Aditya U. Vanarase ALL RIGHTS RESERVED

Transcription

1 11 Aditya U. Vanarase ALL RIGHTS RESERVED

2 DESIGN, MODELING AND REAL-TIME MONITORING OF CONTINUOUS POWDER MIXING PROCESSES by ADITYA U. VANARASE A dissertation submitted to the Graduate School New Brunswick Rutgers, The State University of New Jersey in partial fulfillment of the requirements for the degree of Doctor of Philosophy Graduate Program in Chemical and Biochemical Engineering written under the direction of Prof. Fernando J. Muzzio and Prof. Marianthi Ierapetritou And approved by New Brunswick, New Jersey October 11

3 ABSTRACT OF THE DISSERTATION DESIGN, MODELING AND REAL-TIME MONITORING OF CONTINUOUS POWDER MIXING PROCESSES By Aditya U. Vanarase Dissertation Directors Prof. Fernando J. Muzzio and Prof. Marianthi G. Ierapetritou Continuous processing is an advantageous alternative for the current methods used in the pharmaceutical manufacturing. Important advantages that it offers include smaller equipment footprint, reduced efforts in the scale-up work, and the potential to utilize already continuous processes to make the entire manufacturing more efficient. In the current pharmaceutical manufacturing environment, powder mixing process is carried out in the batch mode. The necessary methods and guidelines to design an equivalent continuous process are not well established. The work presented in this dissertation focuses on the characterization, design and optimization of a continuous powder mixing process for pharmaceutical powders. A systematic study was performed of the effects of process and design variables, and material properties involved in the continuous powder mixing process. The bulk powder flow behavior was characterized using the residence time distribution (RTD) measurement approach. Impeller speed, material bulk density and impeller design greatly influenced the mean residence time. With increasing impeller speed, mechanical fluidization was observed, which significantly affected axial dispersion coefficients. Intermediate rotation rates exerted maximum strain on the material, which leads to maximum homogenization. The strain measurements correlated well with the properties ii

4 of tablets including content uniformity and tablet hardness. Mixing performance was largely dominated by the material properties of the mixture, and the blend uniformity measurement was affected by the sample size analyzed. An experimental protocol was developed to measure the blend uniformity in the in-line mode, and a methodology was further built to quantitatively relate the in-line NIR measurements with the off-line wet chemistry measurements. Considering the shear limitations of the continuous bladed mixer, alternative blending strategies, suitable for blending of cohesive materials were also demonstrated. A combination of a high-shear mixing followed by a low-shear mixing process provided the optimal mixing performance. The predictive understanding of the continuous powder mixing process developed in this dissertation can assist towards the design and development of a fully controlled continuous manufacturing process. iii

5 Acknowledgements First of all, I thank my research advisors, Prof. Fernando Muzzio and Prof. Marianthi Ierapetritou for their guidance, continuous motivation and support throughout the course of my PhD. I especially thank Prof. Muzzio for helping me improve my writing and presentation skills. I also thank the Engineering Research Center (ERC) for proving funding for my research and travel, and for providing a great number of opportunities to interact with the folks from industry. I thank Prof. Benjamin Glasser for assessing my PhD proposal, and helping me build my dissertation. I also thank my external committee members Prof Rajesh Dave and Dr. Ralf Weinekötter for becoming referees for my dissertation defense. I extend my sincere gratitude and appreciation to my collaborators Prof. Rodolfo Romañach, Prof. Rohit Ramachandran and Dr. Janne Paaso. Furthermore, a special vote of appreciation to Prof. Rodolfo Romañach and Dr. Janne Paaso for the technical discussions on the NIR spectroscopy and their assistance in chemometric modeling. I thank all my undergraduate students and summer interns, Albert Gasser, Sabin Mathew, Rizwan Aslam and Rocio Arroyave for assisting me in all the experimental work. Many thanks for the support of my fellow graduate researchers, Amit Mehrotra, Patricia Portillo, Marcos Llusa, Bill Engisch, Juan Osorio, Alisa Vasilenko, Fani Boukouvala, Yijie Gao, Matt Metzger, Brenda Remy, and post-docs Atul Dubey, Eric Jayjock, Athanas Koynov, Kalyana Pingali, and Rafael Mendez. Finally I dedicate my PhD to my parents, and my sister Isha who always supported me and encouraged me to get a doctorate degree. Thank you! iv

6 Table of Contents ABSTRACT OF THE DISSERTATION... ii ACKNOWLEDGEMENTS... iv TABLE OF CONTENTS... v LIST OF TABLES... vii LIST OF FIGURES... viii CHAPTER 1 INTRODUCTION Motivation Mathematical modeling in continuous powder mixing Theoretical developments RTD modeling Experimental characterization of continuous powder blending process Experimental studies on the performance and RTDs of in continuous powder mixers PEPT and PIV studies on the bladed mixers Lubricant blending Blending of cohesive powders in continuous mixing systems On-line process monitoring of powder blending processes CHAPTER CHARACTERIZATION OF POWDER FLOW BEHAVIOR IN THE CONTINUOUS MIXER Equipment and experimental set-up Materials and methods Materials..... Methods....3 RTD, Hold-up and number of Blade passes measurement RTD measurement Hold-up measurement Strain measurement RTD modeling methodology Experimental conditions Results Effects of process parameters on flow behavior Effects of design parameters on flow behavior Effects of material properties on flow behavior Predictive model for blend uniformity suitable for control purposes Conclusions Figures for Chapter Tables for Chapter CHAPTER 3 CHARACTERIZATION OF THE POWDER FLOW BEHAVIOR IN THE CONTINUOUS BLENDER USING DEM Methods Simulation set-up The Discrete Element Method (DEM) Results Data Acquisition and processing Mean residence time Number of blade passes Mean centered variance Conclusions Figures for Chapter Tables for Chapter v

7 CHAPTER 4 CHARACTERIZATION OF THE MIXING PERFORMANCE OF THE CONTINUOUS MIXER Methods NIR Spectroscopy LIBS (Laser Induced Breakdown Spectroscopy) Washburn s method Results APAP mixing Lubricant mixing Conclusions Figures for Chapter Tables for Chapter CHAPTER 5 CONTINUOUS MONITORING OF POWDER MIXING PROCESS BY NIR SPECTROSCOPY Chemometric calibration model development using on-line NIR spectral data Equipment and experimental set-up Materials and pre-blend preparation NIR Spectroscopy Results Conclusions Continuous monitoring using VTT Spectrometer Equipment and experimental set-up Methods Results Conclusions Figures for Chapter Tables for Chapter CHAPTER 6 DEVELOPMENT OF INTEGRATED CONTINUOUS MIXING AND DE-LUMPING PROCESS Mixing effects in low shear (Gericke mixer) and high shear mixing (Quadro - Comil) continuous mixing equipment Equipment Results Conclusions Figures for Chapter CHAPTER 7 CONCLUSIONS AND RECOMMENDATIONS Conclusions Recommendations for future work New Blender designs: Development of an integrated feeder-mixer system with a recirculation tank Figures for Chapter REFERENCES 148 CURRICULUM VITA vi

8 LIST OF TABLES Table -1: Feeder configurations used in the experiments Table -: Materials, supplier and particle size Table -3: Experimental conditions Table -4: Bulk density, Carr Index, dilation, particle size of excipients Table -5: Predictive model for Mean Residence Time... 5 Table -6: Predictive models for Axial Dispersion Coefficient... 5 Table 3-1: Impeller blade configurations Table 3-: DEM Simulation Parameters Table 4-1: Analysis of variance (ANOVA) for the NV (Normalized Variance) Table 5-1: Experimental conditions for continuous mixing experiments Table 5-: Development of calibration models and its initial evaluation Table 5-3: Evaluation of model precision (standard deviation) and model accuracy (RMSEP) at each calibration concentration Table 5-4: Evaluation of continuous mixer experiments at various APAP concentrations Table 5-5: Off-line calibration samples Table 5-6: Off-line (UV absorption) data of sample size, mean concentration, variance, RSD and confidence intervals Table 5-7: Model fitting results for in-line and off-line data vii

9 LIST OF FIGURES Figure -1: (a) Experimental set-up (b) Continuous powder mixer (Gericke GCM-5) Figure -: Effect of rotation rate on RTD. Other parameters: Flow rate 3 kg/h, and blade configuration All forward Figure -3: Effect of rotation rate on (a) mean residence time (b) mean centered variance (c) hold-up and (d) number of blade passes. Other parameters: flow rate 3 kg/h, and blade configuration All Forward Figure -4: Effect of flow rate on RTD at (a) 39 RPM (b) 1 RPM (c) 16 RPM and (d) 54 RPM. Other parameters: Blade configuration All Forward Figure -5: Effect of flow rate on (a) mean residence time (b) mean centered variance (c) hold-up and (d) number of blade passes. Other parameters: blade configuration All Forward Figure -6: Effect of flow rate on (a) hold-up and (b) bulk residence time Figure -7: Effect of weir position on hold-up Figure -8: Effect of blade configuration on RTD at (a) 39 RPM (b) 1 RPM (c) 16 RPM and (d) 54 RPM. Other parameters: flow rate: 3 kg/h Figure -9: Effect of blade configuration on (a) mean residence time (b) mean centered variance (c) holdup and (d) number of blade passes. Other parameters: Flow rate 3 kg/h Figure -1: PLS model for Output variable - Mean Residence Time Figure -11: Loading plot for the PLS model of Output variable - Mean Residence Time, and Input variables - Impeller Speed, Flow rate, Bulk density and Cohesion Figure -1: Variable Importance Plot (VIP) of the PLS model of output variable - Mean Residence Time Figure -13: Effect of Bulk density on mean residence time at 3 kg/hr Figure -14: Effect of impeller speed on the number of blade passes for different excipients Figure -15: PLS Model for Output Variable - Axial Dispersion Coefficient Figure -16: Loading plot for the PLS model of output variable Axial Dispersion Coefficient, and input variables impeller speed, flow rate, bulk density and cohesion Figure -17: Variable Importance Plot (VIP) for the PLS model of output variable - Axial Dispersion Coefficient Figure -18: Effect of cohesion on the axial dispersion coefficient... 5 Figure 3-1: Computer aided drawing of a continuous blender made at the actual scale. Two feeders continuously provide particles in two streams on either side of the impeller which rotates in the direction shown by the curved arrow. A D-shaped semicircular weir was placed at the outlet such that its flat edge was at 45 with the horizontal Figure 3-: Blade patterns used in DEM simulations and experimental validation studies. A) Forward blade pattern with two blades shown; B) Alternate pattern with one forward facing and one backward facing blade, both at Figure 3-3: Simulation snapshots at a) 4rpm, b) 1rpm, c) 16rpm and d) 5rpm. The red and blue particles are fed as two parallel streams of same mean particle size with a normal particle size distribution. The particle bed fluidization begins at approximately 16rpm Figure 3-4: Effect of process parameters on mean residence time (DEM Simulations) Figure 3-5: Comparison between experimental and DEM simulation results for the mean residence time.. 65 Figure 3-6: Effect of operational parameters on the number of blade passes (DEM simulations) Figure 3-7: Effect of operational parameters on the number of blade passes (Experimental) Figure 3-8: Effect of operational parameters on the mean centered variance (DEM simulations) Figure 3-9: Comparison between the DEM simulations and experimental results for the mean centered variance (MCV) Figure 4-1: Schematic of the experimental set-up for LIBS Figure 4-: Experimental set-up for Washburn's method Figure 4-3: Comparison between flow rates: (a) VRR vs. rotation rate ( All Forward Blade configuration) (b) RSD vs. rotation rate ( All Forward Blade configuration) (c) VRR vs. rotation rate ( Alternate blade configuration) (d) RSD vs. rotation rate ( Alternate Blade configuration). Comparison between Blade configurations: (e) RSD vs. rotation rate (3 kg/hr), (f) RSD vs. rotation rate (45 kg/hr) (Note: Comparison viii

10 between the blade configurations is shown only with the RSD, plots of VRR are not shown here in order to avoid redundancy) Figure 4-4: Effect of MgSt concentration: (a) RSDNIR vs. Rotation rate (b) RSDLIBS vs. Rotation rate (c) Tablet hardness vs. Rotation rate (d) Hydrophobicity vs. Rotation rate Figure 4-5: (a) Effect of design parameters on RSDNIR (b) Effect if design parameters on hydrophobicity (c) Effect of blade configuration on tablet hardness Figure 4-6: (a) Feeding positions for MgSt (b) Effect of feed position on blend uniformity at the blender discahrge Figure 4-7: Feed position at the blender inlet (a) RSD vs. blender length (a) Mean concentration vs. blender length... 9 Figure 4-8: Feed position - Center of the blender (a) RSD vs. blender length (b) Mean concentration vs. blender length... 9 Figure 5-1: (a) CDI Spectrometer installed on a powder conveying chute at the mixer discharge (b) Chute Figure 5-: NIR spectra for acetaminophen and for % and 15% (w/w) powder blends Figure 5-3: Scores plot from principal component of analysis of calibration set spectra in nm spectral range Figure 5-4: API content predicted by NIR for cross validation and external validation samples Figure 5-5: NIR predictions from monitoring the continuous mixing process for three representative blends Figure 5-6: Schematic of the multipoint NIR measurement equipment, it consists of a fiber-optic light source, fiber-optic probes and a fiber-optic spectral camera Figure 5-7: (a) Schematic picture (left) and photograph (right) of the multipoint fiber-optic light source. It has 4 output fibers with a ST connector (b) The fiber-optic spectral camera (Spectral camera NIR, Specim Ltd., Oulu, Finland) Figure 5-8: Experimental set-up (a) Above-the-chute configuration (b) Below-the-chute configuration 1 Figure 5-9: The unprocessed calibration set spectra (left) and the spectra after the baseline correction (right)... 1 Figure 5-1: The scatter plot of using the PLS model with the calibration set (left) and after cross-validation (right) Figure 5-11: The scatter plot of cross-validation after averaging the results of each sample over the 5 measurement points of probe number 1 (left) and probe number (right) Figure 5-1: Blend uniformity (RSD) as a function of sample size (NIR Spectroscopy)... 1 Figure 5-13: Blend uniformity (RSD) as a function of sample size (UV Absorption)... 1 Figure 5-14: (a) RSD as a function of sample size (Comparison between NIR data and mathematical model) (b) Linear regression for the best case ( RSD ) Figure 5-15: (a) RSD as a function of sample size (Comparison between UV absorption data and mathematical model) (b) Linear regression for the best case ( RSD ) Figure 6-1: (a) Conical mill (Comil - Quadro Model # 197) (b) Milling chamber with conical round impeller Figure 6-: (a) Schematic of the experimental set-up for mixing in Gericke Continuous mixer (b) Effect of impeller speed on blend uniformity (RSD) Figure 6-3: (a) Schematic of the experimental set-up for mixing in Comil (b) Effect of impeller speed and screen size Figure 6-4: Effect of operational parameters of mill on residence time Figure 6-5: (a) Schematic of the experimental set-up for integrated low and high shear mixing (Low-shear mixing first) (b) Mixing performance after low and high shear mixing Figure 6-6: (a) Schematic of the experimental set-up for integrated low and high shear mixing (High-shear mixing first) (b) Mixing performance after high and low shear mixing Figure 7-1: Schematic of the continuous processing line with a recirculation tank Figure 7-: Schematic of an integrated feeder, mixer and recirculation tank system Figure 7-3: (a) Dynamic response of the continuous mixer for feeder refills (b) Mixer response for recycle flow rates 1 and 5 times of input flow rate ix

