Chemical Engineering Science 57 (2002) 4053 4065

www.elsevier.com/locate/ces

Experimental study of the mixing kinetics of binary pharmaceutical powder mixtures in a laboratory hoop mixer
Sandra Massol-Chaudeura , Henri Berthiauxb; , John A. Doddsb
a LAGEP-UCB b Ecole

Lyon I, UMR CNRS 5007, ESCPE B t 308, Campus La Doua, 69 622 Villeurbanne, France a des Mines dAlbi, Laboratoire de G nie des Proc d s des Solides Divis s-UMR CNRS 2392, e e e e Campus Jarlard-Route de Teillet, 81013 Albi France

Received 21 February 2002; received in revised form 17 April 2002; accepted 15 June 2002

Abstract As increasingly commented by the literature during the last 5 years, estimating the homogeneity of a powder mixture and following powder mixing processes is not a simple task. In this paper, we present the development and statistical validation of a sampling methodology for dening the number of samples required to provide a reasonable estimation of the homogeneity attained in a laboratory scale tumbler mixer. This method is then used to follow the mixing kinetics of a dilute binary powder mixture in a hoop mixer. Special attention is paid to the statistical meaning of the values obtained and the in uence of the physical characteristics such as particle size and shape. The role of the particle shape of the majority powder is particularly emphasised and it is quantitatively demonstrated that spherical particles are harder to mix and more ready to segregate than particles with irregular shapes. The di erent mixing mechanisms at play are identied; the practical limits of use of such tumbler mixers with pharmaceutical powders are discussed. ? 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Powder mixing; Homogeneity; Mixing kinetics; Hoop mixer; Segregation; Pharmaceutical mixtures

1. Introduction Powder mixing is an important unit operation in the wide range of industries involved in solids processing. The end use properties of an increasing diversity of products of the pharmaceutical, food, plastics, ne chemicals industries, mainly depend on the mixing action of dry mixers and=or contacting machines (e.g. for encapsulation, agglomeration, etc.). In the context of the pharmaceutical industry, these properties are usually determined through a formulation procedure involving costly clinical trials to determine the right composition and the right dosage form of the drug. But while a constant e ort is devoted to this aspect, fewand sometime very fewis known about the manufacture process itself. Undoubtedly, this makes the study of mixing and mixtures, including multicomponent ones, a key subject for both academic and industrial product and process engineers. In the last decade, a huge amount of research e ort has been devoted to developing measuring techniques and models in the eld of particle technology. The increase of
Corresponding

author. Tel.: +33-5-63-49-30-00. E-mail address: berthiaux@enstimac.fr (H. Berthiaux).

computer power has led, on one hand to the development of phenomenological models based on particle dynamics simulations, and on the other hand to systemic models based on population balances which are now currently used at appropriate scales (Kaneko, Shiojima, & Horio, 2000; Berthiaux & Dodds, 1999). Technical methods for measuring powder and blend characteristics have also undergone development (Dyakowski, Jeanmeure, & Jaworski, 2000; Lai et al., 2001), some of these being used as on-line process measurements: Electrical capacitance tomography (ECT), optical bre systems, image acquisition and analysis, particle size analysis, : : : . With the continuous improvement of these tools, and as recently pointed out by Scarlett (2001), it can now be said that the road is open for the development of coherent chemical engineering research in general particle technology, and in powder mixing in particular. However, despite this progress, there is still a divide between the increasing research activity and the e ective use of models, techniques with an engineering approach at the production process level. This is again the case in the pharmaceutical industry in which quality insurance is a constant topic, and for which normal practice is still to use batch mixers, even if the previous and the succeeding steps in an

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 2 6 2 - 2

4054

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

operation are continuous, thus generating costs with intermediate storage and increased possibilities of powder segregation. Undoubtedly, the double key factor to control is still the quality of the mixtures through the quality of its own estimation. This point remains a technical and statistical problem, and even in the best cases this is usually based on about 10 30 thief-probe samples taken from a batch process of at least a million possible samples. At the level of regulation, it is commonly asked to reach a certain specication with a limited number of samples in a reproducible way, rather than trying to evaluate rigorously the homogeneity of the whole batch. As a consequence, the major part of the drugs present on the market are produced through additional granulation and encapsulation steps to link the active substance with the excipient, rather than through direct blendinglubricationtabletting steps of lower costs but higher risks. Operational batch mixing time is, in addition, usually xed by empirical procedures that are mainly the fact of the companies themselves. As an example (see the recent paper by Muzzio, Shinbrot, & Glasser, 2002), a typical validation protocol will consist in performing three production-scale batches with a denite sampling procedure and to judge the capability of the process on the basis of its own reproducibility. From the point of view of the products, it is also worth noting that excipients, and mixtures of excipients, in powder form are often used with little investigation of the e ects of their particle size or morphology on mixture homogeneity and stability (but also on dissolution and biodisponibility). In this paper, we present and discuss results of laboratory scale experiments on the kinetics of mixing of typical pharmaceutical powders in a hoop mixer. Attention is given to some of the typical industrial practices mentioned above, examined in the light of the rigor necessary for the correct evaluation of mixture homogeneity (in particular the number of samples to be examined), its in uence on the determination of the optimum mixing time, and the in uence of particle size and shape of a widely used excipient. Due to the particulate systems under consideration, only random mixing is examined and not ordered mixing. 2. State of the art 2.1. Homogeneity of powder mixtures Consider a powder blend, a binary mixture to simplify, consisting of N possible samples, the size of which is equal to the scale of scrutiny (in a pharmaceutical context this would be the weight of a tablet). If these N samples are all exactly taken and analysed, the standard deviation of the distribution curve of the composition in terms of one key component is in-

