Prediction of Bulk Powder Flow Performance Using Comprehensive Particle Size and Particle Shape Distributions

WEILI YU, KOJI MUTEKI, LIN ZHANG, GLORIA KIM Pzer Global Research and Development, Eastern Point Road, MS 8156-007, Groton, Connecticut 06340 Received 8 September 2009; revised 24 March 2010; accepted 3 May 2010 Published online 22 June 2010 in Wiley Online Library (wileyonlinelibrary.com). DOI 10.1002/jps.22254 ABSTRACT: The purpose of this study is to establish a modeling approach that can be used to predict bulk powder owability of pharmaceutical materials from their particle size and shape distributions. To build and validate the model, 23 commonly used pharmaceutical excipients and 38 binary blends were fully characterized for their particle size and shape distributions. The particle size and shape of each sample was characterized by multiple descriptors to fully reect their morphological characteristics. The ow properties of these materials were analyzed using the Schulze Ring Shear Tester at a xed humidity condition. A partial least squares (PLS) approach was used to build the mathematical model. Several different modeling approaches were attempted and the best method was identied as using a combination of formulation composition and particle size and shape distributions of single-component powder systems. The PLS model was shown to provide excellent predictions of powder ow function coefcient (FFC) of up to approximately 20. The results also revealed that both particle size and shape play an important role in determining the powder ow behavior. 2010 Wiley-Liss, Inc. and the American
Pharmacists Association J Pharm Sci 100:284293, 2011

Keywords: image analysis; mathematical model; particle size; shape factor of particle; particle sizing; partial least squares; powder ow; pharmaceutical powders

INTRODUCTION
Optimization of powder ow properties is conducted routinely during solid dosage form development to achieve satisfactory drug product manufacturability. The optimization process may include evaluation of various excipients at different concentrations, and conducting powder ow analysis after each alteration until satisfactory performance is achieved. The number of blends and ow analysis that need to be pursued may vary largely depending on the characteristics of the materials at hand, but the resources required to complete this task is often signicant. It has long been shown that powder ow property is dependent on particle size and particle shape characteristics of a material.110 The recent development and application of novel crystallization technologies allow researchers to modify crystal habit of various materials and assess the impact on power owability.1113 Garekani et al. showed different size and shape of ibuprofen particles crystallized from various solvents. Samples obtained from methanol and ethanol were lath/plate-shaped and exhibited
Correspondence to: Weili Yu (Telephone: 1-860-686-2072; Fax: 1-860-686-5632; E-mail: weili.yu@pzer.com)
Journal of Pharmaceutical Sciences, Vol. 100, 284293 (2011) 2010 Wiley-Liss, Inc. and the American Pharmacists Association

better ow properties than those crystallized from hexane that was needle-shaped.12 The particle shape effect on powder owability was also demonstrated by Kaerger et al. using paracetamol. A decrease in the angle of repose (improved owability) was achieved with the spherical paracetamol particles when compared to micronized particles that were lath-shaped.13 More recently, Liu et al. showed that for needleshaped ibuprofen particles, the powder owability was impacted by not only the average particle size, but also by the breadth of the size distribution. The owability of powders with narrower size distributions increased more signicantly with size.14 Podczeck and Mia studied powder ow properties using eight different powders differing in particle size and shape. Their results showed that while the angle of internal friction was dependent on both particle size and shape, the Jenikes ow factor depended only on particle shape.5,15 Numerous attempts have also been made on modeling ow of bulk powder.1620 Liu et al.14 derived a mathematical model that demonstrated the dependence of powder owability on particle size, and obtained qualitative agreement with experimental data collected using needle-shaped ibuprofen samples. Cleary used discrete element methods and demonstrated a strong effect of particle shape on shear ow.18