11 1 Chapter 1 Introduction 1.1 Motivation Continuous processing is considered as an advantageous choice in many industries, including chemicals, food, household goods, microelectronics, and many others. Its main advantages are better controllability, and, for sufficiently large volumes, lower manufacturing cost by decreased footprint and labor. The pharmaceutical industry, however, due to the rigid nature of its regulatory framework, has remained largely focused on conventional batch manufacturing. However, since the inception of the Process Analytical Technologies initiative (PAT [1]), and more recent, the Quality by Design (QbD) initiative [], significant efforts in designing new manufacturing strategies are underway. Continuous manufacturing for solid dose pharmaceutical products is aimed at improving product quality, reducing manufacturing cost, and essentially provide safer products to the patients. Continuous processing for secondary pharmaceutical manufacturing is an attractive option because processes such as tableting, roller compaction, and capsule filling are already carried out in the continuous mode [3], while mixing, wet granulation, drying, and coating are performed in batch mode; this mixture of batch and continuous steps is a frequent source of inefficiencies. Also, continuous processes can be scaled up simply by time extension, as opposed to batch processes, which require size scale up and which often do not scale-up easily or well. Continuous processing offers other advantages over batch mixing, including smaller equipment size, reduced in-process inventory, less solid handling such as filling and emptying of blenders (potentially reducing undesirable

12 effects like segregation), better control around a well-defined steady state, and higher uniformity of shear application. However, continuous processing has some limitations, including higher initial cost, difficult implementation for low volume products, and reduced process flexibility. Although continuous manufacturing has been heavily implemented in the bulk chemical industry, including processes that involve powder handling applications such as catalysts manufacturing [4], mineral processing [5], and food manufacturing [6], the understanding of continuous powder mixing processes is still limited and only a handful of papers have been published on this topic. Continuous mixing is important in many processes in pharmaceutical manufacturing, including some obvious ones such as API and lubricant mixing, and some less apparent, such as wet granulation, coating, extrusion, and drying, where mixing often plays a critical role. Design of continuous mixing operations requires evaluation of a large parametric space, including selection and design of mixing and feeding equipment; evaluation of operating parameters such as impeller rotation rate and flow rate; characterization of the effects of material properties such as particle size distribution and powder cohesion, and controlling environmental variables such as relative humidity and temperature. This large number of variables (and their interactions) makes it difficult to implement the process for a new entity without detailed studies. Thus, identifying critical process parameters is a key step towards effective implementation of continuous manufacturing. The work presented in this dissertation is focused towards implementing continuous secondary manufacturing pharmaceutical processes, with specific concentration on continuous mixing of powders. Although the work presented here is based on a

13 3 pharmaceutical product manufacturing case, the methods used are general and they apply to other industries as well. The dissertation work is focused on performing a systematic evaluation of the process and design parameters related to the continuous mixing process and also the interconnection between different material properties and continuous mixing performance. The approach used is primarily experimental in nature. DEM simulations of the powder flow behavior in the mixer were also performed, and compared against the experimental results. Formation of undesirable agglomerates [7,8] in powder blends is a commonly faced problem in powder mixing processes. The evaluation of continuous mixers for their potential to mitigate these problems possibly through a combination of different processing approaches is also addressed. The four specific aims of this dissertation are as follows: Specific Aim I: Understand the powder flow behavior in the continuous mixer (Convection, shear and dispersion) (Chapter, Chapter 3) Specific Aim II: Understand effects of process and design parameters and material properties on blend homogeneity and blend properties (Chapter 4) Specific Aim III: Develop fast, effective on-line sensing methods for monitoring blend uniformity in continuous powder blending systems (Chapter 5) Specific Aim IV: Characterization and optimization of an integrated process consisting of continuous mixing and de-lumping to achieve desired blend properties (Chapter 6) In the next section, important theoretical and experimental literature is discussed and a conceptual framework of the dissertation will be established.

14 4 1. Mathematical modeling in continuous powder mixing 1..1 Theoretical developments Detailed information on continuous mixing can be found in reviews such as Pernenkil et al. [9] and Williams and Rahman [1]. Only a few important theoretical developments and experimental investigations related to this dissertation are cited here. The performance measure of continuous blenders was first defined by Beaudry [11] using the Variance Reduction Ratio (VRR), defined as the ratio of the input and output variance in concentration of the key component. Thus, the larger the VRR is, the better the performance of the continuous blender. Theoretical developments on continuous blenders began when Danckwerts [1] introduced the concept of residence time distribution (RTD). Using the RTD function, Danckwerts defined the dynamic output concentration as a function of input concentration. The relationship is given in equation (1-1). C ( i d (1-1) t) C ( t ) E( ) With some mathematical manipulations, the final form of VRR was expressed as given in equation (1-). 1 VRR Fluid i o r r R( r) I( r) dr (1-) Where ' ' i ( t i i ( t ) r) R ( r) and I ( r) E( ) E( r) dr.

15 5 The term is the autocorrelation coefficient of for the time interval. Using the autocorrelation, the time scale of segregation for the incoming feed rate variability can be calculated. William and Rahman [13,14] proposed a numerical method for predicting the VRR using the RTD for both ideal and non-ideal blenders. Thus, both Danckwerts and William and Rahamans definitions express the variability in output concentration as a function of input variability. Danckwerts definition for VRR describes the macro-mixing process but does not contain information about the micro-mixing, which is especially important for granular materials. Weinekötter and Reh (1995) [15] modified Danckwerts definition by adding a term to account for the variance of the ideal random powder mixture. Weinekötter s modified definition of VRR is given in equation (1-3). 1 1 VRR solid VRR fluid out, ideal input (1-3) Weinekötter and Reh [15] also used the concept of scale of segregation introduced by Danckwerts [16] to extract structural information about mixtures. By calculating the power density spectrum of the in-line measurements, the scale of segregation of the mixture can be calculated as defined in equation (1-4). In equation (1-4), is the autocorrelation coefficient. I xx ( ) d (1-4) Thus, the RTD along with the knowledge of input variability seems to be sufficient to predict mixing performance.

16 6 1.. RTD modeling Several methods have been proposed for modeling RTDs. For continuous powder blenders, RTDs have been modeled using a fractional tubularity model [17] (which consists of a CSTR in series with a PFR), a delay and a dead volume, tanks in series [17] and dispersion models [5,18]. The axial dispersion model has been one of the most classical approaches to model axial mixing in both continuous and batch mixers or rotating drums. The one-dimensional diffusion equation used to describe mixing in the axial direction is given by equation (1-5). c t c c u z Dz (1-5) z z This equation has been solved for various systems [18] using appropriate initial and boundary conditions. The most common cases concerned with continuous flow systems include Dankwerts closed-closed and open-open boundary conditions [19]. The important approximation made here is that the particle bed in the system was treated to be a continuum. Fokker Planck Equations (FPEs) were introduced for the case of continuous mixers by Sommer [19] and were solved for each component in the mixture. Kehlenbeck and Sommer [] later showed the applicability of FPEs for particle streams with identical physical properties. Other methods used to model continuous mixers are Markov Chain models [1,], population balance models [3] and discrete element models [4,5]. Berthiaux et al. [] proposed Markov chain models to calculate all the parameters necessary to characterize RTD in a continuous blender. Using this approach, they showed that the ratio of mean residence time and period of fluctuation of the inflow streams is the main

17 7 factor governing VRR in a continuous mixer rather than the variance of RTD. Berthiaux [6] again by using the same methodology of Markov chain modeling simulated the RTD in a continuous mixer that was capable of incorporating effects of different flow conditions as well as tracer properties. The methodology of using the RTD to model the continuous powder blending process accounts only for the axial mixing component. This approach is valid for the case of free flowing (cohesionless) powders. For the case of cohesive powders, a minimum required level of shear stress, and a minimum amount of radial mixing are also required. The RTD modeling approach, as presented in the literature is often not sufficient. In this dissertation, the role of radial mixing is identified, and the analytical method error associated with the blend uniformity measurement is also incorporated into the mathematical model. Correlations which relate the RTD parameters (dispersion coefficient, axial velocity) with the input parameters including material properties, flow rate, shear rate and blade pattern/mixer geometry are in their infancy. To fill this gap, relationships between the process parameters and RTD were developed, and the role of critical material properties was also addressed in this dissertation. Two promising future directions of modeling work include DEM modeling of the continuous mixer geometry and predictive/data driven modeling of RTD as a parallel effort with experimental investigations. The DEM models have been used in the past to simulate industrial mixers such as the double cone blender [7-9], the V blender [3-3], the tote blender [33,34], and also continuous mixers [4,5] in some recent publications. Although the DEM simulation is an attractive option to study such systems with flexible choice of geometry and operating parameters, they suffer some important

18 8 limitations. One of the major difficulties in DEM simulations is the huge computing power required to simulate real case scenarios involving > 1 5 particles. Also, identifying the correct parameters to simulate real material properties is a big challenge. DEM simulations have so far used spherical particles, but in reality particles have shapes other than spherical. Despite these limitations, DEM is one of the best available tools to simulate granular flows [3]. In this dissertation DEM simulations of the powder flow behavior in the continuous mixer were performed, and compared with experimental RTD measurements. 1.3 Experimental characterization of continuous powder blending process Blender geometries reported in the literature, relevant to the continuous blender design used in this dissertation include bladed mixers (batch and continuous) and continuous rotary calciners. The experimental approaches used to investigate these geometries, reported in the literature can be classified in two categories. The first approach consists of measurement of the RTD and the variance of the mixture at the outlet. The other approach consists of characterizing the flow behavior in the mixer using radioactive particle tracking methods such as PEPT (Positron Emission Particle Tracking), or imaging techniques such as the PIV (Particle Image Velocimetry) Experimental studies on the performance and RTDs of in continuous powder mixers Studies of continuous mixing for practical applications have been reported for materials such as zeolite pellets [4], foods (semolina and couscous, chocolate mixtures) [35], Aluminum hydroxide and Irgalite [16], sand and salt [13], etc. William and Rahman [14]

19 9 measured RTDs in a continuous mixer for free flowing granular materials salt and sand. Their proposed model for predicting VRR matched the experimental results very well. Harwood et al. [36] examined seven different continuous mixers and categorized them according to their performance for both cohesive and free-flowing powders. Weinekotter and Reh [15] measured RTDs in a continuous mixer using Al(OH) 3 as a bulk material and SiC or Irgalite as tracers. The main parameter examined in their studies was the effect of feeder fluctuations and the type of mixer. The ability of the continuous mixer to reduce input fluctuations was dependent on both the Peclet number (Pe) and the time period of fluctuation of the feed rate. The continuous mixer was found to be acting as a low pass filter, filtering only the high frequency noise. Sudah et al. [4] conducted RTD experiments in a rotary calciner using zeolite pellets to study the effect of rotation rate, flow rate and angle of inclination on mean residence time, hold-up and axial dispersion coefficient. A dispersion model was used to fit the experimental data to extract the dispersion coefficient and the axial velocity. Mean residence time was related to the angle of inclination and rotation rate but was independent of the flow rate up to 1% fill level. The axial dispersion coefficient also was a function of the speed and the angle of inclination (at high rotation rates), but was weakly dependent on the flow rate. Ziegler and Aguilar [17] studied RTDs in the continuous processing of chocolate in a twin-screw co-rotating mixer and modeled them using a combination of PFR and CSTR (fractional tubularity), and CSTR with delay model. The fractional tubularity model was found to fit the data better than the CSTR with delay model. The mean residence time was found to be directly proportional to the rotation rate and inversely proportional to the feed rate. Sherritt et al. [37] reviewed a significant amount of experimental data for rotary calciners

20 1 and proposed a new design equation for the axial dispersion coefficient in both batch and continuous systems in terms of the rotation rate, fill level, drum diameter and particle diameter. Additionally they examined rotation rate and fill level for a horizontal rotary drum using radioactive tracer material. The axial dispersion coefficient that was measured followed their design equation. However, this correlation takes into account only the particle diameter as the material property. Other material properties such as cohesion and particle density were not included. Also, this correlation applies only for rotary calciners. Generalized correlations for bladed mixers are not well established in the literature. Experimental studies for pharmaceutical powders are recent and include materials such as lactose and acetaminophen mixtures [38], and a mixture involving nine ingredients including three different actives [39]. The important process parameters examined were rotation rate, flow rate, and angel of inclination of the mixer or rotary drum. Portillo et al. [38] reported that lower rotation rates and an upward inclination resulted in better mixing performance. Marikh et al. [4] examined the effect of rotation rate and flow rate on hold-up in a continuous powder mixer using semolina and couscous material. They developed a correlation between mean residence time and rotation rate which could be used as a basis for scale-up criteria. Marikh et al. [35] compared two different types of stirrers for a pharmaceutical mixture, and proposed a more generalized correlation for hold-up as a function of rotation rate and flow rate. They reported that the mixing performance becomes better with increase in the rotation rate. However, at excessively high rotation rates, mixing performance become worse. They also reported the choice of one stirrer over another in producing better quality of mixture. Pernenkil et al. [41]

21 11 examined the Zigzag blender for caffeine and lactose mixture. Operational parameters examined were the shell rotation rate and the intensifier bar rotation rate. They reported the effect of the intensifier bar rotation rate on VRR to be the highest, while the shell rotation rate showed a non-linear effect. Experimental studies for pharmaceutical powders reported so far have been specific to a particular material and the mixer studied. Very few studies have attempted to explain the physics behind the results PEPT and PIV studies on the bladed mixers Several studies are reported in the literature on bladed mixers used for the granular mixing applications such as wet granulation and agitated drying. These mixers are similar to the equipment used in the present case of the continuous blender. Recent experimental studies using PEPT have provided deeper insights into the detailed flow patterns in powder mixers and rotating drums. Stewart et al. [4] compared experimental measurements performed using PEPT and DEM simulations of a granular flow in a vertical bladed mixer. They were able to simulate the overall motion of the bed very well. However they indicated a need for the proper selection of material properties in DEM for a realistic simulation. Laurent et al. [43] performed experiments in a horizontal batch mixer using glass ballotini. They utilized PEPT to track a single particle. Using the PEPT measurements they calculated the RMS velocity and the dispersion coefficient at various positions in the mixer. They also studied the effect of different sizes of tracer particles. The size of tracer particles has negligible effect on the radial flow structure, but the axial motion was significantly affected. The design of the agitator in both continuous and batch mixers is largely based on empiricism or, at best, follows heuristic rules. Laurent and Bridgwater [44] utilized the