dicative of the homogeneity reached. In e ect, a wide distribution curve with a high value will indicate that samples are of very di erent compositions, whilst a narrow curve will show that these are very close in composition. If xi is the fraction of key component in the samples, and if is the observed mean, a macroscopic criteria (Lacey, 1954) for homogeneity is then: : (1) N It is common to use mixing indices which link this value of with those dened for limiting cases of homogeneity, such as the perfect random state R or the totally segregated state 0 . However can be considered an absolute reference for a given powder system. It must be noted that other criteria can also be of interest, such as the autocorrelation function, which takes into account the structure of the mixture (Schoeld, 1970). As mentioned above, more and more interest is being shown in the development of on-line non-invasive techniques for measuring homogeneity at a very ne scale, thus trying to reach an exhaustive sampling procedure at any scale of scrutiny. However, a generalisation of such techniques in the context of industrial plants is still a long way o , and in current practice only an estimation of the real homogeneity can be provided by withdrawing a limited number of samples from the bulk and calculating an estimated standard deviation s. Moreover, the technique most usually employed still involves sampling by inserting thief probes in denite locations, resulting in a certain lack of rigor, as pointed out in a legal decision in the USA (US versus Barr, 1993), and commented later in the scientic community (Berman & Planchard, 1995; Berman, Schoeneman, & Shelton, 1996). Recently, Muzzio and co-authors (Muzzio, Robinson, Wightman, & Brone, 1997) published a comparative study of the e ect of employing di erent probes on the quality of the evaluation of mixture homogeneity. Their conclusion was that practically nothing could be said about the real homogeneity evaluated using these techniques. If one refers to the golden rules (Lantz & Schwartz, 1981) of sampling (sample while the powder ows, sample over the whole width of the ow), it is clear that thief probes violate all these laws. But the problem is not only at the level of the technique used for sampling since a good evaluation of the quality of a mixture must be based on an adequate number of samples of the right size. This is all the more true the greater the heterogeneity of the mixture. In limiting cases, a totally segregated blend will need exhaustive sampling while a perfect mixture can be qualied by analysis of only a single sample. Another consequence of this is that research work in mixing and industrial process control, if not quality control, must be based on sampling procedures, which have been validated. =
N i=1 (xi

)2

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065 Table 1 Typical classication of powder mixers

4055

Mixers Tumbler mixers Convective mixers Fluidized mixers Static mixers

Mechanisms Di usive and shear mixing Convective and shear mixing Di usive and convective mixing Di usive and convective mixing

Examples Cube, double-cone, V mixers, : : : Ribbon mixer, paddle mixer, : : : AirmixJ SulzerJ, silo mixers, : : :

2.2. Powder mixers and mixing kinetics Powder mixers (see Table 1) are relatively simple machines of low capital cost, which are classied according to the mixing mechanisms at play (Harnby, 1992, Chap. 3). In the pharmaceutical industry, which is the domain of reference for this paper, the most widely used types are batch tumbler mixers (together with some specic convective mixers). The main reasons for this are the simplicity to clean after use, the gentle mixing action that it provokes and the avoidance of overheating of the products. It is also usual to process the mixing directly into the container to be used for the storage and transfer of the batch, thus avoiding handling operations which may cause severe changes in the homogeneity of the blends. One typical machine con sists in a cylindrical drum inclined at a 45 angle, xed on a hoop system, which is in turn placed between two rolls driven by a motor. Under these conditions the bulk powder is subjected to rotary cascading ow with a superimposed to and fro motion, giving both radial and axial movements. This may be viewed as a sort of series of reversible surface avalanches. In scientic investigations (Donald & Roseman, 1962a,b) in physics but also in chemical engineering, the most studied mixer is undoubtedly the horizontal rotating drum in which only radial motion, but also segregation, can take place, thus being quite far from e ective chaotic ow found in industrial mixers. Only recently have experimental studies showed that the application of a small-amplitude vertical rocking motion results in a very signicant enhancement of the axial mixing rate in a rotating cylinder (Wightman & Muzzio, 1998a,b). The kinetics of mixing in any type of mixer can be followed by measuring (or estimating) the standard deviation of mixture composition for various mixing times. A typical kinetic curve (see Fig. 1) begins by a sort of exponential decay followed by an asymptote that can be approximated by a series of oscillations. It can be assumed that the rst part of the curve represents the rapid reduction in heterogeneity due to important large scale re-arrangements of the bulk induced by convection and shear mechanisms. These macroscopic motions eventually become more ine cient in ensuring mixing action. This is the case at a ner scale, where di usion mechanisms tend to be the way individual particles (or small packets) are moved to achieve a better mix. Finally, whilst di usion favours mixing, it can also favour segregation by percolation, so that the oscillations in the tail