284

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

PREDICTION OF BULK POWDER FLOW PERFORMANCE

285

Despite all the progresses made to date, quantitative prediction of powder owability of real-world pharmaceutical materials remains challenging. One of the obstacles has been the lack of suitable techniques for quantitative particle shape characterization of bulk powders. In most of the reported studies, particle shape information was obtained using conventional static image analysis techniques (e.g., scanning electron microscopy, or optical microscope) where a relatively small number of particles were analyzed for their shape distribution.21,22 This approach is often tedious and time-consuming, and the sampling errors can be quite signicant.23 In addition, the particle morphologies are often characterized by one or two descriptors which may not be able to fully represent the characteristics of a material.24 In this article, we present a modeling approach developed using the partial least squares (PLS) method to predict powder owability of pharmaceutical materials with complex particle size and shape characteristics. Twenty-three commonly used pharmaceutical excipients and 38 binary blends were fully characterized and were used to build the model. The particle size and shape distributions of the materials were analyzed using laser diffraction and dynamic image analysis (DIA) techniques. The state-of-theart DIA technique employed in the study allows fast analysis of particle size and shape information of millions of particles.2427 The particle size and shape of nonspherical particles often cannot be fully described using a single value.24 To better reect the morphological characteristics associated with
Table 1. Summary of Materials
ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
DOI 10.1002/jps

the materials, the particle size and shape were each represented by three size/shape descriptors. The powder owability was analyzed using a standard Ring Shear Tester under a xed humidity condition.

EXPERIMENTAL
Materials Twenty-three pharmaceutical excipients with various particle size, shape, and ow properties were used in this study, and the materials are summarized in Table 1. All materials were used as received. Thirteen sets of binary powder mixtures were prepared using selected pairs of excipients (components A and B) mixed at volume ratios of 1:3, 1:1, and 3:1, and the compositions are summarized in Table 2. The binary mixtures were blended in a Turbula mixer for 10 min before use. Methods The DIA apparatus QicPic equipped with the RODOS dry powder disperser (Sympatec, Inc., ClausthalZellerfeld, Germany) was used to analyze particle size and particle shape distributions. The QicPic uses rear illumination with a visible pulsed light source that has an exposure time of 1 ns to minimize motion blur. During a measurement, dry powders were fed into the RODOS powder disperser and were dispersed and aerosolized through particleparticle

Name Lactose Monohydrate (FAST FLO1 316) Lactose Anhydrous (SUPER TAB1 21AN) Lactose Anhydrous (DIRECT TABLETING) Lactose-310, NF Monohydrate Lactose 315, NF Monohydrate Mannitol 60 (Pearlitol1 160C) Mannitol (Pearlitol1 300DC) Mannitol 60 Parenteral PF Mannitol Powder Mannitol, Granular 2080 Mannitol (Pearlitol1 200SD) Mannitol Spray Dried (Mannogem EZ1) (nonparenteral use only) Crospovidone (Kollidon1 CL) Povidone (Kollidon1 25) Povidone (Kollidon1 30) L-Ascorbic Acid Sodium Starch Glycolate (ExploTab1 CLV) Sodium Starch Glycolate ExploTab1 Microcrystalline Cellulose (Avicel PH200) Microcrystalline Cellulose (Avicel PH102) Microcrystalline Cellulose (Avicel PH101) Calcium Phosphate Dibasic, Dihyd CellousTM UF-711

Vendor Foremost Farms (Baraboo, WI) DMV-Fonterra Excipients (Norten-Hardenberg, Germany) Foremost Farms Foremost Farms Foremost Farms Roquette Pharma (Lestrem, France) Roquette Pharma Roquette Pharma SPI Pharma Group (Wilmington, DE) SPI Pharma Group Roquette Pharma SPI Pharma Group BASF (Florham Park, NJ) BASF BASF Sigma-Aldrich (St. Louis, MO) JRS Pharma (Rosenberg, Germany) JRS Pharma FMC Corporation (Philadelphia, PA) FMC Corporation FMC Corporation Rhodia (Cranbury. NJ) Asahi Kasei Chemicals Corp. (Chiyoda-kuTokyo, Japan)
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

286

YU ET AL.