22 1 PEPT technique to compare two impeller designs. A design with short paddles showed increase in the dispersion coefficient with the increase in the fill level, as opposed to the decrease observed for the case with long flat blade impeller. Both impellers showed a cellular structure created by radial supports but the short paddle device showed more chaotic behavior. Conway et al. [45] used PIV to measure velocity fields on free exposed surfaces (top and the wall). They characterized the flow and segregation behavior at low as well as high shear rates. The studies that are reported in the literature so far are focused on the operating parameters or the design of mixer. Little attention has been given to the effect of material properties. In a real industrial environment, manufacturing processes needs to be tuned according to the raw material properties. Thus, developing a methodology for selecting the optimum process parameters for a given material property is addressed in this dissertation Lubricant blending One of the important and widely used materials in the pharmaceutical industry is Magnesium Stearate (MgSt) [46]. This material has not been reported so far in the literature in the context of continuous mixing. Lubricant blending is a required step in continuous as well as batch processing scenario, and it needs to be characterized in continuous powder blending systems. A brief literature review on lubricant mixing and its importance in pharmaceutical industry is provided below. Lubricant mixing is a very important and necessary step in pharmaceutical manufacturing. Lubricants are added to the formulation for four primary purposes: 1- It facilitates ejection of the tablet from the die, - It provides internal failure points for the

23 13 tablet during compression, reducing accumulation of stored elastic stress, 3- It acts as an anti-adherent [47] reducing the sticking between the tablet and the punch, and 4. It improves the flow properties of powders [48]. Although they facilitate the manufacturing processes, lubricants introduce some complexities in the product. Lubricants affect the compaction process (Force Displacement curve) as well as post-compact properties including porosity [49] and hardness [5]. They have also been shown to affect tablet disintegration [51,5] and dissolution rate [53], which often are directly related to the bioavailability of the drug. Podczeck and Miah [54] conducted experiments to study the effect of particle size and shape on the angle of friction and the flow function for both lubricated and un-lubricated blends. Yamatoma et al. [51] conducted experiments to relate the disintegration time to the bulk density of the formulation, which is influenced by the presence of lubricant. Van der Watt and de Villiers [55] studied the scale-up of V- mixer by relating the mixer size with crushing hardness of tablets. Ragnarsson et al. [56] related mixing time to the tablet strength. All of the studies mentioned here so far are specific to the system in place and their results do not extend to different materials or mixers. Mehrotra et al. [57] identified the two important parameters, shear rate and strain, irrespective of the system and correlated them with properties of tablets and powder. They conducted experiments in a modified Couette shear cell exposing powder to a controlled and uniform shear environment. Strain was found to be the parameter maximally affecting blend homogeneity, bulk density, flowability of powders and tablet hardness. In this dissertation, measurement of strain (in terms of blade passes) conducted for a continuous blender as a function of different operating conditions was used to establish a

24 14 link between the operating conditions of the blender and the resulting blend/tablet properties Blending of cohesive powders in continuous mixing systems APIs (Active Pharmaceutical Ingredients) which often have an average particle size of less than µm in diameter are highly cohesive, and tend to form agglomerates because of overwhelming Van der Waals forces. Cohesion in powder/granular systems is also caused by the capillary forces [58] which exist due to the presence of moisture in the system. In the present case of dry powders the primary source of cohesion are Van der Waals forces and/or electrostatic forces [59]. Powders which tend to agglomerate are typically mixed under high shear environment such that the agglomerates are de-lumped and an ordered mixture is sometimes created. High shear blending is typically performed using a V-blender with a high-speed intensifier bar or by introducing a mill [8]. In this study, a co-mill was selected as a suitable continuous de-lumping device. In order to qualify the feasibility of a co-mill for the continuous process, its efficiency for mixing and de-lumping needs to be examined. In this dissertation, the effect of co-mill process parameters on mixing performance and its position in the overall continuous processing scheme is examined. 1.4 On-line process monitoring of powder blending processes A typical continuous process for pharmaceutical manufacturing consists of unit operations such as powder feeding, blending, granulation and compaction or capsule filling. In a continuous process, monitoring and ensuring blend uniformity in real time at the blender discharge point are highly critical. Unlike batch manufacturing, it is difficult

25 15 to rework the blending process in a continuous manufacturing scenario. Continuous monitoring of blend uniformity is the first step towards implementing process control for continuous blending operations, or to facilitate rejection of non-uniform powder from the blending operation. Typically in the pharmaceutical industry, following a blending operation, a batch of powder passes inspection only if the variability in the sample concentrations is under a specified limit. This limit is defined in the blend uniformity guidance by FDA [6] as the relative standard deviation between the samples concentrations extracted in the end of a blending process to be under 6%. PAT (Process Analytical Technology) [1] has recently been introduced by FDA as a tool for building predictive understanding of pharmaceutical manufacturing processes. Examples of powder PAT applied to blending processes are abundant in literature; they typically include Near Infra-Red NIR spectroscopy [61], Raman spectroscopy [61,6] and Laser Induced Fluorescence LIF [63]. Most of the PAT work for blend uniformity monitoring exists for batch blending, which includes commonly used blenders such as V-blender [64], Bin blender [63], Y- mixer [64] and Nauta mixer [65]. Examples of PAT applied to continuous blending processes are scare in the literature. Near infrared (NIR) spectroscopy is one of the most commonly used analytical techniques for monitoring pharmaceutical processes [66]. The NIR spectral region extends from 78 to 5 nm, and NIR spectra consist of absorbance bands corresponding to overtones and combinations of fundamental C H, N H, S H and O H molecular vibrations. NIR methods have been developed to monitor a number of pharmaceutical unit operations including granulation [67], drying [68] and crystallization

26 16 [69]. The NIR spectra after following an adequate pre-processing provide both the physical as well as chemical signature of the material. Some studies include quantitative methods for the determination of drug concentration during the blending process [7][71,7], while other studies include qualitative approaches to evaluate the end-point as the changes in the spectra are reduced during the blending process [73]. A series of studies [74-76], describing the use of NIR spectroscopy as a PAT tool in the design and implementation of a blending process are available in the literature. The use of NIR spectra to develop a control chart based on Hotelling's T statistic was also demonstrated [77]. Methods reported in the literature to monitor blend uniformity are mostly generic, which include quantitative methods such as a PLS (Partial Least Squares) modeling, or qualitative methods such as Principal Component Analysis (PCA) of the spectra acquired during blending process, monitoring the pooled standard deviation between spectra, monitoring the dissimilarity between the process spectra and spectrum of an uniform mixture or individual components. PAT for blending has been reported primarily as a tool for monitoring evolution of RSD during the blending process and detect the blending end point. The final blend uniformity measured using a PAT method, and the blend uniformity measured using an off-line method based on wet-chemistry are often poorly correlated, rather the methods to directly link off-line and on-line blend uniformity measurements are not very well established in the literature. In order to develop a relationship between the off-line and on-line measurements, it is necessary to quantify the error in the measurement method and the sample size being analyzed, and relate that with equivalent offline measurements.

27 17 Sekulic et al. [78] demonstrated a methodology to estimate the effective sample mass being analyzed by a fiber optic probe in a blending process, by comparing the spectral variance of fiber optic measurements against a priory relationship developed between spectral variance and sample mass using offline micro-sample cups. Very few studies have considered this aspect while applying PAT to the blending processes. PAT applications for continuous blending include some recent studies using NIR spectroscopy [79] and LIF [71,78]. In the continuous blending process, typically powder is in a state of motion, and inherently there is always a certain degree of spectral averaging involved in the measurement. Blend uniformity, quantified as the RSD (Relative Standard Deviation) between the in-line measurements, is dependent on the degree of averaging. It is necessary to select the averaging window depending on the sample size of interest (one unit dose), which is often not the case in the current industrial practice. In this dissertation a methodology is presented using a case study based on pharmaceutical powders, allowing quantification of the error associated with the in-line measurements, as well as the relationship between in-line and off-line blend uniformity measurements. The methodology was developed using a multi-point fiber optic probe NIR monitoring system.

28 18 Chapter Characterization of powder flow behavior in the continuous mixer Powder mixing processes are primarily governed by intrinsic powder flow mechanisms such as convection, shear and dispersion [8]. RTDs have been used extensively in the past to characterize macroscopic flow behavior in non-ideal reactors for liquid systems [18]. In this chapter, RTDs in a continuous powder mixer were measured to characterize the bulk powder flow behavior. Effect of process variables (impeller speed, flow rate), design variables (impeller design, mixer outlet (weir) design) and material properties (bulk density, particle size, and cohesion) on the RTD were examined. RTDs were measured experimentally by providing a tracer impulse and measuring the concentration of the tracer at the mixer discharge. The effects of aforementioned variables were compared using the statistical RTD parameters (mean residence time and mean centered variance) and strain exerted on the powder. RTDs were also characterized computationally using DEM simulations of the powder flow behavior in the continuous mixer. The comparison between experimental and computational measurement of RTDs is presented in Chapter 3. RTDs can also be used for predicting the axial mixing behavior of the system. A reduced order model for the powder flow behavior in the continuous mixer was developed by applying 1-D Fokker Plank Equations (FPEs). In this chapter, the RTD data was fitted using an analytical solution of the FPE solved for appropriate boundary conditions for the case of continuous powder mixer. The methodology for model development and data-fitting is also described in this chapter.

29 19.1 Equipment and experimental set-up The experimental set-up is shown in Figure -1. A commercial continuous mixer manufactured by Gericke was examined. The design of the mixer is illustrated in Figure -1(b). The continuous mixer (Model GCM-5) is.3 m long and.1 m in diameter. The impeller consists of 1 triangular shaped blades, equally spaced along the axis of rotation. The first and the last blades of the impeller are designed differently due to their position close to the end walls of the mixer vessel. The angle of the blade with the shaft can be changed. In the present study, blade angles both in the forward and backward directions were examined. Forward direction means that the blade imposes a forward flow on the powder along the axis of the mixer; backward direction implies flow in the reverse direction. The mixer is equipped with a weir (Figure -1 (b)), which is a semicircular disc placed at the exit of the mixer to control powder hold-up. The weir can be rotated to change the fill level of the powder in the mixer. Additional details are shown in Figure -1 (a). Loss-In-Weight (LIW) feeders manufactured by Schenck AccuRate were used in the experiments. Performance of LIW feeders, which could affect the blend homogeneity at the mixer discharge, is dependent on the type tooling used. In the present study, the size of the discharge opening, screw design, and hopper design were selected based on manufacturer's recommendation. The particular settings used for feeding Acetaminophen (APAP) and Avicel are presented in Table -1.

30 . Materials and methods..1 Materials The materials used in this study, their mean particle size and the supplier are listed in Table -. Excipients, including micro-crystalline cellulose (Avicel PH-, Avicel PH- 11), Fast Flo Lactose and Dicalcium Phosphate were used as model materials to study the effect of material properties on the powder flow behavior in the mixer. Acetaminophen was used as a tracer for measuring RTDs for micro-crystalline cellulose and lactose, whereas Caffeine was used as a tracer for Dicalcium phosphate... Methods Chemical composition of the powder samples was analyzed using NIR spectroscopy. Chemometric calibration models were built for individual tracer-excipient combinations. A general protocol for building calibration models involves following steps. Calibration samples with known concentration of tracer are prepared by accurately massing the individual components. Calibration samples are then scanned multiple times to acquire representative spectra. The NIR analyzer Antaris (Thermo Fisher) was used in this study. Ominc software was used to acquire the calibration spectra, and TQ Analyst was used to build chemo-metric models. A PLS algorithm was then applied to build the calibration curve. Each calibration model was validated using a cross-validation program (leave one spectra out at a time). More information on calibration model development using NIR is presented in Chapter 5.

31 1.3 RTD, Hold-up and number of Blade passes measurement.3.1 RTD measurement Residence time distributions in the mixer were measured by the impulse response method. Initially, the bulk material (Avicel) was fed in the mixer until steady state flow was reached. Tracer (APAP or Caffeine) was inserted manually in the inflow stream as an instantaneous pulse. Approximately 11 g of the tracer material was used in the experiments. The amount of tracer was selected such that the concentration of APAP at the exit of the mixer would be over several orders of magnitude above the detection limits of NIR method, in order to resolve the long tail of the residence time distribution. Samples were subsequently collected at various times from the outlet of the mixer. The samples collected were analyzed by NIR spectroscopy to determine the concentration of tracer in them. Thus, for each experiment, a dataset of concentration vs. time was collected. Using this data, the residence time distribution function ( E ( t)), the mean residence time ( ) and the mean centered variance ( ) were calculated. As established in prior literature [81] these parameters completely describe axial mixing in a continuous flow system. Mathematical relationships for each of these terms are given below. Residence Time Distribution (RTD) ( E ( t)) : E( t) c( t) c( t) dt (-1) Mean Residence Time (MRT) ( ) : te( t) dt (-)

32 Mean Centered Variance (MCV) ( ) : ( t ) E( t) dt (-3) RTD parameters, mean residence time (MRT) and mean centered variance (MCV) were used to quantify the flow behavior in the continuous mixer. For a perfectly mixed system, which is often described as a perfectly mixed CSTR (Continuously Stirred Tank Reactor), the residence time distribution function E (t) is given by e t /. For such a system, the MCV is equal to one. For a perfectly unmixed system, also described as a perfect PFR (Plug Flow Reactor) E (t) is given by a Dirac delta function which is zero everywhere except at t. At this condition the MCV is equal to zero. Flow behavior of real cases is expected to be in between these two extreme conditions. Thus, a higher value of MCV indicates better mixing condition. Since the RTD arises from velocity differentials in the axial direction, MCV calculated here refers to the mixing in axial direction. Mixing behavior in the continuous blender can be described as a combination of axial and radial mixing. Axial mixing is important in order to mitigate the variability introduced by the feeding process. Radial mixing is necessary to mix the initially unmixed ingredients to the required degree of homogeneity..3. Hold-up measurement Powder hold-up in the mixer is important because it determines the average residence time, and thus the total average strain experienced by the powder as it travels through the mixer. Hold-up was measured by simultaneously monitoring the weight of the powder

33 3 collected at the outlet of the mixer and that of the powder being fed. Powder was collected in a collection bucket resting on a scale at the exit of the mixer. The weight of the powder being fed was monitored by the built-in scale present in loss-in-weight feeders. The difference between the two weights (at the outlet and at the inlet) at a given time gives the hold-up. In preliminary measurements, hold-up is initially zero; it increases with time, and finally reaches a plateau. The mixer operating under constant hold-up was considered to be operating at steady state. In the absence of stagnant regions, hold-up measurement is complementary to the measurement of mean residence time calculated from the RTD curve. As mentioned, hold-up measurements can be used to calculate the bulk residence time [hold-up (kg) / flow rate (kg/h)] of the powder in the mixer..3.3 Strain measurement In the continuous mixer, energy input is provided by rotating the impeller. This energy is dissipated in the convective transport of the powder, random fluctuations of the velocity of the particles (granular temperature), friction between the impeller and the powder and the mean strain (velocity gradients in the powder). Since the impeller rotation rate has an effect on the residence time (which will be explained in the results section), the strain is proportional to the product of the shear rate and the residence time, which in turn is proportional to the number of blade passes in the mixer. Using the residence time, the number of blade passes at various experimental conditions was calculated as follows: Number of blade passes = Rotation rate (RPM) Residence time (s)/6.