Fig. 1. Typical powder mixing kinetic curve showing the preponderant mechanisms.

of the kinetic curve can be attributed to a competition between mixing and segregation. In the late 1950s, Rose (1959) developed a kinetic model that takes into account the e ect of segregation on the shape of the curve. This basically consists in dening a mixing degree M and a segregation degree linked by dM = A(1 M ) B ; dt
2

= 1 M:

(2)

In the latter A and B are constants (they may depend on the material and on the operating conditions), while M is dened as M =1 s s0
2

(3)

Eq. (2) can be solved to give the following set of results, depending on the initial disposition of the products ( 0 if the denser one is placed at the bottom): M = {1 [B=A + (1 B=A) exp(At=2)]2 }; M = {1 [B=A (1 + B=A) exp(At=2)]2 }; 0; 0: (4)

4056

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

Fig. 2. Importance of the number of samples required for a correct estimation of homogeneity (to a 95% condence interval) according to Schoeld (1970, 1976).

Also, optimum mixing times corresponding to each case can be found: tm = tm = 2 ln[1 A=B]; A 2 ln[1 + A=B]; A 0; 0: (5)

sure that the mixing states are really di erent from each other. For example, if the number of samples n is relevant, a condence interval Ic for the estimated standard deviation s at a certain risk can be derived from a 2 distribution function table (see Raasch & Sommer, 1990 or Massol-Chaudeur, 2000). But in the industrial practice, the number of samples is usually xed (10 30 samples) and the condence interval on the standard deviation becomes of a limited signicance: Ic = s n1
2 =2; n1

It must be denoted that other formalisms based on modelling the re-arrangements of packets of particles (Lai, Fan, & Akao, 1978) or by using Markov chain theory for the case of static mixers (Wang & Fan, 1977) have also been proposed. However, it is worth noting that most previous studies used thief probes for sampling, but also only estimated the standard deviations with a reduced number (typically 30 out of some thousands) of samples, whatever the mixing time. Fig. 2, due to Schoeld (1976), illustrates the problem associated with the number of samples required for a correct estimation. It is clear that little can be said about the real standard deviation if the number of samples is not validated independently from the study of the mixing kinetics. The use of a constant number of samples over the whole kinetic curve is also an important feature, as heterogeneity is greater at shorter times than for longer times. This has two practical consequences: When a constant number of samples is withdrawn from the bulk at each point on the curve, the accuracy of the estimation is lower for higher degrees of heterogeneity, typically for small mixing times than for longer mixing times. In other words, an identical accuracy for all points in the kinetic curve will need much more samples for smaller times than for greater ones. The determination of the optimum mixing time needs a comparison between the error bars, or the standard deviation of the standard deviation, at any mixing time to be

;s

n1
2 1 =2; n1

(6)

It follows that any experimental study of powder mixing kinetics involving estimations of mixture homogeneity which aims at determining, say the mixing time, requires an independent and rigorous validation of the sampling process.

3. Experimental set-up and methods 3.1. Particulate systems The binary mixtures under study were chosen as being similar to typical pharmaceutical systems, with the additional constraint of not using active ingredients in the laboratory for safety reasons. Sodium saccharinate was chosen to simulate the active drug substance, and three excipients di ering in shape and size, but of the same chemical nature (lactose -monohydrate) were used. They will be further referred to as LFF (spray dried lactose), G70 (ground lactose of the greatest particle size) and G140 (ground lactose of the smallest particle size). A 1% w=w drug concentration was xed to represent to low dosage pharmaceutical mixtures.

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

4057

Fig. 3. SEM photographs of the di erent powders used: sodium saccharinate (a); lactose LFF (b); lactose G70 (c); lactose G140 (d). Table 2 Some physical properties of the powders used

Sodium saccharinate d10 ( m) d50 ( m) d90 ( m) Span (d90 d10 )=2d50 Aerated bulk density Tapped bulk density Density Shape Flow function (kPa) Comment 20 100 225 1.02 0.74 0.99 1.70 Cubic 3.25 Cohesive

Spray-dried lactose LFF 60 115 200 0.61 0.58 0.66 1.54 Spherical 1.94 Free- owing