Table 2. Mixture Compositions

ID Code M1-1M1-3 M2-1M2-3 M3-1M3-3 M4-1M4-3 M5-1M5-3 M6-1M6-3 M7-1M7-3 M8-1M8-3 M9-1M9-3 M10-1M10-3 M11-1M11-3 M12-1M12-3 M13-1M13-2 Component A Fast Flo 316 Fast Flo 316 Fast Flo 316 Fast Flo 316 Fast Flo 316 Fast Flo 316 Fast Flo 316 Lactose S/D0 Avicel PH 200 Mannitol 2080 Mannitol 2080 Mannitol 2080 Mannitol 2080 Component B Avicel PH101 Lactose 310 Kollidon 30 Lactose SuperTab 21AN Lactose Direct Tableting Lactose 310 MCC UF711 Avicel PH101 Avicel PH101 Avicel PH102 Mannitol Powder Avicel PH101 MCC UF711

from instrument to instrument. Three particle size descriptors available through the QicPic were found relevant in this study, namely the Equivalent Projection Circle (EQPC), minimum Feret diameters (Fmin), and maximum Feret diameter (Fmax). Similarly, three particle shape factors obtained from the QicPic were employed in the current study, including the aspect ratio (AR), sphericity, and convexity. The denitions of these size and shape factors are summarized in Table 3 and were kept consistent with those used by the QicPic instrument to allow easy implementation of the proposed modeling methodology by interested readers. From the denitions, it

and particlewall collisions before entering the measurement zone with a speed of up to 100 m/s. Additional details of the instrument can be found in the literature.25,27 In the current study, at least 100,000 particles were analyzed during each measurement to minimize statistical biases resulted from sampling. The particle size distributions of all the materials were also analyzed using the HELOS (HeliumNeon Laser Optical System, Sympatec, Inc.) laser diffraction instrument equipped with the same RODOS dry powder disperser. Measurement conditions including sample feed rate, dispersing air pressure and frame rate for image-capturing (QicPic only) were adjusted from sample to sample to accommodate different particle characteristics. Appropriate methods for each material were determined by conducting method development studies on both instruments before collecting the reported data. At least three replicates were conducted with each sample using the selected method. Powder ow properties were analyzed using the Schulze Ring Shear Tester RST-XS (Dietmar Schulze Schuttgutmesstechnik, Wolfenbuttel, Germany).28 Each sample was equilibrated at 50 2% RH in a humidity-controlled glove box for 1214 h before testing. At the beginning of a measurement, the powder was preconsolidated at 4 kPa and sheared until a steady shear stress was reached. The powder was then subjected to a series of normal stress ranging from 1 to 2.6 kPa with a preconsolidation step applied (at 4 kPa) before each load.28 From the yield locus obtained from this procedure, the ow function coefcient (FFC) was calculated. Three replicates were conducted with each sample with RSD less than 10%. The average FFC values were used in this study.

Table 3. Denition of Particle Size and Shape Descriptors Used by the QicPic
Descriptors EQPC Fmin Denition Diameter of a circle that has the same projected area as the particle The minimum distance between pairs of tangents to the particle projection in some xed direction. For irregularly shaped particle, Fmin can thus be signicantly smaller than EQPC The maximum distance between pairs of tangents to the particle projection in some xed direction. Fmax can thus be signicantly larger than EQPC The ratio of Fmin to Fmax. The value is between 0 and 1 The ratio of the perimeter of the equivalent circle, PEQPC, to the real particle. Since the equivalent circle gives the smallest possible perimeter at a given projection area, the value of sphericity is between 0 and 1. The smaller the value, the more irregular the shape of the particle. Surface roughness may also increase perimeter of a particle, and thus decrease sphericity The ratio of the projection area of a particle to the area of the convex hull (the latter is calculated from particle projection). The value describes the compactness of a particle. The more concave region a particle has, the smaller the value of convexity The gure below shows a particle with a projection area A (gray) with a concave area B. Convexity is dened as Convexity A/(A B)