34 4.4 RTD modeling methodology Reduced order model of the continuous mixing process was developed by applying the 1- D FPE. One dimensional diffusion equation that describes mixing in axial direction is given by equation -4. c t U c z D z z c (-4) For the present case, following boundary conditions were used to solve the PDE. Danckwerts open-open vessel boundary condition [18] is given in equation -5 and -6. At the entrance: D z c(, t) c z z c( Uc(, t), t) D z c z z Uc(, t) (-5) At the exit: D c( L z, t) c z z L c( L Uc( L, t), t) D z c z z L Uc( L, t) (-6) After applying the boundary conditions, the analytical solution of the above PDE is given by equation -7. (1, ) t t C exp / Pe, D z L. Pe (1 ) 4 / Pe (-7) In equation (-7), is the normalized time, t where is the dead time of the system and is the mean residence time. All the RTD datasets were fitted using equation (-7). The fitting parameters in equation (-7) are mean residence time ( ), dead time t ), pulse (

35 5 strength ( C ) and the Peclet number (Pe). A non-linear regression program from MATLAB was used for the data fitting exercise. The methodology used for model fitting is can also be found in Gao et al. [8]. Once the parameters are estimated, the total residence time ( Total t ), and axial dispersion coefficient ( D z ) can be calculated..5 Experimental conditions In the first set of experiments, the effect of process parameters and mixer design parameters was examined using Avicel PH- as the model material. The parametric space for the continuous mixer consists of manipulated (independent) process parameters (impeller rotation rate and flow rate) and manipulated (independent) design parameters (blade configuration, angle of weir). The experimental conditions examined are presented in Table Results.6.1 Effects of process parameters on flow behavior The following sections focus on the effect of rotation rate and flow rate on the powder flow behavior in the continuous mixer Effect of rotation rate on flow behavior Figure - shows the effect of rotation rate on the residence time distribution. As the rotation rate increases, the mean residence time (τ) decreases (Figure -3 (a)). The mean centered variance was found to increase with increasing impeller rotation rate (Figure -3 (b)). Inverting the mean centered variance gives the equivalent number of ideal stirred tanks in series that would give a similar RTD. For the mixer of interest here, the number

36 6 of ideal stirred tanks decreased from about 6 to 3 as the rotation rate increased. The number of stirred tanks in series indicates the degree of dispersion of the input pulse; fewer tanks indicate faster dispersion, which also means faster axial mixing. Thus, increasing rotation rate gives rise to better axial mixing. However, faster rotation rate also decreases the time available for mixing (lower residence time) and increases the variability in the residence time distribution (less uniform shear treatment). These are opposing effects in achieving good mixing. In fact, for the present case, at 54 RPM, mean residence time was found to be extremely low. Therefore, the 54 RPM rotation rate may not be the best operating condition even though the degree of dispersion is the highest. Hold-up in the mixer (Figure -3(c)) decreases from.57 to.4 kg as the rotation rate increases from 39 to 54 RPM (Fr =.7 3.3). For rotation rates of 16 and 54 RPM, the powder bed was fluidized in the mixer. In order to estimate the amount of strain the powder experiences in the mixer, the number of blade passes was calculated as a function of rotation rate. The relationship between these variables is depicted in Figure -3 (d). The number of blade passes was found to be at maximum at the intermediate rotation rates, entirely due to the significant change in hold up at higher speeds. This indicates that between 1 and 16 RPM, the powder undergoes maximum amount of mechanical work (total strain). It is important to estimate total strain applied during processing since it may affect powder flow behavior, bulk density of the powder and also the blend uniformity [57]. All these variables are important attributes of powder blends Effect of flow rate on flow behavior Flow rate is a key process parameter that is directly related to the capacity of the manufacturing system. Although it was stated that continuous processes can be scaled up

37 7 simply by time extension, in certain cases where higher or lower production rates are required, throughput also needs to be changed. Residence time distributions under different flow rate and rotation rates are compared in Figure -4. RTD curves seem to overlap with each other as the flow rate increases. Figure -5 shows the effect of flow rate on the RTD parameters, hold-up and the number of blade passes. The effect of flow rate on the mean residence time is influenced also by the impeller rotation rate. As shown in Figure -5 (a), at low rotation rates, the increase in the flow rate decreases the mean residence time. With increasing rotation rate, this effect diminishes. For the two flow rates examined (3 and 45 kg/h), mean centered variance (Figure -5 (b)) did not show any particular trend. This indicates that in the range of 3 45 kg/h, similar degree of axial mixing was obtained. Figure -5 (c) shows the effect of flow rate on hold-up for our experiments. Increasing the flow rate increases hold-up. Since the hold-up depends on mixer capacity, at different rotation rates, the dependence of bulk residence time on flow rate varies. To clarify the effect of flow rate on residence time, hold-up was measured for a wider range of flow rates (5 6 kg/h) at three rotation rates (39, 16 and 54 RPM). The bulk residence time was also calculated from the hold-up measurements. The results are presented in Figure -6 (a - b). Remarkably, bulk residence time is not affected by the flow rate at very high rotation rates (54 RPM), which means that under these conditions, mixing performance (such as it might be) is independent of throughput. However, residence time decreases with increasing flow rate at lower rotation rates (39 RPM and 16 RPM). Observations from Figure -5 (a - c) and Figure -6 reveal the relationship between the process parameters and mixer capacity. At the highest rotation rate, hold-up

38 8 increases linearly with increase in flow rate. As the impeller rotation rate decreases, mixer capacity becomes the limiting parameter, which leads to the non-linear behavior of hold-up as a function of flow rate. Similar explanation follows for the trends of residence time or number of blade passes. At higher rotation rates, the number of blade passes is not affected by the change in flow rate; at lower rotation rates, increase in flow rate decreases the number of blade passes (Figure -5(d)). Our results can be compared with studies from the literature. Sudah et al. [4] showed similar results for rotary calciners; mean residence time was found to be independent of feed rates up-to 1% fill levels. In the present case, at high rotation rates (54 RPM), fill level is well below the capacity of the mixer, which yields similar results. For a continuous mixer with a similar impeller design, Berthiaux et al. [6] has shown that the mean residence time increases with increases in flow rate. The differences can be attributed to the different size of the mixer and also to differences in material properties. A design criterion for scale-up could perhaps be developed from these observations, recognizing that hold-up should increase with increase in flow rate, but that this effect interacts with rotation rate and mixer size..6. Effects of design parameters on flow behavior The following sections focus on the effect of weir angle and blade configuration on the powder flow behavior in the continuous mixer Effect of weir angle on flow behavior The weir angle is an additional design parameter that can be used to change the residence time of the powder in the mixer. Powder hold-up in the continuous mixer was measured

39 9 for different weir angles. Figure -7 shows the effect of the weir angle on the hold-up. In the interest of clarity, the results are presented only for the Alternate blade configuration at 39 RPM; trends are similar for other blade patterns and speeds. While the impeller is rotating in the mixer and the blender is operating, the powder bed fill level exhibits a gradient around the axis of the mixer. Rotating the weir to the same angle at which powder bed is tilted gives the maximum hold-up. At 39 RPM, hold up was found to be the maximum at a weir angle of. At high rotation rates (1 RPM and 16 RPM) and for other blade configurations, a rotated weir also achieved the maximum hold-up. However, the angle of rotation of the weir corresponding to maximum hold-up varied between and 45 for different cases. At the highest rotation rate (54 RPM) the powder bed is at least partially fluidized in the mixer, and the tilted powder bed was not directly observable, but again a rotated weir showed the highest hold-up in the mixer. In further experimental investigations to study the effect of other parameters, the weir position was kept constant at..6.. Effect of blade configuration on flow behavior Two blade configurations (Table -3), All Forward and Alternate were examined to determine whether mixing performance could be enhanced. Figure -8 shows the residence time distribution functions for the two blade configurations at different rotation rates. All the results presented in Figure -8 and Figure -9 corresponds to the 3 kg/h flow rate. At high rotation rates (16 RPM and 54 RPM), the mean residence time (Figure -9(a)) is greater for the Alternate blade configuration than for the All Forward configuration. However, at low rotation rates (39 RPM and 1 RPM), this effect diminishes. This indicates that the mixer capacity again becomes the limiting

40 3 parameter at lower rotation rate, leading to the observed behavior. Interestingly, hold-up was found to be higher for the Alternate blade configuration than the All Forward configuration at all the rotation rates (Figure -9(c)), indicating higher bulk residence time. This suggests that Alternate blade configuration creates recirculating zones in the mixer. The mean centered variance (MCV) showed a complex dependence between the two blade configurations. As shown in Figure -9 (b), the Alternate blade configuration showed a higher MCV value at 54 RPM than the All Forward configuration. At lower rotation rates (39 16 RPM), MCV was not distinguishable between the two blade configurations. Although the designs of the two blade configurations were significantly different, a similar degree of dispersion was obtained. Only the rotation rate was found to affect the axial dispersion. At this point, more work is necessary to study the effect of blade configuration on RTD. As shown in Figure -9 (d), the number of blade passes was higher for the Alternate blade configuration than for the All Forward configuration, indicating that a higher amount of total shear is being applied by the Alternate blade configuration..6.3 Effects of material properties on flow behavior RTDs in continuous mixers are dependent on the mixer geometry, impeller speed, flow rate and material properties of the powder. In the previous sections relationships between RTD and process (flow rate, impeller speed) and design parameters (impeller design, weir angle) are reported. In this case the original parametric space of process parameters (flow rate, impeller speed) was extended to include different materials. The materials

41 31 used in this DoE are presented in Table -4. Essentially two bulk material properties (bulk density and cohesion) were correlated with the observed RTD parameters Material properties Material flow properties were characterized using two methods, namely Carr Index (C.I.) and dilation Carr Index Carr Index is an indicator of the compressibility of the powder; higher the C.I. is, more compressible is the powder. C.I. was calculated using equation (-8) by measuring the bulk and tapped densities of powder. C.I. measurements are listed in Table -4. C 1 1 B T Dilation Dilation is a complementary measurement to the C.I. which is also an indicator of the compressibility of the powder. Dilation was measured using the GDR (Gravitational Displacement Rheometer [83, 84]). Dilation is calculated as the percent change in volume of the powder bed as function of time. Details on the experimental set-up of dilation measurement can be found in Faquih et al. [84]. The procedure for dilation measurement can be briefly described as follows. Powder was filled upto 4% volume in a transparent acrylic cylinder, the cylinder was tapped on a tapping machine such that powder is compacted to its tapped density. The cylinder was then mounted on the GDR, which rotates the cylinder such that powder tumbles inside the cylinder. The tumbling or avalanching motion of the powder was monitored in the cross-sectional direction using a

42 3 camera. The images captured by the camera were subsequently analyzed using an image analysis program to measure the change in the cross sectional area of the powder. By making an assumption of uniform cross sectional area across the length of the cylinder, powder dilation can be measured. Dilation measurement, although complementary to the C.I., is more accurate; because it is measured under fully dilated state of the powder. Common errors involved in the bulk density measurement, such as powder compaction while filling the cylinder or inaccurate volume measurement due to uneven powder surface can be avoided in the dilation measurement. Dilation is an indicator of the level of cohesion present in the powder; greater the dilation, more cohesive the powder. This fact has been proven through experimental as well as computational studies [84]. The correlation between the Flow Index (F.I.) measured by the GDR, and dilation is highly linear, and it has been demonstrated that the F.I. is proportional to cohesion when consolidation state of the powder approaches to zero [85]. Thus in our experimental studies dilation was used as the material property that represent cohesion in the system. Dilation values of the materials used our study are listed in Table Effect of material properties on mean residence time In the previous sections ( and.6.1.) it was shown that residence time is a strong function of impeller speed. Flow rate also affects residence time, but it shows an interaction with the impeller speed. At lower impeller speeds, higher flow rates show lower residence times, whereas at higher impeller speeds effect of flow rate is negligible. In order to quantify the effects of process variables and material properties, a PLS analysis was performed on the dataset consisting of an output variable mean residence time and four input variables impeller speed, flow rate, bulk density and cohesion. PLS

43 33 analysis was performed using multi-variate data analysis software - ProSensus Multivariate. PLS analysis showed that 85% of the total variance was captured using just the first principal component (Figure -1). Addition of another principal component did not increase the R value of the model. The relative trends between all the input and output variables are presented using a loading plot (Figure -11). Residence time was found to increase with decrease in impeller speed, decrease flow rate, increase in bulk density and decrease in cohesion. The relative importance of input variables is shown using a VIP (Variable Importance Plot) in Figure -1. Impeller speed and bulk density were found to be the two most important variables that affect mean residence time, whereas flow rate and cohesion were relatively less important. Bulk density was found to be the key material property that affected mean residence time; cohesion did not show a clear trend. The relationship between bulk density and residence time is shown separately in Figure -13. Increases in the bulk density lead to an increase in the mass hold-up in the continuous mixer which increases residence time. This relationship is a strong function of impeller speed. At lower impeller speeds (4, 1 RPM) where the flow regime in the mixer is quasi-static, bulk density has a strong influence. At higher speeds (16,5 RPM) while powder is fluidized in the mixer, bulk density has minimal effect on the residence time. At higher speeds, the forces exerted by the impeller exceed the gravitational settling forces. It is shown in Chapter 4 that the number of blade passes greatly influences the mixing performance. As shown previously the number of blade passes goes through a maximum as the impeller speed increases. The optimum speed that provides the maximum number of blade passes was found to be dependent on the relationship between residence time

44 34 and impeller speed. It was observed that for powders with higher bulk density, the optimal speed was smaller. As shown in Figure -14, for Avicel PH-11 and Avicel PH- the optimal speed is ~ 16 RPM, whereas for Fast Flo Lactose and Dicalcium Phosphate the optimal speed is ~ 1 RPM. This result can be used as a guideline to select impeller speed based on the bulk density of the material Effect of material properties on axial dispersion coefficient The second important fitted parameter in the RTD model (Equation (-7)) is the axial dispersion coefficient. In the previous section, the effect of process parameters (flow rate and impeller speed) on the axial dispersion coefficient was studied for the case of Avicel PH-. It was shown that the axial dispersion coefficient increases with increase in speed.however, flow rate did not show any clear effect. In this case study, another PLS model was developed considering axial dispersion coefficient as the output variable and the same set of input parameters (impeller speed, flow rate, bulk density and cohesion). As shown in the bar plot (Figure -15), only 5% of the total variance was captured using this model. The relative trends between input and output variables are shown in Figure -16. Axial dispersion coefficient was found to increase with increase in speed, increase in cohesion, decrease in bulk density and increase in flow rate. The relative importance of input variables was analyzed by using a VIP plot. As shown in the VIP plot (Figure -17), impeller speed and cohesion were found to be the most important variables affecting axial dispersion coefficient, and bulk density and flow rate had the least influence. In this DoE, bulk density did not show a clear effect on the axial dispersion coefficient, because variation in bulk density was primarily driven by the variation in true density (Dicalcium Phosphate =.31 g/cm3, Lactose =1.5 g/cm3, MCC= 1.6 g/cm3).

