%

II

The first thing that meets the eye, is the inaccuracy of the fits, but we will come to discuss that in the next paragraph. What is of importance here, is that peak I is less sensitive to noise, that is, it retains the information about the diffusivities better than peak II if the noise gets worse. This is especially so if it is realised that for a given r, the signal-to-noise ratio is much smaller for peak I than for peak II. Not only do the fits of peak I give more accurate values of the Dij, but also are the 37

Chapter 3 estimated standard errors (ESEs) smaller on average. Moreover, it appears that the binary dispersion function (30) can also be fitted well to peak II, which indicates that the information on the two superpositioned peaks is weak. Therefore, it is worthwhile to choose the sample compositions such, that the recorder signal has a peak I appearance, but this may not always be possible. Of the diffusivities listed in the table, only those at r-values of 1010 V are usable. For the higher noise levels, the ESEs are so big that the values of the diffusivities are statistically insignificant. That the fitted values of the Dij are not too far off, is a result of the ideal peaks: apart from the (almost) perfectly white noise there are no disturbing influences. For the example peaks, the maximum signal-to-noise ratios at a r of 1010 V are as high as 4105 (peak I) and 5106 (peak II). Since such levels can not be expected in practice, the conclusion must be that one peak simply is not enough to determine the diffusivities with low errors. Fits to two peaks simultaneously, on the other hand, give perfect results with low ESEs. This is illustrated in table 4. If the peaks are good, a fit of more than two peaks gives only slight improvement. Table 4. Two-peak fit results for both peaks I and II for varying r. Calibrated detector, sample
compositions known.

r (V) 10 106 105


10

D1,1 D1,1 3.00 510 % 3.00 4101% 2.96 5%


5

D1,2 D1,2 0.500 410 % 0.520 4% 0.700 3101%


4

D2,1 D2,1 1.30 310 % 1.29 3% 1.17 3101%


4

D2,2 D2,2 5.00 8105% 5.05 8101% 5.45 8%

But there is more! From table 2, we see that we need two data sets to fit the diffusivities for an uncalibrated detector. With the result for the calibrated detector in mind, we might expect that we may need more than two sets, but this is not so. The linearity [S ] can be fitted together with [D ] and the two s without apparent loss of accuracy of the fitted value of [D ]. (In fact, it slightly improves them.) Thus we kill two birds with one stone: we avoid the work of calibration, and we keep errors in the experimentally determined values of [S ] out of our other parameters. 3.3 CONCLUSIONS The first conclusion that can be drawn from the foregoing is that for normal substances, no special precautions need to be taken for diffusivity measurements with the Taylor dispersion method. For normal capillary dimensions, the coil diameter can be relatively small, and flows can be chosen that result in acceptable residence times of the pulses. However, in extreme cases such as macromolecular substances and unusual combinations of density, viscosity and diffusivity, it may be necessary to decrease the flow rate through the capillary to obtain good accuracy. The second conclusion is that it is possible to measure multicomponent diffusivities with a detector which can produce only one independent signal. As more is known about the samples and the detector linearity, fewer measurements are needed to produce reliable results.

38

TAYLOR DISPERSION THEORY NOTATION a1,..., a9 A b, b1,..., b4 c D d De disc[D] f J0 K1,..., K4 L constants in coiled-capillary dispersion model, () constant, i8 2.1701105, () fitting parameters, (m) concentration, (mol m3) Fick diffusivity, (m2 s1) inner capillary diameter, (m) Dean number, Re / , () discriminant of [D], tr 2 [D] 4 | D |, (m4 s2) detector signal, (V) zeroth order Bessel function of the first kind constants in Prices formulas eqs. (4) and (5), (mol) length of the capillary, (m) second moment of the cross-section averaged concentration profile, (m2) number of moles (mol) pressure, (Pa) radial Pclet number, usup d / D = Re Sc, () noise function, (V) Reynolds number, usup d / , () Schmidt number, / D, () detector linearity, (V) detector offset, (V) trace of [D], i Dii, (m2 s1) time, (s) velocity, (m s1) volume, (m3) axial capillary co-ordinate, (m) ith root of the derivative of J0 difference, deviation coil diameter, (m) accuracy, () fluid dynamic viscosity, (Pa s) dispersion coefficient, (m2 s1) coil / capillary diameter ratio, / d, () fluid mass density, (kg m3) standard deviation of the response peak, (m) standard deviation of the normally distributed noise function, (V) mean residence time, L / usup, (s) volumetric flow rate, (m3 s1)