Fmax

Aspect ratio Sphericity

Convexity

RESULTS
Particle Size and Shape Descriptors Terminologies used to describe particle size and particle shape by image analysis techniques vary
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011 DOI 10.1002/jps

PREDICTION OF BULK POWDER FLOW PERFORMANCE

287

is clear that the use of different size and shape descriptors makes little difference when analyzing particles that are perfectly spherical. For nonspherical particles, however, each of the size and shape descriptors reects a different aspect of a materials morphological property. For instance, while the AR reects the elongation of a particle and deviation from a sphere, the convexity describes the compactness of a particle. Both these morphological features may in turn affect the powder ow properties as will be shown later. Since bulk powder properties are often affected by mass (or volume) based distribution, both the particle size and shape distributions presented in this study are volume weighted. When weighted by volume, the presence of ne particles in a distribution can be diminished to some extent as the nes typically occupy a relatively small portion of the volume of the bulk powder. In applications where presence of nes is important, number-weighted distribution can be used to provide increased resolution of nes. Figure 1ac demonstrates the effect of using different particle size and shape descriptors when analyzing Lactose 310, and the Scanning Electron Micrograph of the material. The SEM results revealed that the material was a mixture of equant and elongated particles with smooth surface. Consequently, the particle size analyzed using different size descriptors showed the trend of Fmax > EQPC > Fmin. The AR ranged from 0.4 to 1.0, reecting the mixture of equant and elongated morphologies of the particles. The sphericity and convexity ranged from 0.6 to 1.0, consistent with the particle morphology, and also indicated that the particles had relatively smooth surface and compact structures.

Particle Size and Shape of Binary Mixtures The particle size and shape distributions of the binary powder mixtures were measured both experimentally and computed using those of the individual components. Overall, good agreement between the experimental and computed distributions was obtained with all samples. Examples of such comparisons were given in Figure 2ae, where distributions of the EQPC (particle size) and AR (particle shape) of mixture samples M6, M10, and M11 at three different volume ratios were shown. Note that some minor differences exist between the experimental and computed data and this could be attributed to inherent

Figure 1. (a) Comparison of the particle size distributions of lactose 310 analyzed using three size descriptors, Fmax, Fmin, and EQPC. (b) Comparison of the particle shape distributions of lactose 310 analyzed using three shape descriptors, aspect ratio, sphericity, and convexity. (c) SEM micrograph of lactose 310.
DOI 10.1002/jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

288

YU ET AL.

Figure 2. Comparison of the experimental and computed particle size and shape distributions of binary powder mixtures. Particle size and shape distributions were plotted using EQPC and Aspect Ratio, respectively. The scatter plot represents the experimental data, while the solid and dashed lines represent computed data. (a) EQPC distribution of mixture M6-1, M6-2 and M6-3; (b) Aspect ratio plot of mixture M6-1, M6-2 and M6-3; (c) EQPC distribution of M10-1, M10-2 and M10-3; (d) Aspect ratio distribution of M10-1, M10-2 and M10-3; (e) EQPC distribution of M11-1, M11-2 and M11-3; (f) Aspect ratio distribution of M11-1, M11-2 and M11-3.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011 DOI 10.1002/jps