45 35 Since true density has the least significance on the axial dispersion coefficient, the effect of bulk density is not clearly seen. If the materials have similar true densities, bulk density and cohesion are typically inversely related. The relationship between dilation and axial dispersion coefficient is shown in Figure -18. Speed and cohesion were found to interact significantly. At lower speeds (4,1 RPM), cohesion did not affect the axial dispersion coefficient; whereas at higher speeds (16, 5 RPM), higher cohesion lead to higher values of the axial dispersion coefficient. At higher speeds, the movement of powder is in the form of agglomerates as opposed to individual particles, which essentially increases the axial dispersion coefficient..6.4 Predictive model for blend uniformity suitable for control purposes In our previous publication [8], a methodology was developed to predict blend uniformity (RSD) as a function of incoming feed rate variability and process parameters. As shown in equation (-9), the total variance ( ) in concentration observed at the total,solids mixer discharge can be decomposed as the variance due to incomplete axial mixing ( ( fluid ) and the variance due to the non-ideal behavior of the powder mixing process Ideal _ feed ). total, solid fluid ideal _ feed (-9) The variance due to the non-ideal powder mixing process can be due to several sources including incomplete transverse mixing, sample size, segregation, agglomeration, or measurement method errors. Usually, the second component of variance is empirical and needs to be characterized individually for each mixture. However the variability due to incomplete axial mixing can be computed by predicting the concentration of the output

46 36 stream as a function of incoming concentrations and process parameters. The mathematical equation proposed by Danckwerts is provided in equation (-1). Cout, fluid ( t) Cin( t ) E( ) d (-1) In order to use equation (-1) as a predictive model for control purposes, a mathematical model for the RTD, and a separate empirical model for RTD parameters as a function of process parameters is necessary. As described in the section.4, RTDs were modeled using FPEs, and RTD parameters (mean residence time, axial dispersion coefficient) were estimated. Empirical relationships were developed between the RTD parameters and the impeller speed. The model parameters and the coefficient of correlation for different pharmaceutical excipients are listed in Table -5and Table -6. As shown in Table -6, a linear model worked reasonably well for predicting mean residence time as a function of impeller speed; and an exponential model showed the best possible fit for predicting the axial dispersion coefficient (Table -6). These models are suitable for constant flow rate conditions, in real situations the dynamics of flow rate also need to be incorporated. A promising way to develop a dynamic predictive model that includes both the process parameters (flow rate and impeller speed) is to use an approach similar to Response Surface Modeling (RSM). In the present case, due to lack of sufficient data, the feasibility of RSM was not investigated..6.5 Conclusions The mean residence time decreases with increase in the rotation rate, but the degree of dispersion increases. Intermediate rotation rates exert the maximum number of blade passes on the powder, thus maximizing strain and homogenization.

47 37 Increasing the flow rate also decreases the mean residence time, but this effect diminishes with increases in rotation rate. At the highest rotation rate (54 RPM), flow rate did not affect mean residence time. This result indicates that high RPMs could be suitable for high flow rates since they do not lead to significant decrease in density (which was observed for low flow rates). Out of the two blade configurations examined, the Alternate blade configuration showed greater powder hold-up than the All Forward blade configuration. The degree of dispersion (MCV) did not show any particular trend between the two blade configurations since it was also dependent on rotation rate and flow rate. Bulk density was found to be the key material property that affects mean residence time; cohesion did not show a clear effect. The effect of cohesion was not seen because the true densities of the materials included in the DoE were significantly different. The effect of cohesion needs to be studied using materials of nearly equal true densities, either by varying the particle size or adding minor ingredients that affect cohesion. The optimal speed, which exhibits the highest number of blade passes was found to be lower for materials with greater bulk densities. This result provides a design rule for selecting impeller speed based on the material bulk density. The axial dispersion coefficient was affected by cohesion (or dilation), and not by the bulk density of the material. Cohesion and impeller speed showed a significant interaction. At lower impeller speeds, cohesion hardly affected the axial dispersion coefficient, however at higher impeller speeds, greater cohesion lead to higher axial dispersion coefficients. At lower impeller speeds, while the flow conditions in the mixer are quasi-static, convection is the main transport mechanism and the dispersive transport

48 38 due to particle-particle interactions is relatively less significant. At higher impeller speeds, while the powder is fluidized in the mixer, particles travel in the form of agglomerates. An increase in cohesion leads to increase in agglomeration behavior in the powder, which essentially affects the axial dispersion coefficient.

49 E(t) (sec-1) 39.7 Figures for Chapter Figure -1: (a) Experimental set-up (b) Continuous powder mixer (Gericke GCM- 5) Figure -: Effect of rotation rate on RTD. Other parameters: Flow rate 3 kg/h, and blade configuration All forward Time (sec) 39 RPM 1 RPM 16 RPM 54 RPM

50 τ (sec) 4 Hold-up (kg) # Blade passes Figure -3: Effect of rotation rate on (a) mean residence time (b) mean centered variance (c) hold-up and (d) number of blade passes. Other parameters: flow rate 3 kg/h, and blade configuration All Forward Rotation rate (RPM) (a) σ τ N= Rotation rate (RPM) N=3 (b).7 (c) 1 (d) Rotation rate (RPM) Rotation rate (RPM)

51 E(t) (sec-1) E(t) (sec-1) E(t) (sec-1) E(t) (sec-1) 41 Figure -4: Effect of flow rate on RTD at (a) 39 RPM (b) 1 RPM (c) 16 RPM and (d) 54 RPM. Other parameters: Blade configuration All Forward Time (sec) 3 Kg/hr 45 kg/hr (a) Time (sec) 3 kg/hr 45 kg/hr (b) (c) (d) 1 Time (sec) Time (sec) 3 kg/hr 45 kg/hr 3 kg/hr 45 kg/hr

52 Hold-up (kg) # Blade passes τ (sec) 4 Figure -5: Effect of flow rate on (a) mean residence time (b) mean centered variance (c) hold-up and (d) number of blade passes. Other parameters: blade configuration All Forward Rotation rate (RPM) (a) σ τ Rotation rate (RPM) (b) 3 kg/hr 45 kg/hr 3 kg/hr 45 kg/hr.7 (c) 1 (d) Rotation rate (RPM) 1 3 Rotation rate (RPM) 3 kg/hr 45 kg/hr 3 kg/hr 45 kg/hr

53 Hold-up (kg) Hold-up (kg) Bulk res. time (sec) 43 Figure -6: Effect of flow rate on (a) hold-up and (b) bulk residence time Flow rate (kg/hr) (a) Flow rate (kg/hr) (b) 39 RPM 16 RPM 54 RPM 39 RPM 16 RPM 54 RPM Figure -7: Effect of weir position on hold-up Deg Deg 45 Deg No Weir Weir position

54 E(t) (sec-1) E(t) (sec-1) E(t) (sec -1 ) E(t) (sec -1 ) 44 Figure -8: Effect of blade configuration on RTD at (a) 39 RPM (b) 1 RPM (c) 16 RPM and (d) 54 RPM. Other parameters: flow rate: 3 kg/h Time (sec) (a) Time (sec) (b) All Forward Alternate All Forward Alternate.9 (c).9 (d) Time (sec) Time(sec) All Forward Alternate All Forward Alternate

55 Hold-up (kg) # Blade passes τ (sec) 45 Figure -9: Effect of blade configuration on (a) mean residence time (b) mean centered variance (c) hold-up and (d) number of blade passes. Other parameters: Flow rate 3 kg/h Rotation rate (RPM) (a) σ τ Rotation rate (RPM) (b) All Forward Alternate All Forward Alternate.8 (c) 14 (d) Rotation rate (RPM) Rotation rate (RPM) All Forward Alternate All Forward Alternate

56 46 Figure -1: PLS model for Output variable - Mean Residence Time Figure -11: Loading plot for the PLS model of Output variable - Mean Residence Time, and Input variables - Impeller Speed, Flow rate, Bulk density and Cohesion

57 Mean residene time (sec) 47 Figure -1: Variable Importance Plot (VIP) of the PLS model of output variable - Mean Residence Time Figure -13: Effect of Bulk density on mean residence time at 3 kg/hr y = x R² =.9786 y = x R² = y = 7.837x R² =.448 y = 1.594x R² = Bulk density (g/cc) 4 RPM, 3 kg/hr 1 RPM, 3 kg/hr 16 RPM, 3 kg/hr 5 RPM, 3 kg/hr Linear (4 RPM, 3 kg/hr) Linear (16 RPM, 3 kg/hr) Linear (1 RPM, 3 kg/hr) Linear (5 RPM, 3 kg/hr)

58 # Blade passes 48 Figure -14: Effect of impeller speed on the number of blade passes for different excipients Impeller Speed (RPM) Avicel PH11 Fast Flo Lactose Avicel PH Dicalcium Phosphate Figure -15: PLS Model for Output Variable - Axial Dispersion Coefficient

59 49 Figure -16: Loading plot for the PLS model of output variable Axial Dispersion Coefficient, and input variables impeller speed, flow rate, bulk density and cohesion Figure -17: Variable Importance Plot (VIP) for the PLS model of output variable - Axial Dispersion Coefficient

60 (Dz)Axial dispersion coefficient (cm /sec) 5 Figure -18: Effect of cohesion on the axial dispersion coefficient % Dilation 4 RPM, 45 kg/hr 1 RPM, 45 kg/hr 16 RPM, 45 kg/hr 5 RPM, 45 kg/hr

61 51.8 Tables for Chapter Table -1: Feeder configurations used in the experiments Material Feeder Flow rates Nozzle-size (mm) Screw type Avicel PH- Schenck 5-6 kg/hr Auger AccuRate APAP Schenck AccuRate kg/hr 36 Helix Table -: Materials, supplier and particle size Material Supplier Particle size (µm) Avicel PH- FMC Biopolymer 9 Avicel PH-11 FMC Biopolymer 34.1 Fast Flo Lactose Foremost Farms USA 1.1 Dicalcium Phosphate TLC Ingredients 186. Acetaminophen (APAP) Mallinckrodt 3 (micronized) Caffeine TLC Ingredients 36 Table -3: Experimental conditions Process parameters Flow rate Rotation rate Mixer design parameters Blade configuration name: Blade direction, blade angle (Angle with the shaft) Weir angle 3, 45 kg/hr 39 RPM, 1 RPM, 16 RPM, 54 RPM 1. All Forward All blades directing forward, blade angle deg. Alternate Alternate blades directing in forward and backward, blade angle deg deg Table -4: Bulk density, Carr Index, dilation, particle size of excipients Material Bulk Density (g/cm 3 ) Carr Index (C.I.) % Dilation d5 (µm) Avicel Avicel Fast Flo Lactose Dicalcium Phosphate (CaHPO 4 )

62 5 Table -5: Predictive model for Mean Residence Time Flow rate (kg/hr) Material Model Coeff. of Correlation (R ) 3 Avicel PH N Avicel PH N Fast Flo Lactose -.77N Dicalcium Phosphate -.336N Avicel PH N Avicel PH N Fast Flo Lactose -.47N Dicalcium Phosphate -.33N Table -6: Predictive models for Axial Dispersion Coefficient Flow rate (kg/hr) Material Model Coeff. of Correlation (R ) 3 Avicel PH N. 444e Avicel PH-. 9. N.8e Fast Flo Lactose N. 871e Dicalcium Phosphate. 8. N. 415e Avicel PH N. 314e Avicel PH N 1.3e.9 45 Fast Flo Lactose N. 77e Dicalcium Phosphate. 4. N. 667e.97

63 53 Chapter 3 Characterization of the powder flow behavior in the continuous blender using DEM This chapter presents a study of the powder flow behavior in a continuous mixer using DEM modeling. The aim is to establish and employ a predictive model and validate it against experiments. Once a relationship between a real case scenario and DEM simulations is established, DEM can be used as to optimize and design continuous powder mixing processes. The comparison between simulations and experiments was made using RTD as the main response. 3.1 Methods Simulation set-up Computer simulations of the solids mixing process were performed using DEM, a method by Cundall and Strack [86]. For maximum accuracy, computer aided drawings of the blender with 1:1 size ratio were created using Pro/Engineer software. The drawings were imported into a commercial DEM-based simulation program called EDEM (a commercial software package marketed by DEM Solutions Inc) (Figure 3-1). At the inlet of the blender two feeders were providing a continuous supply of particles on either side of the impeller at a uniform feed rate. The two streams were completely segregated from each other. At the other end of the blender, a semicircular weir was placed in order to increase the hold-up and facilitate back mixing. The weir was placed such that its straight edge made a 45 angle with the horizontal. The impeller rotation was counterclockwise when viewed along the axis of rotation from the outlet end. Two impeller blade patterns were designed using the blade angles shown in Table 3-1. The first pattern has twelve

64 54 blades with all forward facing (in the direction of flow) angles except for the last blade, and the second pattern has twelve blades with alternating forward-reverse angles. The impeller blade patterns are shown in Figure 3-. A full factorial experimental design consisting of two feed rates (of 3 kg/h and 45 kg/h), four impeller speeds (4rpm, 1rpm, 16rpm and 5rpm) and two blade configurations was examined in this study. The blender was fed with particles continuously in the form of two streams on either side of the blade at the inlet. Table 3- shows the simulation parameters used in this study The Discrete Element Method (DEM) DEM is a technique for simulating the behavior of granular materials with each particle treated as a discrete unit as opposed to continuum models where the material is treated as a featureless medium with smooth properties defined by continuous functions. In the DEM method, the motion of each particle is tracked based on the calculated positions and velocities, which are a result of the forces present in the system. Forces on particles are of two types contact forces and body forces (equation (3-1)). The contact forces are due to inter-particle or particle-boundary collisions. The boundary can be any physical object in the system, such as walls, impellers, and baffles. Forces are resolved into normal and tangential components that are independent of each other. F F F (3-1) Total contact body In equation (3-1) is the resultant force on a given particle due to its interactions with other particles and/or boundary elements as well as due to the effect of external force fields such as the gravitational field, cohesive or electrostatic interactions. The term accounts for all the normal and tangential contact forces and denotes