Ground lactose G70 30 115 220 0.85 0.67 0.93 1.54 Cubic 2.50 Free- owing

Ground lactose G140 20 70 140 0.86 0.63 0.90 1.54 Cubic 2.81 Cohesive

The particle size distributions (PSD) were determined using a MastersizerJ. LASER di raction instrument. Sodium saccharinate was passed through a 200 m stainless steel sieve before use, nevertheless it has the widest PSD. The spray dried lactose LFF and the large size ground lactose G 70 have very similar PSD, even though the former is a little narrower. Observation by scanning electronic microscopy (see Fig. 3) shows that due to their di erent origins, all these products have di erent morphologies. The G70 and G140 particles seem to have a cubic shape whereas the LFF particles are clearly spherical. Therefore, comparison of mixing characteristics of blends made with G70 or G140 should provide an analysis of the e ect particle size in the absence of morphological factor. But also, since their size range is close enough, the in uence of morphology in the absence the particle size factor can be examined by comparing results obtained with LFF and G70. Bulk densities were determined by measuring the mass of powder contained in a measuring cylinder of 100 cm3 lled by aerated ow and after 180 standard taps. Flow

characteristics were determined by means of an annular shear tester by successively placing the di erent powders into a 130-mm-diameter ring, rotating at a shear velocity of 0:15 mm=s. This test showed that the spray dried lactose is the most free- owing powder, while sodium saccharinate and G140 seem to be more cohesive. The main characteristics of these powders are summarised in Table 2. 3.2. Mixer and mixing conditions The powders were mixed in an ErwekaJ AR402 hoop mixer comprising a 5 l stainless steel tumbler drum placed in a hoop with a 45 angle (see Fig. 4). This system is rotated by two rollers driven by a stepping motor. The speed of rotation is kept constant for each experiment and equal to 15 rpm. The total powder load for each experiment is also constant and equal to 40% of the total apparent vessel volume. This lling ratio corresponds to a powder mass of 1300 g for mixtures containing LFF or Granulac70 and a powder mass

4058

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

would be 130 mg (in fact the calibrated size of many of the current thief probes), this dened 510 samples for the band. All these samples were carefully taken, dissolved, diluted, and nally analysed by HPLC, with the aim of generating su cient data to apply statistical rules of sampling. 4. Results and discussion 4.1. Validation of the sampling procedure First, it is important to note that the golden rules of sampling are not violated with this experimental methodology. In fact, the one-dimensional layer which is obtained here not only allows sampling over the whole width of the ow stream, but also to freeze the mixture in the same conguration of that of the ow itself. Fig. 5 shows the raw results obtained for 500 samples, without considering the extremities of the layer. It seems that these are quite grouped around the mean, which means that the homogeneity of the mixture would probably have satised a pharmaceutical standard (a typical gure is: = 6%). The values found for the (weight based) true mean and true standard deviation are as follows: = 0:96 102 ; = 5:44 104 : With the data generated by this single (but also very tedious) experiment, it is possible to realise virtual sampling of any number n of samples, with the help of a table of random numbers. However, as all samples are of di erent compositions, the value obtained from one sampling run is not su cient to be representative of the overall sampling procedure, principally if n is small. Thus, an important number of runs must be made to correctly estimate the e ect of taking only n samples out of 500. In this study, 100 sampling runs of collecting n samples were done, assuming that this number is statistically relevant (in fact it can be demonstrated that the results follows a normal law). The results of this procedure can be appreciated in Fig. 6, in terms of mean and standard deviation. As expected, the dispersion of the data is greater when fewer samples are considered.

Fig. 4. Laboratory hoop mixer used in this work.

of 1380 g for mixtures containing Granulac140. The vessel is loaded in a horizontal position with the active component placed as a horizontal layer between two equal volume layers of lactose. At the end of any mixing experiment, the mixture is gently discharged through a funnel on to a conveyor belt operating at constant speed (5 m min1 ). This gives a regular powder layer of constant thickness and width, the whole discharged mixture corresponding to nearly 20 times the length of a single belt. From this, samples of equal size are taken by means of a hand vacuum pump. Each sample is then dissolved and diluted in 50 ml of bu er solution at pH = 8 and analysed by HPLC. For LFF plus active mixtures, the following mixing time were chosen, so as to give a denite number of revolution of the mixer which is also important when opening the tumbler: 0; 32; 64; 96; 128; 192; 256 and 512 s. For the other products, the study was restricted to : 0; 128; 256 and 512 s. 3.3. Sampling protocol Apart from the kinetic study, and as explained above, it is necessary to develop a specic and experimental procedure for estimating the number of samples to be considered. Let n0 be the number of samples which is necessary to take for providing a good (say, at a certain precision) estimation of the homogeneity of one single layer (band). The basis of our reasoning lays in assuming the following: as the whole discharged mixture corresponds to twenty times the length of one single layer, a good estimation of the overall blend homogeneity would result in taking 20n0 samples. For this reason, prior experimental work consisted in preparing a mixture with LFF and the active ingredient corresponding to the length of one single layer, with the objective of extrapolating the results to 20 belt lengths. Assuming that a typical scale of scrutiny in pharmaceutical processes

Fig. 5. Distribution of the compositions of the samples along the band.