m2
n P Pe r Re Sc S s tr [D] t u V z Greek symbols i r

39

Chapter 3

Subscripts A cap EC i inj m max NLG sup V Superscripts E T REFERENCES [1] [2] [3] [4] [5] [6] [7] [8]

Aris capillary Erdogan & Chatwin several counters injected measurement maximum Nunge, Lin & Gill superficial volume excess transpose

[9] [10]

Alizadeh A., Nieto de Castro C.A., Wakeham W.A. The theory of the Taylor dispersion technique for liquid diffusivity measurements. Int. J. Thermophys., 1(3), 1980, pp 243284. Aris R. On the dispersion of a solute flowing through a tube. Proc. Roy. Soc. A, 235, 1956, pp 6977. Braam W.G.M. Scheidingsmethoden: chromatografie. 1985, WoltersNoordhoff, Groningen, NL. Chen H.R., Blanchard L.P. Can. J. Chem., 53, 1975, p 228. Chen H.R., Blanchard L.P. Can. J. Chem., 53, 1975, p 476. Dean W.R. Note on the motion of fluid in a curved pipe. Phil. Mag., 4, 1927, pp 208223. Dean W.R. The stream-line motion of a fluid in a curved pipe. Phil. Mag., 5, 1928, pp 673695. Erdogan M.E., Chatwin P.C. The effects of curvature and buoyancy on the laminar dispersion of solute in a horizontal tube. J. Fluid Mech., 29, pp 465484. Janssen L.A.M. Axial dispersion in laminar flow through coiled tubes Chem. Eng. Sci., 31, 1976, pp 215218. Nunge R.J., Lin T.-S., Gill W.N. Laminar dispersion in curved tubes and channels. J. Fluid Mech., 51, 1972, pp 363383.

40

TAYLOR DISPERSION THEORY [11] [12] Press W.H. et. al. Numerical Recipes. 1986, Cambridge University Press, Cambridge, UK. Price W.E. Theory of the Taylor dispersion technique for three-component-system diffusion measurements. J. Chem. Soc. Faraday Trans. 1, 84(7), 1988, pp 24312439. Rutten Ph.W.M. Diffusion in liquids. 1992, Delft University Press, Delft, NL. Snijder E.D. Metal hydrides as catalysts and hydrogen suppliers. 1992, Enschede. Taylor G.I. Dispersion of soluble matter in solvent flowing slowly through a tube. Proc. Roy. Soc. A, 219, 1953, pp 186203. Taylor G.I. Conditions under which dispersion of a solute in a stream of solvent can be used to measure molecular diffusion. Proc. Roy. Soc. A, 225, 1954, pp 473477. Tijssen R. Effect of column-coiling on the dispersion of solutes in gas chromatography. Part 1: theory Chromatographia, 3, 1970, pp 525531. Tijssen R. Effect of column-coiling on the dispersion of solutes in gas chromatography. Part 2: generalized theory Chromatographia, 5, 1972, pp 286295. Tijssen R. On the dispersion of macromolecules in dilute solutions measured with the bandbroadening technique. Can. J. Chem. Eng., 55, 1977, pp 225226. Tijssen R. Axial dispersion and flow phenomena in helically coiled tubular reactors for flow analysis and chromotography Anal. Chim. Acta, 114, 1980, pp 7189. Ven-Lucassen I.M.J.J. van der, Kieviet F.G., Kerkhof P.J.A.M. Fast and convenient implementation of the Taylor dispersion method. J. Chem. Eng. Data, 40, 1995, pp 407411.

[13] [14] [15] [16]

[17]

[18]

[19]

[20]

[21]

41

You might also like