PREDICTION OF BULK POWDER FLOW PERFORMANCE

289

experimental variability and possible presence of particle porosity effect.24 The impact of using the experimental or computed distributions on FFC prediction will be further elaborated in Modeling Section. Powder Flow Properties Powder owability is mainly characterized by its unconned yield stress (su) in dependence on the consolidation stress (sc). The stress that causes an unconned powder bed to ow is dened as the unconned yield stress. The consolidation stress is the stress that compacts or consolidates a powder system. Apparently, the unconned yield stress (su) changes with the consolidation stress (sc), and the ratio between sc and su, dened as the FFC, is a numerical description of the owability of a powder system: FFC s c =s u The larger the FFC value, the better the powder ows. Powder ow behaviors can be classied as the following:29,30 FFC < 1 not owing; 1 < FFC < 2 very cohesive; 2< FFC < 4 cohesive; 4< FFC < 10 easy owing; 10< FFC free owing. Powders with a wide range of ow properties were selected in this study as shown in Figure 3.
Figure 3. Distribution of the ow function coefcient of materials used in this study.

MODELING
Data Structure The data structure used in this study is demonstrated schematically in Figure 4a. Xsin and Xbin represent the particle size and shape distribution matrix of the single and binary powder systems, respectively. Rsin and Rbin are the compositions (i.e., volume %) of the single and binary powder systems, respectively. It is evident that Rsin is an identity matrix as it shows the composition of the single-component systems. The Ysin and Ybin are the FFC vectors of the single and binary powder systems, respectively. The 61 powder systems (including both single and binary) were used to build and validate the model. Partial Least Squares (PLS) Modeling A PLS approach was employed for the modeling since PLS model can effectively project correlated high dimensional data matrices into low dimensional subspace and provide good estimates on Y.3134 MatLab was used to program the PLS model.
DOI 10.1002/jps

Based on the data structure shown in Figure 4a, a key question is how to build the PLS model which provides the best estimates on the FFC. There are three different potential PLS modeling approaches: (A) use only the formulation compositions (i.e., Rsin and Rbin); (B) use the measured particle size and shape distributions (i.e., Xsin and Xbin); (C) use the formulation compositions (i.e., Rsin and Rbin) and the particle size and shape distributions of the singlecomponent systems (i.e., Xsin). In model (C), as demonstrated in Figure 4b, the particle size and shape distributions of the binary blends were computed from those of the single-component systems and are expressed as Rbin Xsin (calculated by ideal mixing rule35). For the single component systems, Rsin Xsin is the same as Xsin because Rsin is an identity matrix. For all the above PLS models, due to obvious nonlinearity in the FFC data, a logarithmic transformation of the FFC was used. To evaluate the different modeling approaches, R2 (the fraction of the cumulative sum of squares of Y explained by the model t) and Q2 (the fraction of the cumulative sum of squares of prediction obtained from cross-validation) are used. The Q2 value is a more important indicator on models predictability. The number of latent variables is selected so as to maximize the Q2 value. The R2 and Q2 values for the above PLS models are shown in Table 4. The PLS model (C) provided better estimates on FFC (in terms of Q2), compared to the PLS models (A) and (B). The reason model (A) provided the worst result is likely due to the lack of information in X for predicting FFC (i.e., formulation compositions alone cannot predict
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

290

YU ET AL.

Figure 4. (a) Experimental data structure that contains 23 single-component and 38 binary powder mixtures. (b) Data structure used to build the PLS model. The particle size and shape distributions of the binary mixtures were computed using those of the individual components.

FFC reliably). It is suspected that the improvement model (C) showed over model (B) is probably due to reduction of inherent variability embedded in the measurement of particle size and shape of binary blends. The most important advantage model (C) provides is that it allows quick estimate of the FFC of formulated blends even in the absence of the particle size and shape distributions of the blends, as the distributions will be computed using those of the blend components. It thus opens the door to virtual selection and optimization of formulation components and compositions.32 For these reasons, the modeling approach in (C) was used in this study. Another important task during modeling was to select the most appropriate X variables. The X

variables (particle size and shape) available for modeling were summarized in Table 5. A total of 275 X variables were compiled for each material. Several modeling trials were conducted using various X variables and the results are summarized in Table 6. The