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

4059

Fig. 6. Distribution of the means and standard deviations observed for 100 runs and for 3, 7 and 20 samples taken (respectively in gure a-1, a-2 and a-3 for the means, and in gure b-1, b-2 and b-3 for the standard deviations).

Further, and as mentioned above, the estimation is linked to a certain degree of condence. Let us choose this acceptable condence as being 5% of the real mean, a gure that is motivated by current industrial practice. The 95% condence intervals, either for the mean and the standard deviation, are depicted in Fig. 7. It can be seen that using at least 4 samples gives (in 95% of the cases) the estimation of the mean at a condence interval of 5%. Returning to the assumption given at the end of Section 2, we conclude that at least 80 samples must be considered for the estimation of the mixture quality for the overall contents of the laboratory mixer to within 5%. This nding poses problems if it is scaled at the level of an industrial mixer containing

nearly 1 million of unit doses per batch, as it suggests that several thousands of samples (or tablets) should be taken and analysed for an appropriate estimation, even in the case of a good mix. Undoubtedly, this would drive to the necessity of considering a di erent operating procedure in the process, such as the use of a premix stage, or a change in particle sizes. The overall procedure of mixing sampling and analysis was further validated by performing three experiments in exactly the same conditions: LFF excipient, 80 samples randomly taken, mixing time 512 s. The means were compared two by two using a Student law, while a FischerSnedecor test was used to compare the variances (Caporal-Gautier

4060

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

4.2. Initial mixtures In each experiment the rst measurements correspond to a test performed without using the mixer. The di erent products are introduced in the tumbler as explained above and are directly discharged through the funnel onto the conveyor belt. The study of these initial mixtures is of interest as it is indicative of the ow properties of the mixtures, as well as qualifying the e ect of discharging the powder through the funnel. First, Fig. 8 shows that the compositions of the samples of the mixtures prepared with LFF are closer to the mean composition than are the other ones. In fact, it seems that a good mix is already achieved just by discharging the blend. This could be attributed to the free- owing nature of LFF, which easily destroys the initial segregated structure, so that the funnel acts as a sort of post-mix process. Nevertheless, and as it will be demonstrated further, this initial standard deviation is always higher than those obtained for greater times, thus indicating that even if the funnel introduces a certain disturbance, no e ect on the accuracy of the experimental data can be attributed to this. In the case of the ne ground lactose (G140), the initial structure of the mixture, constituted by superposed layers, can still be detected after discharge through the funnel. For example, the compositions are much greater than 1% between the 20th and the 50th samples. Such an e ect can also be observed with the coarse ground lactose (G 70), but at a smaller scale due to a higher free- owing behaviour of this product with respect to G140. Undoubtedly, the higher cohesion of these two powders means that the corresponding blends are more stable than those produced with LFF. This particular point was conrmed later

Fig. 7. 95% condence intervals for the observed means (a) and standard deviations (b) as a function of the number of samples taken.

et al., 1992). The results (see Table 3) show that it is not possible to discriminate between any of the three runs performed, thus ensuring then their overall reproducibility.

Table 3 Means and variances calculated for three repeated experiments with LFF=sodium saccharinate mixtures (a). Fischer and Student tests showing the equality of the values obtained (b) (a) Run number Sample number Degree of freedom Variance Mean (b) Runs compared (i=j) Experimental F value (variance i=variance j) Tabulated F value F[0:05; 79; 79] Conclusion Calculated t value Tabulated t value t[0:05; 79; 79] Conclusion 1 80 79 3:71 107 9:97 103 2 with 1 1.39 1.45 Fexp Ftab 1.341 1.978 ttab tcalc ttab 2 80 79 5:01 107 9:83 103 2 with 3 1.47 1.45 Fexp Ftab 1:803 1.978 ttab tcalc ttab 3 80 79 3:39 107 1:00 103 1 with 3 1.09 1.45 Fexp Ftab 0.474 1.978 ttab tcalc ttab

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

4061

Fig. 8. Compositions of the three initial mixtures just after discharge onto the belt.

Fig. 9. Kinetics of LFFsodium saccharinate mixing showing the associated 95% condence intervals.