65 55 the sum of all body forces. This resultant force is computed for each particle at a high frequency (a time-step being typically of the order of 1 µsec.) and the new particle position is computed by numerically solving the equations of motion. Using Newton s law the position of a particle i that has j number of contacts with its surroundings is related to the resultant force by equation (3-). m x I i i i i j j F R n ij i F F t ij t ij l k F body body (3-) In equation (3-), is the mass of the particle of radius, is its position, its acceleration, and is its moment of intertia. and are the normal and tangential components of the contact force on the particle due to its th contact respectively. The term accounts for all body forces acting on the particle using a summation index k. In this study, gravity is considered to be the only body force acting on the particles (k = 1, ). The rotational components of motion are the angular displacement θ, angular acceleration, and sum of all torques due to body forces using summation index. The contact forces in DEM are calculated using a suitable contact model. There are several types of contact models available for use in DEM simulations [87-89]. They can be broadly classified into the following categories - molecular dynamics like potentialbased contact models [9] and the more commonly used linear viscoelastic [91], nonlinear viscoelastic [9,93] and hysteretic [94-96]contact models. This study uses a contact model introduced by Tsuji et.al.[93] which is based on Hertzian contact theory [86,87,96,97]. This model utilizes a soft particle approach. The distance between the centers of each pair of particles or particle-boundary is computed at every time step. A

66 56 contact is detected if the distance between the centers of particles (in case of a particleparticle contact) is less than sum of the particle radii or the distance between a boundary and the center of a particle (in case of a particle-boundary contact) is less than the particle s radius. A very small overlap is allowed in each of the normal and tangential direction. Normal Forces The normal force due to a contact that resulted in a normal overlap is given by: F n k n 3/ n n ' n 1/ 4 n (3-3) In the above equation, refers to the normal stiffness coefficient, is the normal damping coefficient, and is the rate of deformation. The normal stiffness coefficient is obtained by equation k n ( R 3(1 eqv ) E ) (3-4) In the above equation, E is the particle s Young s modulus and is the material Poisson ratio. is defined as the effective radius of the contacting particles. If two contacting particles have radii and the effective radius is obtained by R eqv R R R i i j R j (3-5) In case of a particle-boundary collision, the Poisson ratio of the boundary and the Elastic modulus of the boundary are also required. For a particle of radius colliding with the boundary, the stiffness coefficient is then calculated as:

67 57 k n (1 4 R i ) E (1 3 b ) E b (3-6) With the knowledge of the normal stiffness coefficient and a chosen coefficient of restitution e, the normal damping coefficient is calculated as: mk n n ln e (3-7) ln e Tangential Forces The tangential force is calculated in a similar fashion as its normal counterpart. The tangential contact force also consists of elastic and damping components. When a tangential overlap of has been detected and there is a corresponding normal overlap of due to the same contact, the tangential force is calculated as: t t t t ' t 1/ 4 n F k k (3-8) In the above equation, is the tangential stiffness coefficient, and is the tangential damping coefficient. The tangential stiffness coefficient is calculated [98] by: R eqv G 1/ k n n (3-9) where G is the particle s Shear modulus. It is related to the elastic modulus E as: G E (1 ) (3-1) The tangential displacement (or overlap) is calculated by time-integrating the relative velocity of tangential impact, between two colliding entities (interparticle or particlewall contact): t dt v t rel (3-11) between two entities having velocities and is calculated by resolving the absolute relative velocity [ using, the tangential component of the unit vector

68 58 connecting the centers of the colliding particles (or center of a particle and its contact point with the geometry for particle-geometry contact) and adding the effect of angular velocities as: v t rel [ v v ]. nˆ R R (3-1) i j t i i j j The tangential force is limited by the Coulomb condition, which states that the tangential force should be less than the normal force scaled by the coefficient of static friction as. If, in the simulation, the tangential force obtained from equation (3-8) exceeds the Coulomb limit for any pair-interaction, the slip is accounted for by resetting the tangential displacement is to. 3. Results 3..1 Data Acquisition and processing The continuous blending simulation was performed by rotating the impeller at a constant speed and feeding two streams of particles at the inlet, simulating two feeders. The blender was allowed to fill up until the mass hold-up reaches a near-constant value, indicating that a steady state is achieved. The simulation was then run further for another 5-3 seconds at the steady state to allow for a time-window for data collection. Figure 3-3 shows the simulation snapshots at 4, 1, 16 and 5rpm while the blender is operating under steady state. An increasing fluidization was observed in the 16 to 5rpm range. The RTDs were computed using the following methodology: The particles were continuously created at the inlet and two equal streams (meaning same mean particle size, size distribution and feed rate). After the steady stage was achieved, the particles that

69 59 were created at in a 1-second window were tagged. Those particles were tracked until they crossed the weir at the outlet of the mixer. The time taken by each tagged particle to cross the weir was recorded as the residence time of that particle. The simulations were run at steady state for long enough time such that at least 95% of the tagged particles were retrieved at the outlet. A histogram was then created using 1-second interval bins. Thus a curve of frequency (number of particles) vs. time was obtained, which was used further to calculate the RTD, and the necessary RTD parameters. 3.. Mean residence time Computational results showing the effect of operational parameters on the mean residence time are presented in Figure 3-4. A good qualitative agreement was achieved between the experimental results (Chapter ) and DEM simulations. The relative effects of impeller speed, flow rate, blade configuration and their respective interactions were also captured reasonably well in the simulations. However, the quantitative comparison between the two sets of results showed some interesting differences. Experimental measurements of the mean residence times were higher than those computed from DEM simulations in the lower range of residence times. An opposite behavior was observed under higher residence times (Figure 3-5) where experimental values were lower than the simulations. These differences can be attributed to the differences between the material properties of real powders, and the particle properties assigned in DEM simulations. Lower values of residence times (- 4sec, Experimental) belong to the impeller speed of 5 RPM. At such a high speed, fluidization of particles is created in the blender. Under fluidization, powder hold-up in the continuous blender significantly decreases. Although the total flow rates are kept equal in the experimental studies and the simulations, the

70 6 number of particles present in each system differs significantly. The mean particle size used in DEM simulations ( mm) is much greater than the particle size of the powder used in experiments (Avicel PH-, d5~µm). The difference in the particle size leads to much greater differences in the number of particles present between two systems (1 9 orders of magnitude). Under fluidization conditions, while the hold-up is very low, in the simulation case, the number of particles present in the blender is extremely low. In that case individual particles might be getting impacted by the impellers (rather than groups of particles, which is usually the case in real situations). While the particles are impacted individually by the impeller, they experience much greater forces, which lead to lower residence times than observed experimentally. Despite of all these differences between experiments and simulations, DEM seems to a good qualitative tool to characterize the relative effects of operating and design parameters Number of blade passes As described in Chapter 4, the number of blade passes, which is proportional to the total strain applied on the material during blending, correlates well with the observed mixing performance (RSD), hydrophobicity of the blend and tablet hardness. Using the residence time measurements from DEM simulations, the number of blade passes was calculated. The results from simulations and experiments are shown in the figures Figure 3-6 and Figure 3-7 respectively. In this case also, qualitatively, the relative effects of operating and design parameters are captured well in DEM simulations. However in DEM simulations, the maximum number of blade passes was observed to be at 1 RPM, while in experiments it was at 16 RPM. The differences between the two cases are a direct result of the residence time measurements described in the previous section. The nature

71 61 of the relationship between residence time and impeller speed is relatively linear in experiments, while in the simulations two distinct regions can be identified Mean centered variance The effect of operational parameters on the MCV, captured from the DEM simulations is shown in Figure 3-8. The impeller speed showed a strong effect on the MCV. Except for the Alternate blade and 45 kg/h case, the MCV increased with increase in speed and exhibited plateau between 16-5 RPM. At lower impeller speeds (4, 1 RPM), when the particles are not fluidized, the impeller speed seems to be the only important variable; effects of flow rate and blade configuration are minimal. Under fluidization (16, and 5 RPM), except for the Alternate blade and 45 kg/h case, the MCV shows a slight decrease with increase in speed. In this range, other variables show non-monotonic effects. These results indicate that the MCV depends primarily on the flow condition in the mixer (Fluidization or quasi-static flow). In the experimental studies, powder is not completely fluidized at 16 RPM. The transition regime for fluidization lies between 16-5 RPM. This essentially leads to an increase in the MCV with increase in impeller speed until 5 RPM (Figure -3). In simulations, while the fluidization occurs early between 1-16 RPM, a plateau in the MCV values was observed between 16 5 RPM. A comparison between the experimental and simulation MCV values is shown in Figure 3-9. As a result of the differences in the onset of fluidization, a poor match (R =.43) was obtained between the experimental and the simulation results.

72 6 3.3 Conclusions A good qualitative agreement was achieved between the experimental and simulation results. The relative trends between the operational (flow rate, speed) and design variables were faithfully captured in DEM. A different fluidization behavior was observed in DEM simulations compared to experimental results. The onset of fluidization in real case scenarios was observed between 16-5 RPM, whereas in DEM simulations it was observed between 1-16 RPM. The primary reason for this behavior seems to be the presence of fewer particles present in DEM simulations than the real case scenarios. The differences in fluidization behavior lead to differences in the optimal operating zone between the two cases. In DEM simulations, at 1 RPM, the maximum number of blade passes was observed as opposed to 16 RPM in experimental cases. The mean residence time in experimental cases was found to be higher than the simulations under higher impeller speeds, while it was lower than simulations under lower impeller speeds. At lower speeds, while the particles are not fluidized in the mixer, tend to slip between the particle-particle and particle-wall contacts. This behavior leads to higher residence times than observed in the simulations.

73 Figures for Chapter 3 Figure 3-1: Computer aided drawing of a continuous blender made at the actual scale. Two feeders continuously provide particles in two streams on either side of the impeller which rotates in the direction shown by the curved arrow. A D-shaped semicircular weir was placed at the outlet such that its flat edge was at 45 with the horizontal. Figure 3-: Blade patterns used in DEM simulations and experimental validation studies. A) Forward blade pattern with two blades shown; B) Alternate pattern with one forward facing and one backward facing blade, both at. (a) (b)

74 64 Figure 3-3: Simulation snapshots at a) 4rpm, b) 1rpm, c) 16rpm and d) 5rpm. The red and blue particles are fed as two parallel streams of same mean particle size with a normal particle size distribution. The particle bed fluidization begins at approximately 16rpm. (a) (b) (c) (d)

75 DEM Simulation Mean Residence time (s) 65 Figure 3-4: Effect of process parameters on mean residence time (DEM Simulations) Impeller Speed (RPM) All Forward 3 kg/hr All Forward 45 kg/hr Alternate 3 kg/hr Alternate 45 kg/hr Figure 3-5: Comparison between experimental and DEM simulation results for the mean residence time Experimental All Forward 3 kg/hr All Forward 45 kg/hr Alternate 3 kg/hr Alternate 45 kg/hr

76 Number of blade passes Number of blade passes 66 Figure 3-6: Effect of operational parameters on the number of blade passes (DEM simulations) Impeller Speed (RPM) All Forward 3 kg/hr All Forward 45 kg/hr Alternate 3 kg/hr Alternate 45 kg/hr Figure 3-7: Effect of operational parameters on the number of blade passes (Experimental) Impeller Speed (RPM) All Forward 3 kg/hr All Forward 45 kg/hr Alternate 3 kg/hr Alternate 45 kg/hr

77 MCV (DEM Simulations) Mean Centered Variance (-) 67 Figure 3-8: Effect of operational parameters on the mean centered variance (DEM simulations) Impeller Speed (RPM) All Forward, 3 kg/hr All Forward, 45 kg/hr Alternate, 3 kg/hr Alternate, 45 kg/hr Figure 3-9: Comparison between the DEM simulations and experimental results for the mean centered variance (MCV) MCV (Experimental) All Forward, 3 kg/hr All Forward, 45 kg/hr Linear Regression Alternate, 3 kg/hr Alternate, 45 kg/hr

78 Tables for Chapter 3 Table 3-1: Impeller blade configurations Impeller blade configuration Name Blade Angles (deg.) Pattern-1 (11 Forward Blades) 5,,,,,,,,,,, -3 Pattern- (4 Alternate blades) 5,,, -,,,, -,, -,, - 3 Table 3-: DEM Simulation Parameters Particle Properties Shear Modulus: e+6 N/m Poisson s Ratio:.5 Density: 15 Kg/m 3 Diameter: mm Normal Size distribution with S.D. =. (Truncated at lower limit of 7% and a higher limit of 13%) Particle-Particle Interactions Coefficient of Static Friction :.5 Coefficient of Rolling friction :.1 Coefficient of Restitution:.1 Blender Walls Blades Material: Glass Shear Modulus: 6 GPa Density: Kg/m 3 Poisson s Ratio:.5 Material: Steel Shear Modulus: 8 GPa Density: 78 Kg/m 3 Poisson s Ratio:.9 Particle-Blade Interactions Coefficient of Static friction:.5 Coefficient of Rolling friction:.1 Coefficient of Restitution:. Particle-Wall Interactions Coefficient of Static friction:.5 Coefficient of Rolling friction:.1 Coefficient of Restitution:.1

79 69 Chapter 4 Characterization of the mixing performance of the continuous mixer Chapter was focused on the characterization of macroscopic flow behavior in the continuous mixer using RTD. However, RTD does not provide the complete description of the system. Micro-scale properties of mixtures, including the scale of segregation of the mixture and blend homogeneity, are not directly captured in the RTD. This chapter will focus on the characterization of the powder blend homogeneity as a function of process and design parameters of the continuous mixer. Two cases, involving blending studies for APAP and MgSt are presented. A third case study of blending of highly cohesive powders, which involves mixing and de-lumping, is presented in Chapter 6. In lubricant blending, along with the blend homogeneity, other blend properties such as hydrophobicity and post-blending flow properties, and tablet properties such as tablet hardness and dissolution profile, are important variables. The hydrophobicity of the blend is affected by the total strain applied in blending, shear rate and lubricant concentration [99]. Over-lubrication often leads to reduced tablet hardness [1] and poor dissolution [11,1]. In this case study of lubricant blending, effects of process and design parameters of the continuous blending process on the blend uniformity, hydrophobicity, content uniformity in tablets and tablet hardness were examined. 4.1 Methods NIR Spectroscopy A Nicolet Antaris NIR spectrometer (Thermo Fisher) was used to quantify APAP in the samples. Spectral data was collected using the software Omnic and TQ Analyst was