Table 5. Summary of the X Variables Used in the PLS Model

No. 1 2 3 4 5 6 7 8 9 Description of X Variables Distribution of EQPC Distribution of Fmin Distribution of Fmax Distribution of aspect ratio Distribution of sphericity Distribution of Convexity Aspect ratio vs. EQPC Distribution of HELOS particle size The 10th, 50th, and 90th percentile values obtained from distributions of EQPC, Fmin, Fmax, aspect ratio, sphericity, and convexity Total The Number of X Variables 31 31 31 20 20 20 31 73 18

Table 4. Result of Three Different PLS Models

R2 (%) PLS model (A) PLS model (B) PLS model (C)
a

Q2 (%) 31.8 60.9 71.0

PCa 2 4 4

91.0 88.5 86.2

The number of PC was selected so as to maximize the Q2 value.

275
DOI 10.1002/jps

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

PREDICTION OF BULK POWDER FLOW PERFORMANCE

291

Table 6. Evaluation of Different Modeling Approaches That Differed in X Variables

Model # 1 2 3 4 5 6 7
a b

X variablesa All (19) All but #9 All but 8 17 1, 2, 4, and 7 1 and 2 8

R2 (%) 86.2 85.8 78.4 77.3 78.9 75.7 48.7

Q2 (%) 71.0 67.3 63.7 59.3 54.3 56.7 31.7

PCb 4 4 3 3 4 2

most signicant impact on the FFC are plotted in the Variable Importance in the Projection (VIP) plot in Figure 6. The results revealed that the Fmin, EQPC, and AR had the most impact on the FFC prediction. Interestingly, all the top 20 particle size and shape variables were obtained from the QicPic measurements. These results conrmed that both particle size and particle shape play an important role in determining powder ow properties. Model Reliability and Maintenance As for any predictive model, it is important to be able to assess uncertainties in the predicted FFC values to prevent erroneous predictions being made unintentionally. For instance, when the input data contain false measurements or lies outside the region on which the model was built, the predicted FFC values will generally have a large uncertainty. In the current model, the square prediction error (SPE) using latent variables of input matrix is used to represent the reliability of predicted FFC values. The SPE is calculated as a sum of squared errors between the actual ^ X and the X projected to the latent variable space.32 If the SPE of a new material exceeds certain threshold (e.g., 99% SPE condence limit of training samples), the prediction will be agged and experimental approach may be needed in determining the FFC of the material. When such an incident occurs, it is an indication of the false measurement or need of model update and maintenance. Addition of the new data to the model database will help to extend the applicable range of the model and improve model robustness. Model Application The capability of estimating powder ow properties using simply the particle size and shape distributions makes it possible to quickly screen through a large number of formulations with very little resource. The fact that the particle size and shape distributions of commonly used excipients can be used to establish the model in the rst place means that these data will be readily available at the time of the prediction. Often the only new data that need to be generated before the prediction are the particle size and shape distributions of the drug itself. Particle size and shape distributions of formulated blends can be easily computed from those of the individual components and used to predict the ow property of the virtual formulations. The model thus enables quick evaluation and screening of a large number of virtual blends with varying compositions for their ow property. This approach is perhaps particularly useful for direct compression processes where the ow properties of formulated blends often present a challenge, but are critical to manufacturability and drug product quality.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

Please see Table 5 for descriptions of the corresponding X variables. The number of principal components was so selected to maximize the Q2 value.

results showed that the best model (Model 1) was obtained when all the 275 X variables were incorporated. It is interesting to note that when using data obtained from the QicPic alone (Models 36), the model performance decreased only slightly comparing to Model 1. But when using only the particle size data obtained from the laser diffraction (HELOS), the model performance decreased signicantly (Model 7). The results indicate that for nonspherical particles (1) the bulk powder owability is impacted by both particle size and shape and (2) expressing particle size and shape in various ways help to better capture the morphological characteristics of the particles that are relevant to powder ow.