by specic experiments performed in a segregation cell and reported elsewhere (Berthiaux, Muerza, Massol-Chaudeur, & Fages, 1999). However, despite the fact that the funnel introduces some disruption of the mixtures, the initial standard deviations of all of these are high enough (and greater than those obtained for the other times) to conclude that the sampling device does not really mask the e ect of mixing. 4.3. Following the mixing kinetics for LFFsodium saccharinate blends First, it is necessary to compare two by two the experimental variances obtained. Therefore, a FischerSnedecor test was again applied, the results being summarised in Table 4. If one keeps in mind the tabulated value of this F-test (Ftab = 1:45), it can be concluded that all variances are di erent except for those found at 64 and 96 s, as well as for 192 and 256 s. For these two couples, it cannot be said, always to within 5% error, that homogeneity is better in one case rather than in the other. Fig. 9 shows the kinetic curve obtained with the associated 95% condence interval calculated by Eq. (6). It is found that, the higher the mixing time, the smaller the standard deviation, and the narrower the condence interval. On the basis of a current standard (coe cient of variation inferior to 6%), it can be said that the optimum mixing time is equal to 250 s. On the other hand, no signicant e ect can be detected of the mixing time on the homogeneity reached after 200 s, nor on the competition between mixing and segregation (which would have resulted in oscillations). This latter result is in
Table 4 Results of the Fischer test for LFFsodium saccharinate mixtures Runs Compared F value Equality 0s with 32 s 4.25 No 32 s with 64 s 2.21 No 64 s with 96 s 1.25 Yes 96 s with 128 s 1.73 No 128 s with 192 s 2.86 No 192 s with 256 s 0.84 Yes 256 s with 512 s 2.07 No Fig. 10. E ect of the number of samples considered on the accuracy of the kinetic curve (LFFsodium saccharinate mixtures).

contradiction with many of the published works mentioned above. This e ect can be attributed to the larger number of samples considered here (80) with respect to previous studies (between 10 and 30, even for mixers of a greater volume). But it can also be a consequence of the highly diluted mixture under consideration. The displacement of shear planes induced by the motion of the mixer is of a higher scale than the state of dilution of the blend. As a consequence, the mixer becomes ine cient to achieve a very good mix as practically no di usive mixing, but also no segregation by percolation, can take place. To demonstrate the importance of taking so many samples, experimental standard deviations were calculated again by randomly choosing n samples out of the 80 samples really taken, thus introducing a slight bias. Fig. 10 shows

4062

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

this e ect with ve di erent values of n . If this number is greater than 30, the kinetic curve seems to be similar to that obtained with 80 samples, but of course all the condence intervals (not reported here) are narrower in this latter case. If the number of samples considered is smaller than 30, the standard deviation decreases and increases irregularly with respect to the mixing time. For example, in the case of n =20, a FisherSnedecor F-test shows that all the variances must be considered as equal after 96 s of mixing (except for the test at 192 s). Therefore, considering only 20 samples, could have lead to the conclusion that operating the mixer for 96 s is su cient to achieve a good mixture, while it is demonstrated above that the operational mixing time is 250 s. Furthermore, examining the oscillations of this curve could also have lead to the conclusion that a competition between segregation and mixing was taking place. Again, we must emphasise the extreme importance of developing and validating a specic sampling procedure before drawing any sort of conclusion in powder mixing kinetics studies. On another hand, when the operational mixing time has been dened and validated (e.g. at the laboratory scale), the number of samples to be considered can certainly be drastically reduced. In the present case, if 250 s is chosen as the reference, the sampling of 20 (and even 10) unit doses gives a rather good picture of the actual homogeneity. This is particularly important if one wants to refer to the industrial practice and standard of withdrawing a limited number of samples from the mixture. 4.4. In uence of the size and shape of the excipient Bearing in mind the results obtained with LFF, Fig. 11 shows the results obtained for the two other excipients in exactly the same conditions. Table 5 gives further information about the crossing of the condence intervals (not reported here for convenience and clarity). When comparing the two ground excipients, it seems that the di erence in particle size has an insignicant e ect on mixing kinetics, and as a consequence on the required operational mixing time (approximately 2 min). The same values are obtained for any of the mixing times under consideration, which indicates that G140 and G70 can probably be employed without any specic care. Furthermore, and as can be seen from Fig. 12 (a and b), 10 samples (and probably less) would have been enough to provide a very good estimation of the actual homogeneity in both cases, which is also a proof that a good mixture quality is reached. This is also conrmed by a FisherSnedecor F-test, showing that standard deviations calculated for any mixing times are all equal, whatever the number of samples n . On the other hand, comparing G70 with LFF at constant particle size, a certain di erence can be observed as the condence intervals do not cross each other. Despite the fact that few runs were made with the ground product, this clearly suggests that particle morphology has a signif-

Fig. 11. In uence of the size and shape of the excipient on the mixing kinetics.

icant in uence on the kinetics of mixing in such a mixer. This is perhaps clearer from a graphical point of view in Fig. 13, where semi-logarithmic plots are considered, and from which three mechanisms can be identied: (I) a convective region, corresponding to the displacements of blocks of powder, (II) a shear region corresponding to a sliding effect induced by the avalanches, (III) a di usion region corresponding to re-arrangements at the level of small packets of particles. As commented before, it seems that the avalanche motion of the bulk, together with the diluted state of the mixture, do not allow di usive mixing whatever the type of excipient. Further, it seems that the shear region is wider for the spray-dried lactose than for the ground products, which results in a sort of greater slowing down factor for the LFF. This could be attributed to the fact that the avalanche stopping distance is greater for spherical particles than for cubic ones, so that the blends undergo motion from the top to the bottom of the mixer without e cient mixing, in the middle part. It would perhaps be advisable to increase the speed of rotation, or to place some inserts in the tumbler, in order to provoke an additional convective e ect. But in any case, it can be seen that the operational mixing time is di erent for ground and spray-dried products, which also means that the current industrial practice of changing the excipient (of the same chemical nature) while maintaining constant mixing parameters is particularly hazardous if spherical particles (such as LFF) are used. In order to have an overall quantication of the in uence of the excipient, we nally derive the following criteria P80 formed by the coe cient of variations, and in which k is the number of mixing times under consideration, n being the number of samples taken: P80 =
k i=1 ((CVi; n