80 7 used for calibration model development. The instrument measures the spectrum in the range of 4 cm 1 to 1, cm 1 wave numbers. The regression method used was partial least square (PLS). A Norris derivative filter was used to treat the spectral data. The coefficient of correlation (R ) for the model obtained was.9831 and root mean squared error of prediction (RMSEP) was.39, which indicates good fitting of the spectral data LIBS (Laser Induced Breakdown Spectroscopy) A schematic for the experimental set-up for LIBS is shown in Figure 4-1. LIBS is based on the principle of atomic emission spectroscopy of laser induced plasma. LIBS utilizes a highly energized pulsed laser beam which is focused on a very small area of the sample. In a single shot of the laser pulse, a tiny amount of material is ablated, which generates plasma at very high temperatures. At such a temperature, material is broken down into excited atomic or ionic species. Such excited states decay by emitting radiation in UV, visible and NIR region. By resolving the white light by an optical spectrometer, a spectrum consisting of atomic and molecular bands is obtained. In recent years, LIBS has been applied for quantitative analysis of pharmaceutical products for measuring content uniformity [13-15] as well as coating thickness uniformity [9,15]. Tablets used in this study contain APAP, MgSt and micro-crystalline cellulose. Three spectral lines corresponding to Mg were identified. The peak height at the wavelength nm was used to quantify MgSt. Spectral data was collected at 13 sites and 11 shots per site in each tablet. The intensity between different sites and shots was averaged for each tablet. Uniformity of MgSt distribution can be characterized as the RSD between different sites in each tablet or the

81 71 RSD between average intensities for different tablets. In the present experimental work, inter-tablet variability was calculated by computing the RSD between average intensities measured for tablets at each experimental condition. Micro-distribution of MgSt has been shown to affect bulk density of the powder, tablet strength etc. LIBS signal intensity is known to be affected by the matrix changes [16]. The effect of matrix changes as a function of process parameters on LIBS signal intensity was also measured Washburn s method According to Washburn theory [17] when a porous solid is brought into contact with a liquid the rise of the liquid into the pores of the solid obeys the following relationship: T C cos M (4-1) The terms are defined as follows: T = time after contact, η = viscosity of liquid, C = material constant characteristic of solid sample, ρ = density of liquid, γ = surface tension of liquid, θ = contact angle, M = mass of liquid adsorbed on solid If an experiment is performed where the mass of the adsorbed liquid is measured with time, provided that the powder is uniformly packed and does not dissolve during the experiment, a graph of Time vs. Mass should yield a straight line whose slope is (η / C ρ γ cosθ), is defined as the hydrophobicity. As the slope increases, hydrophobicity increases, in other words cosθ decreases and contact angle increases. The experimental set-up is as shown in Figure 4-. This method measures the amount of a water-based solution that permeates through a powder bed with respect to time. Experiments are conducted as follows:

82 7 1. The powder bed is packed in a chromatographic column (the bottom is made of sintered glass), which is tapped to obtain a consistent density throughout the column.. The bottom of the column is immersed into the solution, which rises through the powder bed by the action of capillary forces. As the powder becomes hydrophobic, the penetration by the solution slows down or stops altogether. 3. The weight of the column increases as the solution penetrates the powder bed. The weigh is monitored and frequently recorded. 4. The curves obtained should reproduce accurately. 4. Results 4..1 APAP mixing To determine the homogeneity of the stream coming out of the mixer, samples were retrieved from the outlet stream. The relative standard deviation (RSD, Equation (4-), which is also known as the coefficient of variability and is the most common mixing index used in industry, was computed. For each experimental run, samples were collected from the outlet. Scintillation glass vials were used to store and analyze the samples. Concentration of acetaminophen in each sample was measured using a NIR spectroscopy analytical method. RSD between the acetaminophen concentrations was calculated using the usual relationship.

83 73 RSD s C N 1 ( C i N 1 C ) (4-) In equation (4-), C is the average concentration of the total samples (N) collected in each mixing run and C i is the concentration of each sample; s is the estimate of the standard deviation obtained using the sample concentrations. When RSD is smaller, concentrations of the individual samples are closer to the mean concentration, which indicates better blend uniformity. In continuous blending systems, blend uniformity is often represented in the form of VRR (Variance Reduction Ratio). The term VRR, as given by equation (4-3), was introduced by Danckwerts [1], and is a ratio of the variance in concentration at the mixer inlet to the variance in concentration at the outlet. VRR represents the efficiency of the mixer to reduce incoming fluctuations in the feed rate. VRR s s i o (4-3) In equation (4-3), s i is the variance in concentration at the inlet and s o is the variance at the outlet. A smaller value of s o (or higher value of VRR) indicates better mixing performance. In the present case s i changes only when the total input flow rate is changed. For each of the mixing runs, both indices RSD and VRR were calculated. The first set of experiments were performed using a low API dose formulation (3% APAP, 97% Avicel PH-). In order to determine the statistical significance of each parameter for blend homogeneity, analysis of variance (ANOVA) was performed. In

84 74 doing this, we are treating the variance as an averageable response, which in a practical sense, it is. An alternative method would be to do pairwise comparisons using an F-test (or similar) but this approach is impractical for multiple levels because proper definition of an overall α level for the entire data set is unclear. A full factorial design of three parameters, rotation rate, flow rate and blade configuration was used. The response used in the analysis was the normalized variance (NV), NV = RSD. The results from ANOVA are presented in Table 4-1. Rotation rate was the most statistically significant variable affecting mixing performance (p = ). Following that, the effect of blade configuration was statistically significant (p = ). Flow rate was found to be statistically not significant (p =.85) variable. The interaction between Rotation rate and blade configuration was also statistically significant (p = ). At 54 RPM, the All Forward blade configuration shows higher NV than the Alternate blade configuration; at 39 RPM, NV nearly coincides for both of the blade configurations Effect of impeller rotation rate The rotation rate is one of the important process variables. In a continuous process, for a given throughput capacity, the rotation rate is the only manipulated variable that can be easily changed online to control the blend homogeneity. Four rotation rates (39, 1, 16 and 54 RPM) were examined for the All Forward blade configuration. As depicted in Figure 4-3, the highest VRR (Figure 4-3 (a)) and smallest RSD (Figure 4-3 (b)) were observed at intermediate rotation rates (1 16 RPM). For Alternate blade configuration as well, intermediate rotation rates (1 RPM 16 RPM) exhibited lowest values of the RSDs (Figure 4-3 (d)). The effect of rotation rate was found to be statistically significant (p = ). At intermediate rotation rates, the maximum

85 75 number of blade passes (maximum strain) is exerted on the powder, which leads to better mixing performance. The effect of rotation rate on the number of blade passes is shown in Figure 4-3 (d) Effect of flow rate The other critical operational variable is the flow rate through the system. During operation, this variable is typically determined by the capacity of other process components (for example, the tablet press), thus it is critical during process design to determine that the mixer is properly sized to achieve optimum performance at the intended flow rate. As shown in Figure 4-3 (d), for the Alternate blade configuration, mixing performance was better at the lower flow rate (3 kg/h). The difference in RSD between the two flow rates was not statistically significant at very high and low rotation rates. Such relative differences between RSDs are somewhat made clear by the measurement of strain. For the All Forward condition, both flow rates show similar mixing performance at higher rotation rates (16 RPM and 54 RPM) (Figure 4-3 (b)). At low rotation rates (39 RPM and 1 RPM), experiments conducted at higher flow rate exhibited lower RSD (Figure 4-3 (b)). However, over the entire range of RPM, the effect of flow rate was statistically not significant with p =.85. This result is actually quite useful, indicating that for the range of flow rates studied here, mixing performance is robust with respect to flow rate. However, increase in total flow rate has an impact on fill level in the mixer and also on the input variability, because feeders operate more accurately at larger flow rates. Increasing the flow rate improves feeder performance, which leads to lower variance in

86 76 the input concentration. Variance at the input decreases from s i = 1.1 to s i =.41 for the increase in flow rate from 3 kg/h to 45 kg/h, which essentially leads to decrease in the VRR (Figure 4-3 (a) and (c)). However, the RSD at the discharge is relatively less affected by this large change in input variability (Figure 4-3 (b) and (d)). Thus, the conclusion, for this particular case, is that the variability contributed from feeding is almost completely filtered out by the continuous mixer, provided that enough residence time is available. The final RSD of the mixture is largely dominated by the sample size and the inherent material properties, and how exposure to shear in the mixer affects them. Considering that acetaminophen is a cohesive material with large electrostatic response, the high variability in the final blend could be due to agglomeration [8,18] or electrostatic effects [19], both of which can worsen mixing performance Effect of blade configuration A clear effect of the blade configuration was not observable for the conditions examined here, possibly because the blade configuration effect interacted with the other parameters. For the lower flow rate (3 kg/h) case, the Alternate blade configuration showed a better mixing performance than the All Forward configuration at 1, 16 and 54 RPM (Figure 4-3 (e)). For the higher flow rate (45 kg/h) case, the Alternate blade configuration showed better mixing performance at 16 and 54 RPM; however at 39 and 1 RPM, the All Forward blade showed better mixing performance (Figure 4-3 (f)). Statistically, the effect of the blade configuration on the blend uniformity was significant (p = ) p =.8. Experimental results at this point are not sufficient to provide a mechanistic explanation for the observed effect of blade configuration. The entire experimental investigation performed to assess mixing performance showed that

87 77 the lowest achievable RSD was about.8 (Figure 4-3 (b)). In this case study the analytical method used for quantification of APAP was NIR spectroscopy, which utilizes very small sample size (~ 1 mg). As explained in section 4.1.1, the prediction error for NIR calibration model was.39, which further indicates that for a 3% APAP case, the lowest achievable RSD would be.39/3 = ~.8. In conclusion, the observed mixing performance was the best possible mixing performance that could be achieved at the sa mple size used in the study, and its value was entirely due the inherent material properties of APAP. 4.. Lubricant mixing Lubricant mixing experiments were conducted using MgSt as a lubricant (and tracer) and a pre-blend of Avicel and APAP as a bulk material. The experimental protocol is similar to that of API blending except for a few additions. The parameters investigated in this case study include blade configuration, rotation rate, % MgSt and the feed position for MgSt. The lubricant homogeneity in the powder was analyzed by NIR Spectroscopy. The extent of lubrication in the powder blend was also characterized by conducting wettability measurements using Washburn s technique. In addition, powder blends collected after lubricant blending were compressed using a Carver press by applying a constant pressure. The micro-distribution of MgSt in the tablet was measured using LIBS. Finally, crushing tablet hardness was also measured to characterize the bonding strength of tablets as a function of blending protocols.

88 Effect of impeller rotation rate and MgSt concentration The effect of rotation rate and % MgSt on blend uniformity, uniformity in the distribution of MgSt in tablets, tablet hardness and hydrophobicity is shown in Figure 4-4. The blend uniformity measured using NIR (RSD NIR ) does not change significantly between 39 to 16 RPM. However at 5 RPM, mixing performance becomes worse (Figure 4-4 (a)). The uniformity of MgSt distribution in tablets was characterized using LIBS (Figure 4-4 (b)). While doing LIBS measurements, all the intensity measurements corresponding to different sites and shots for a tablet were averaged. The RSD was calculated between the average intensities of such tablets. For 1% and % MgSt concentration levels, RSD measured using LIBS (RSD LIBS ) was the lowest at the intermediate rotation rates. For the.5% MgSt case, RSD increased with increase in rotation rate. These results suggest that the uniformity of MgSt distribution is dependent on sample size as well as operational parameters. The sample size in the NIR measurement is larger compared to the LIBS measurement (1cm.5cm spot size in NIR vs. 5 µm diameter spot in LIBS). For the case of 1% and % MgSt, at 39 and 16 RPM, similar level of macro-mixing is achieved, however micro-mixing is better at 16 RPM. At 5 RPM, due to lower residence times, mixing is poor on both macro as well as micro level. For the case of.5% MgSt, similar phenomenon on the macro level is observed. At micro-level a different behavior than the other two cases was observed. This can possibly be related to a sampling problem considering the very concentration of MgSt. As shown in Figure 4-4(c), the tablet hardness decreased at intermediate rotation rates for at levels of MgSt. This observation correlates very well with the relationship between the number of blade passes and impeller speed. At 16 RPM, where the number of blade

89 79 passes is maximum, the lowest tablet hardness is observed. The relationship between hydrophobicity and rotation rate was a function of MgSt concentration (Figure 4-4(d)). At % MgSt concentration, hydrophobicity was maximum at the intermediate rotation rate; otherwise it decreased with increase in rotation rate. Except for the % MgSt case, the hydrophobicity measurement did not show any correlation with the number of blade passes. This leads to the conclusion that for this particular formulation, hydrophobicity is not significantly affected by the blending parameters Effect of design parameters In order to further understand the mixing behavior of MgSt, mixing experiments with different blade configurations and weir positions were conducted. The blend uniformity measurements under different conditions are shown in Figure 3-5. The Alternate and the All forward blade configuration showed similar mixing performance at rotation rates 39 and 16 RPM. However, at higher RPMs, alternate blade showed better mixing performance, which was attributed to higher residence time. A few experiments were also conducted without the presence of weir in order to expose the blend to a shear level as low as possible. Removing the weir increased the RSD approximately by a factor of two. The hydrophobicity tests were conducted on these blends; results are presented in Figure 4-5(b). Although the mixing performance was adversely affected by the absence of the weir, hydrophobicity of the blends was not affected to a significant degree. This observation again confirms that for the conditions examined here, hydrophobicity of the blend is relatively insensitive to blending parameters. Tablet hardness was measured for different blade configurations. The Alternate blade configuration exhibited lower tablet

90 8 hardness than the All Forward blade configuration. In this case also, residence time or the number of blade passes measurements show good correlation with tablet lubrication Effect of feed position As indicated earlier, MgSt needs to be blended under the lowest possible shear. Feeding MgSt farther along the axis of the blender was examined as a possible design variable. Three different feeding positions as shown in Figure 4-6(a) were examined. The blend uniformity for MgSt blending was measured under these conditions. As shown in Figure 4-6(b), the RSD increases as the feed position gets farther from the blender inlet. At a feed position at the blender inlet, and at the center of the blender, RSD is not significantly different, which indicate complete mixing of MgSt. However, feeding MgSt near the exit of the blender produced a blend with very high RSD which showed presence of MgSt agglomerates. Clearly, feeding position close to the blender outlet is not desirable for blending MgSt. In order to further clarify the effect of the feed position, the RSD profile along the blender length was measured. The blender was stopped while it was operating at steady state, and samples along the length of the blender were collected. At each axial position, 5-1 samples were collected. Figure 4-7(a) shows, RSD profile for the case of the feed position at the blender inlet, which exhibits a plateau at approximately 5% of the blender length. The mean concentration of MgSt became steady at 75% of the blender length (Figure 4-7 (b)). Since the mean concentration of MgSt is not uniform, there seems to be a sampling problem, possible caused by the fact that the samples retrieved were very few. For the case of the feed position at the center of the blender, as shown in Figure 4-8(a), the plateau in the RSD is observed at approximately 75% of the length. Given that