DISCUSSIONS
Comparison of Predicted and Experimental FFC Figure 5a shows the results of the predicted FFC values of the 61 materials plotted against their experimental values. The diagonal solid line represents where Y equals X. Model 1 was used for the prediction, in which a logarithmic transformation of FFC was incorporated. Prediction result without logarithmically transforming the FFC was given in Figure 5b for comparison purposes. It is clear that the logarithmic transformation improved model predictability signicantly. The results also indicated that the model works the best in the lower FFC range. As FFC exceeds $20, the model appears to underestimate the FFC values. While more studies may be necessary to improve the model predictability in the higher FFC range for academic interest, practically, powders with FFC larger than 10 are already considered free owing and thus less likely to have ow issues. Impact of Particle Size and Shape on FFC The PLS modeling provides quantitative analysis of the importance of the X variables on the FFC prediction. The top 20 variables that showed the
DOI 10.1002/jps

292

YU ET AL.

Figure 5. (a) Comparison of the experimental and estimated FFC values with logarithmically transforming the FFC values. (b) Comparison of the experimental and estimated FFC values without logarithmically transforming the FFC values.

SUMMARY
A PLS model was successfully developed to estimate the FFC of bulk powder materials using comprehensive particle size and particle shape distributions. The model was established using 61 single-component and binary pharmaceutical particle systems whose morphological characteristics were analyzed and expressed
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011

using multiple size and shape descriptors. The model revealed that among the particle size and shape descriptors, the Fmin, EQPC, and AR showed the most signicant impact on the FFC prediction. While the result conrmed the earlier ndings reported in the literature that both particle size and shape play an important role in determining powder ow, it also revealed the importance of expressing particle size/
DOI 10.1002/jps

PREDICTION OF BULK POWDER FLOW PERFORMANCE

293

Figure 6. Result of the top 20 causal variable ranking from the PLS model.

shape using multiple descriptors to better capture the morphological characteristics of a powder system. The model allows estimation of particle size and shape distributions of formulated blends from those of the individual components. Once an initial database is established with commonly used pharmaceutical excipients, the model can be used to screen through a large number of formulations to optimize bulk powder ow properties with little resource cost. The model ags predictions with large uncertainties, so that erroneous experimental data or inappropriate use of the model can be noted promptly to avoid false predictions.

REFERENCES
1. Sun C, Grant DJW. 2001. Effects of initial particle size on the tableting properties of lysine monohydrochloride dihydrate powder. Int J Pharm 215:221228. 2. Guo A, Beddow JK, Vetter AF. 1985. A simple relationship between particle shape effects and density, ow rate and Hausner ratio. Powder Technol 43:279284. 3. Ridgway K, Rupp R. 1971. The mixing of powder layers on a chute: The effect of particle size and shape. Powder Technol 4:195202. 4. Johanson JR. 1972. Modeling ow of bulk solids. Powder Technol 5:9399. 5. Podczeck F, Mia Y. 1996. The inuence of particle size and shape on the angle of internal friction and the ow factor of unlubricated and lubricated powders. Int J Pharm 144:187 194. 6. Chan LCY, Page NW. 1997. Particle fractal and load effects on internal friction in powders. Powder Technol 90:259266. 7. Hancock BC, Carlson GT, Ladipo DD, Langdon BA, Mullarney MP. 2002. Comparison of the mechanical properties of the crystalline and amorphous forms of a drug substance. Int J Pharm 241:7385. 8. Rasenack N, Muller BW. 2002. Ibuprofen crystals with opti mized properties. Int J Pharm 245:924.
DOI 10.1002/jps