CVi; 80 )=CVi; 80 )2 : k

(7)

This criterion gives an estimate the relative error if the number of samples considered is less than 80. In the present case, k is equal to 4 as the mixing times that can be considered are: 0; 128; 256 and 512 s.

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065 Table 5 Results of the Fischer test for G140 and G70 sodium saccharinate mixtures

4063

Excipient Runs Compared F value Equality 0s with 128 s 196.2 No

G140 128 s with 256 s 1.07 Yes 512 s with 256 s 4.2 No

G70 0s with 128 s 320 No 512 s with 128 s 1.37 Yes

Fig. 13. Semi-logarithmic plots of the kinetic curves obtained for the three excipients, showing the mixing mechanisms at play.

Fig. 12. E ect of the number of samples considered on the accuracy of the kinetic curve for G140 sodium saccharinate (a) and G70 sodium saccharinate (b) mixtures.

As a rst approach, the results (see Fig. 14) show that P80 seems to have a linear relation with the number of samples for all the excipients. However, the slope of the line is much higher for the case of LFF than for G70 or G140, these being practically equal to one another. Because all curves are forced to coincide at the point [80, 0], this means that LFF sodium saccharinate blends are globally less robust than the others. Again, this particular excipient seems to exhibits a higher sensitivity to segregation, a certain resistance to mixing, and an accurate characterisation of the state of mixing will always require a higher number of samples (or reversely: a lower accuracy on the estimation is to be expected when using pharmaceutical standards for sampling).

Fig. 14. Evolution of the criteria P80 indicating higher sensitivity of LFFsodium saccharinate mixtures to separate.

5. Concluding remarks In this work, an experimental method is developed and used to provide an accurate determination of batch mixing kinetics in a certain type of mixer, and for certain types of powders. While it is sure that no sampling technique is universal, the methodology presented could be employed with probably few changes at the laboratory scale for other types of mixers and other products, including multiconstituent blends. But the key point in such kinetic studies and in the absence of on-line and non-destructive measurement, is the

4064

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065

statistical accuracy of the calculations to provide a good estimation of the various mixing states. In particular, errors are all the more important the shorter the mixing time, which in turn makes di cult the search for the practical optimum mixing time. Scaling up is another part of the problem, as when considering higher volume mixers (remember the scale of scrutiny is still the same) it is unreasonable to expect that thousands of samples could be analysed to repeat the same procedure as developed here. Whilst it may be possible to achieve a complete study at the laboratory level, it is probable that at larger scales it will consist in verifying the optimum rather searching for it. When referring to current industrial practice, attention should also be drawn to the changes in mixing behaviour found when one type of excipient is changed for another of the same chemical nature but with di erent physical characteristics of the particles. Mixtures prepared with nearly spherical particles are found to exhibit a certain resistance to a quick mixing and a higher segregability in mixers that induce the shear mechanism. Indeed, the direct way is not advisable for blends made with such an excipient with respect to the other cubic ones that are more likely to be handled. This is the tribute to pay for ensuring the free- owing character of the blends, which is also another part of the industrial specications. In any case, it should be observed that the gain by optimising batch mixing operations, for example on the basis of the operational mixing time, is not the key factor for the industry, despite it being important for reproducibility of the process. In the pharmaceutical industry a mixer is operated for less than 1 h per day, the rest of the time (some 7 h) being devoted to other tasks such as emptying, cleaning and the validation of the cleaning, : : : . When other stages in the processes are continuous, it is clear that a technological change from the present batch mixers to continuous mixers would generate a higher prot than saving some few minutes of daily operation, reproducibility being the fact of the stability of the feeders. The need for developing new products, with low dosage of the active ingredient, will probably emphasise the need for developing new, automated, and on-line controlled processes, which will in turn change the regulatory framework of avoiding any process modication.

n s t tm xi xm Greek symbols

number of samples taken, dimensionless estimated standard deviation, dimensionless time, s optimum mixing time, s composition of sample i, dimesionless estimated mean, dimensionless

risk associated to a condence interval, dimensionless segregation potential dened by Rose, dimensionless constant in Roses formula, dimensionless true mean, dimensionless true standard deviation, dimensionless distribution function for samples, dimensionless