9. Listiohadi DY, Hourigan JA, Sleigh RW, Steele RJ. 2005. Properties of lactose and its caking behavior. Aust J Dairy Technol 60:33. 10. Fatah N. 2009. Study and comparison of micronic and nanometric powders: Analysis of physical, ow and interparticle properties of powders. Powder Technol 190:4147. 11. Nada A, Al-Saidan SM, Mueller BW. 2005. Improving the physical and chemical properties of ibuprofen. Pharm Technol 29. 12. Garekani HA, Sadeghi F, Badiee A, Mostafa SA, RajabiSiahboomi AR. 2001. Crystal habit modications of ibuprofen and their physicomechanical characteristics. Drug Dev Ind Pharm 27:803809. 13. Kaerger JS, Edge S, Price R. 2004. Inuence of particle size and shape on owability and compactibility of binary mixtures of paracetamol and microcrystalline cellulose. Eur J Pharm Sci 22:173179. 14. Liu LX, Marziano I, Bentham AC, Litster JD, White ET, Howes T. 2008. Effect of particle properties on the owability of ibuprofen powders. Int J Pharm 362:109117. 15. Jenike AW. 1961. Gravity ow of bulk solids. Utah Eng Exp Stn Bull. Salt Lake City, Utah: University of Utah. p 108. 16. Kachrimanis K, Karamyan V, Malamataris S. 2003. Articial neural networks (ANNs) and modeling of powder ow. Int J Pharm 250:13123. 17. Luding S. 2005. Shear ow modeling of cohesive and frictional ne powder. Powder Technol 158:4550. 18. Cleary PW. 2008. The effect of particle shape on simple shear ows. Powder Technol 179:144163. 19. Fraige FY, Langston PA, Chen GZ. 2008. Distinct element modelling of cubic particle packing and ow. Powder Technol 186:224240. 20. Wu J, Chen J, Yang Y. 2008. A modied kinematic model for particle ow in moving beds. Powder Technol 181:7482. 21. Bao N, Shen L, Feng X, Lu X, Yanagisawa K. 2004. Shape and size characterization of potassium titanate bers by image analysis. J Mater Sci 39:469476. 22. Li M, Wilkinson D, Patchigolla K. 2005. Comparison of particle size distributions measured using different techniques. Particulate Sci Technol 23:265284. 23. Jillavenkatesa A, Dapkunas SJ, Lum L-SH. 2001. Particle size characterization, Washington: U.S. Government Printing Ofce. Special Publication 9601. 24. Yu W, Hancock BC. 2008. Evaluation of dynamic image analysis for characterizing pharmaceutical excipient particles. Int J Pharm 361:150157. 25. Witt W, Kohler U. 2004. List J. Partec 2004. Nurnberg: Partec. 26. Witt W, Kohler U, List J. 2005. Particulate Systems Analysis 2005. Stratford-upon-Avon, UK. 27. Witt W, Kohler U. 2007. List J. Partec 2007. Nurnberg. 28. Schulze D. 2008. Powders and bulk solids: Behavior, characterization, storage and ow, 1st ed. Berlin: Springer. 29. Rhodes M. 2008. Introduction to particle technology, 2nd edition. West Sussex: John Wiley & Sons Ltd. 30. Jenike AW. 1964. Storage and ow of solids, Bull. No. 123 of Utah Eng Exp Stn, Salt Lake City, Utah: Univ Utah. 31. Hoskuldsson A. 1988. PLS regression methods. J Chemom 2: 211228. 32. Muteki K, MacGregor J, Ueda T. 2006. Rapid development of new polymer blends: The optimal selection of materials and blend ratios. Ind Eng Chem Res 45:46534660. 33. Martens H, Tormod N. 1991. Multivariate calibration, 1st ed. Chichester, West Sussex, England: John Wiley & Sons Ltd. 34. Burnham JA, MacGregor JF, Viveros R. 1996. Frameworks for latent variable regression. J Chemometrics 10:3145. 35. Muteki K, MacGregor J, Ueda T. 2007. Mixture designs and models for the simultaneous selection of ingredients and their ratios. Chemometrics Intelligent Lab Syst 86:1725.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 100, NO. 1, JANUARY 2011