Acknowledgements The authors would like to thank the Pierre Fabre Group and the region Midi Pyr n es for nancial support. e e References
Berman, J., & Planchard, J. A. (1995). Blend uniformity and unit dose sampling. Drug Development and Industrial Pharmacy, 21(11), 12571283. Berman, J., Schoeneman, A., & Shelton, J. T. (1996). Unit dose sampling: A tale of two thieves. Drug Development and Industrial Pharmacy, 22(11), 11211132. Berthiaux, H., & Dodds, J. A. (1999). Modelling ne grinding in a uidised bed opposed jet mill: Part I-Batch grinding kinetics. Powder Technology, 106, 7887. Berthiaux, H., Muerza, S., Massol-Chaudeur, S., & Fages, J. (1999). D veloppement dun test pour evaluer la s gr gabilit des m langes e e e e e pulv rulents. R cents Progr s en G nie des Proc d s, 13(63), e e e e e e 153158. Caporal-Gautier, J., Nivet, J. M., Algranti, P., Histe, M., Lallier, M., NGuyen-Huu, J. J., & Rusotto, R. (1992). Guide de validation analytique-I. M thodologie. STP Pharma Pratiques, SFSTP e commission report, 2 4, pp. 205 239. Donald, M. B., & Roseman, B. (1962a). Mixing and demixing of solids particlesPart I: Mechanisms in a horizontal drum mixer. British Chemical Engineering, 7(10), 749753. Donald, M. B., & Roseman, B. (1962b). Mixing and demixing of solids particlesPart II: E ects of varying the operating conditions of a horizontal drum mixer. British Chemical Engineering, 7(11), 823827. Dyakowski, T., Jeanmeure, L. F. C., & Jaworski, A. J. (2000). Applications of electrical capacitance tomography for gassolids and liquidssolids owsA review. Powder Technology, 112, 174192. Harnby, N. (1992). The selection of powder mixers. In Harnby, Edwards, & Nienow, (Eds.), Mixing in the process industries (2nd ed.). Oxford: ButerworthHeinemann. Kaneko, Y., Shiojima, T., & Horio, M. (2000). Numerical analysis of particle mixing characteristics in a single helical ribbon agitator using DEM simulation. Powder Technology, 109, 5564.

Notation A B Ic i M N constant in Roses formula, s1 constant in Roses formula, s1 condence interval, dimensionless index corresponding to a sample, dimensionless Rose mixing index, dimensionless overall number of samples, dimensionless

S. Massol-Chaudeur et al. / Chemical Engineering Science 57 (2002) 40534065 Lacey, P. M. C. (1954). Developments in the theory of particle mixing. Journal of Applied Chemistry, 4, 257268. Lai, F. S., Fan, L. T., & Akao, Y. (1978). The convective mixing process and striated mixture. Journal of Powder and Bulk Solids Technology, 2, 3851. Lai, C. K., Holt, D., Leung, J. C., Cooney, C. L., Raju, G. K., & Hansen, P. (2001). Real time noninvasive monitoring of dry powder blend homogeneity. AIChE Journal, 47(11), 26182622. Lantz, R. J., & Schwartz, J. B. (1981). Mixing. In Lieberman, & Lachman (Ed.), Pharmaceutical dosage forms: Tablets. New York: Marcel-Dekker. Massol-Chaudeur, S. (2000). Caract risation de l tat de m lange de e e e poudresCas de m langes faiblement dos s. Ph.D. thesis, INPL e e Nancy. Muzzio, F. J., Robinson, P., Wightman, C., & Brone, D. (1997). Sampling practices in powder blending. International Journal of Pharmaceutics, 155. Muzzio, F. J., Shinbrot, T., & Glasser, B. J. (2002). Powder Technology in the pharmaceutical industry: The need to catch up fast. Powder Technology, 132, 17. Raasch, J., & Sommer, K. (1990). Anwendung von statistischen Prufverfahren im bereich der mischtechnik. Chemical Engineering, 62(1), 1722.

4065

Rose, H. E. (1959). A suggested equation relating to the mixing of powders and its application to the study of the performance of certain type of machine. Transactions of the Institution of Chemical Engineers, 37, 47. Scarlett, B. (2001). How will Chemical Engineering in uence pharmaceutical engineering practice? Plenary lecture at the Albi International Rencontres in Pharmaceutical Engineering. Schoeld, C. (1970). Assessing mixtures by autocorrelation. Transactions of the Institution of Chemical Engineers, 48, T28T34. Schoeld, C. (1976). The denition and assessment of mixture quality in mixtures of particulate solids. Powder Technology, 15, 169180. United States versus Barr Laboratories Inc. (1993). Food and Drug Law Reports, 4:4. Wang, R. H., & Fan, L. T. (1977). Stochastic modelling of segregation in a motionless mixer. Chemical Engineering Science, 32, 695701. Wightman, C., & Muzzio, F. J. (1998a). Mixing of granular material in a drum mixer undergoing rotational and rocking motionsI: Uniform particles. Powder Technology, 98, 113124. Wightman, C., & Muzzio, F. J. (1998b). Mixing of granular material in a drum mixer undergoing rotational and rocking motionsII: Segregating particles. Powder Technology, 98, 125134.