91 81 these RSD profiles are under 39 and 16 RPM, while the feed position for MgSt is at the center of the blender, a further increase in blender RPM might shift the position of the plateau. It was presented in Chapter that the number of blade passes significantly decrease down when powder is completely fluidized in the mixer. Therefore, it is likely that the feed position at the center of the blender would not be suitable at high impeller speeds. Given that feeding at the inlet of the blender does not lead to over-lubrication, there seems to be no further advantage feeding MgSt at the center of the blender. 4.3 Conclusions Mixing performance was largely dominated by the material properties of the mixture and the extent of total shear (strain) applied in the mixer. Rotation rate was found to be the most significant process parameter affecting mixing performance. Intermediate rotation rates showed the best mixing performance. The effect of blade configuration on the blend homogeneity was statistically significant. This effect was also interacted with the rotation rate and the flow rate. More investigation is required for better understanding of the physical phenomenon. Some insights were also gained regarding the mixing mechanisms in the continuous mixer. For cohesive powders, a certain minimum shear rate is often required to break large clumps or agglomerates present in the mixer. In a batch mixer, this is usually achieved by applying shear using an intensifier bar. Although the uniformity of shear in a batch mixer is questionable, by providing the appropriate mixing time, total strain can be controlled. In batch mixing, the shear rate depends both on blender rotation speed and blender size, further complicating scale-up of batch processes. In a continuous mixer, shear rate and

92 8 total shear are also difficult to control independently, since other parameters such as holdup and bulk density also change by changing the shear rate (RPM). In the present case, when the shear rate is increased, the hold-up decreases. Increasing the shear rate changes flow conditions inside the mixer from a dense powder bed stirred by the impeller regime to a fluidized bed regime. Once the powder is fluidized, bulk density drops significantly, which decreases the total strain applied to the powder; as a result of fluidization, void regions are formed in the powder bed. Increasing the flow rate reduces such void regions, which in turn increases the strain applied on the powder. In the work reported here, relationships between hold-up, strain and process/design parameters were identified. In conclusion, the lowest RSD which represents the best possible mixing performance was achieved at the intermediate rotation rates and Alternate blade configuration. The lowest value of RSD was a result of the sample size analyzed and due to the inherent material properties of APAP. Measurement of strain partially helps in understanding the mixing performance at various experimental conditions. In the second case study, blending of MgSt was studied. Four responses including blend uniformity for MgSt, uniformity for MgSt distribution in tablets, tablet hardness and blend hydrophobicity were measured at each of the operating conditions. Blend uniformity, tablet hardness and MgSt distribution in tablets were strongly affected by the impeller speed and design variables (impeller design, weir position). However, hydrophobicity was found to be less sensitive to the blender parameters. The strain measurement (number of blade passes) showed good correlation with tablet hardness and MgSt uniformity in tablets. The sample size being analyzed in the NIR and LIBS

93 83 measurement also showed its effects on the RSD trends. The micro-distribution of MgSt measured by LIBS was found to be the best at 16 RPM. A reasonably good macro mixing however was achieved between RPM. At the highest impeller speed (5 RPM), the residence time drops down to a significantly low value which leads to poor macro as well as micro mixing performance.

94 Figures for Chapter 4 Figure 4-1: Schematic of the experimental set-up for LIBS Figure 4-: Experimental set-up for Washburn's method

95 VRR RSD VRR RSD 85 Figure 4-3: Comparison between flow rates: (a) VRR vs. rotation rate ( All Forward Blade configuration) (b) RSD vs. rotation rate ( All Forward Blade configuration) (c) VRR vs. rotation rate ( Alternate blade configuration) (d) RSD vs. rotation rate ( Alternate Blade configuration). Comparison between Blade configurations: (e) RSD vs. rotation rate (3 kg/hr), (f) RSD vs. rotation rate (45 kg/hr) (Note: Comparison between the blade configurations is shown only with the RSD, plots of VRR are not shown here in order to avoid redundancy) Rotation rate (RPM) (a) Rotation rate (RPM) (b) 3 kg/hr - All Forward 3 kg/hr - All Forward 45 kg/hr - All Forward 45 kg/hr - All Forward (c) 1 3 Rotation rate (RPM) (d) 1 3 Rotation rate (RPM) 3 kg/hr - Alternate 3 kg/hr - Alternate 45 kg/hr - Alternate 45 kg/hr - Alternate

96 RSD RSD 86 Figure 4-3 (Continued) (e) 1 3 Rotation rate (RPM) (f) 1 3 Rotation rate (RPM) 3 kg/hr - All Forward 45 kg/hr - All Forward 3 kg/hr - Alternate 45 kg/hr - Alternate

97 Hardness (kpa) Hydrophobicity (min /gm) RSD NIR LIBS RSD 87 Figure 4-4: Effect of MgSt concentration: (a) RSDNIR vs. Rotation rate (b) RSDLIBS vs. Rotation rate (c) Tablet hardness vs. Rotation rate (d) Hydrophobicity vs. Rotation rate Rotation rate (RPM) % MgSt 1% MgSt.5% MgSt (a).5 (b) Rotation rate (RPM) 1% Mg % Mg.5% Mg (c) 1 3 Rotation rate (RPM) (d) 1 3 Rotation rate (RPM) % Mg 1% Mg 1% MgSt % MgSt.5 %MgSt

98 Hardness (kpa) RSD Hydrophobicity (min/g ) 88 Figure 4-5: (a) Effect of design parameters on RSDNIR (b) Effect if design parameters on hydrophobicity (c) Effect of blade configuration on tablet hardness.5. (a).5. (b) Rotation rate (RPM) Alternate blade (No Weir) Forward blade W Alternate blade W 1 3 Rotation rate (RPM) Alternate blade (No Weir) Alternate Blade ( W) Preblend (before Lubrication) Forward blade ( W) RPM 1% Fwd Blade 1% Alternate Blade (c)

99 RSD 89 Figure 4-6: (a) Feeding positions for MgSt (b) Effect of feed position on blend uniformity at the blender discahrge Lubricant Preblend (Avicel + API) (a).6.5 (b) In Mid Outlet MgSt Feed Position

100 RSD % MgSt RSD % MgSt 9 Figure 4-7: Feed position at the blender inlet (a) RSD vs. blender length (a) Mean concentration vs. blender length (a) (b) Axial position Axial position 39 RPM 16 RPM 39 RPM 16 RPM Figure 4-8: Feed position - Center of the blender (a) RSD vs. blender length (b) Mean concentration vs. blender length (a) (b) Axial position Axial position 39 RPM Inside the blender 39 RPM Inside the blender 16 RPM Inside the Blender 16 RPM Inside the Blender 39 RPM - At the exit 39 RPM At the exit 16 RPM - At the exit 16 RPM At the exit

101 Tables for Chapter 4 Table 4-1: Analysis of variance (ANOVA) for the NV (Normalized Variance). df SS MS F P RPM Flow rate Blade configs RPM*Flow rate Flow rate*blade config Blade config*rpm Error 19.. Total 31.

102 9 Chapter 5 Continuous monitoring of powder mixing process by NIR Spectroscopy Continuous monitoring of blend uniformity is the first step towards implementing process control for continuous blending operations, or to facilitate rejection of non-uniform powder from the blending operation. In this chapter two case studies are presented where on-line NIR analyzers (CDI and VTT) were used to monitor the continuous powder blending process. As described in Chapter 4, it is required to specify the sample size being analyzed by the NIR analyzer while reporting any blend uniformity measurement. In the continuous blending process, typically powder is in a state of motion and inherently there is always a certain degree of spectral averaging involved in the measurement. Blend uniformity, quantified as RSD (Relative Standard Deviation) between the in-line measurements, is dependent on the degree of averaging. It is necessary to select the averaging window depending on the sample size of interest (one unit dose), which is often not the case in the current industrial practice. Once the correct averaging window for the in-line measurements is determined, the blend uniformity measured from the in-line measurements can be directly compared with the off-line measurements. In order to be able to do such a comparison, quantification of error associated with in-line and off-line measurement is necessary. This chapter is broadly divided in two parts. In this first, a methodology for building chemometric calibration models using the on-line spectral data is presented. Here calibration model building exercise was based on the data acquired using a CDI on-line NIR analyzer. In the second part, a methodology is presented that allows quantification of the error associated with the in-line measurements, as well as the relationship between in-

103 93 line and off-line blend uniformity measurements. This methodology was developed using a multi-point fiber optic probe NIR monitoring system provided by VTT. 5.1 Chemometric calibration model development using on-line NIR spectral data Equipment and experimental set-up The experimental set-up designed for monitoring blend uniformity at the discharge of the continuous blender is shown in Figure 5-1 (a). A chute was installed at the outlet of the blender. The connecting piece used to mount the CDI Spectrometer on the powder conveying chute is shown in Figure 5-1 (b). The level of powder on the chute was dependent on the flow rate and the cohesivity of the powder. The angle of inclination was adjusted such that the distance of the flowing powder from the spectrometer remained approximately constant Materials and pre-blend preparation The materials chosen for this study were micro-crystalline cellulose (Avicel-PH 1, FMC BioPolymer), acetaminophen (micronized, Mallinckrodt Inc.), colloidal silicon dioxide (Carbosil), and magnesium stearate (Mallinckrodt Inc.). Pure acetaminophen was pre-blended with silicon dioxide (3%) in a V-blender with intensifier bar rotating at 1 rpm to improve the flow properties since it posed difficulties in feeding through loss-in-weight feeders due to its highly cohesive nature. Avicel was also pre-blended with magnesium stearate (1% (w/w)). All experiments were performed at a constant total flow rate with a constant rotation rate of the agitator in the continuous blender. Different

104 94 experimental conditions consisted of different levels of APAP. The feed rates used in the experimental runs are presented in Table NIR Spectroscopy Preparation of calibration samples Calibration samples used in this study were prepared in a V-blender (4 quarts, Patterson Kelley). The calibration samples consisted of 6 g blends where acetaminophen (APAP) ranged in concentration from to 15% (w/w), with samples prepared at approximately 1.5% intervals. APAP was used diluted with Carbosil and Avicel lubricated with magnesium stearate in the same manner as in the continuous mixing experiments. The following sections describe spectral acquisition and development of the calibration model Instrumentation A CDI (Control Development Inc., South Bend, IN) Blend Uniformity Analyzer NIR spectrometer was used to acquire the spectra of the flowing powders. This instrument includes a 5.4 W dual tungsten halogen light source, and an indium gallium arsenide (InGaAs) diode array that is thermoelectrically cooled and has 56 elements to cover the nm spectral area. Reference spectra were obtained with an AutoCal feature that includes an internal Hg/Ar lamp used for wavelength calibration and a spectralon reference plate that is automatically placed by the spectrometer in the optical path for both wavelength and baseline reflectance calibrations. Each sample was analyzed by the NIR spectrometer in dynamic mode with an integration time of 5 ms and co-adding 3

105 95 scans of the powder flowing through a chute specially designed for the continuous mixer. The spectrometer transferred the spectra via wireless communication to a laptop computer Estimation of sample size for NIR measurement The sample size analyzed during the online monitoring was estimated as follows. The NIR illumination spot was approximately 5 mm in diameter. The penetration depth for the NIR radiation was assumed to be 1 mm [11,111]. The velocity of the powder was determined to estimate the length of time over which the NIR radiation interacted with the sample. The velocity was determined by measuring the average residence time with a small amount of colored tracer material introduced at the beginning of the chute, and measuring the time it took to travel across the chute using a chronometer. The measurement with tracer was repeated 1 times, and the average velocity for a 3 inclined chute was found to be 18 cm/s. Each spectrum was an average of 3 scans, with a scan time of 5 ms for an estimated sampling time of 16 ms. Hence, the scanning length (velocity*time) was estimated to be.88 cm. Having the length, spot diameter and thickness of penetration measured, the volume of the sample was calculated to be.73 cm 3. The bulk density of APAP Avicel mixture was measured to be.36 g/cm 3. Thus, sample size (mass) was calculated to be.6 g for each co-added spectrum. For the calibration samples, a total of 1 co-added spectra per sample were acquired, depending on the flow rate. The blends were emptied three times down the chute to collect three sets of spectra with the calibration blends. In each experiment, three consecutive spectra were averaged.

106 96 The calibration blends were also used to develop an off-line calibration model to predict the drug concentration and further evaluate the method's precision. Samples of the 3%, 6% and 8% APAP blends were collected for offline analysis. Samples weighing approximately 1 g were retrieved from the chute after 3, 6, 9 and 1 s. A total of five NIR spectra were obtained for each sample collected. After acquiring each spectrum, the powder in the area illuminated was removed and this procedure was repeated four times. The drug content was then predicted by the off-line calibration model Software and NIR data processing Spectra were collected with the Spec 3 (version ) software provided by Control Development (South Bend, IN). The Pirouette 4. software developed by Infometrix (Bothell, WA) was used to evaluate and extract information from the spectra obtained. All calibration models were developed using the partial least squares (PLS) regression method. First and second-derivative (1st der and nd der) spectra were obtained by using the Savitzky Golay algorithm with a 15-point moving window and a second-order polynomial. Standard normal variate (SNV) was also used as spectral pretreatment. PLS calibration models were constructed by cross-validation, using the leave-class-out method. Calibration and external prediction sample sets were chosen via scores plot of a principal component analysis (PCA) of the first two PCs. Both sample sets encompassed the complete concentration range. The quality of the models was assessed in terms of root mean square errors of cross validation and prediction (RMSECV/RMSEP).

107 Results Exploratory data analysis Figure 5- shows the NIR spectra of pure APAP and Avicel plus a blend of 15% (w/w) APAP. The changes in APAP concentration are reflected on certain bands (113, 1 and 149, 165 nm). Certainly, the last two mentioned bands presented the larger spectral changes; however, these ranges also showed the most intense Avicel bands. The first band (113 nm) is due to APAP and is not affected by overlapping with the excipient spectral bands. The second band (1 nm) is due to Avicel. Figure 5-3 shows the scores plot (PC-1 and PC-) of a principal component analysis calculated with the entire calibration data set ( 15%APAP), after SNV and first derivative spectral pretreatment for the nm range. The main source of variability (PC-1, 65.6% spectral variance) is the APAP concentration. The second PC explained 8.6% of X-spectral variance. Each cluster of spectra is clearly separated according to the concentration. The ordering of all clusters depended on the concentration confirming that the nm spectral range is adequate to further develop a calibration model for APAP determination Development of NIR calibration model The calibration model was developed with spectra of the blends flowing through the chute with a set of 11 calibration blends ranging in drug concentration from to 14.46% (w/w) and varying in steps of about 1.5% (w/w). This experiment was repeated three times. The calibration model was first developed using the set of spectra from the first experiment with the calibration blends. These spectra were then used to predict the APAP concentration in the same samples from the second and third run. During the method

108 98 development stage, several wavelength ranges were evaluated; however, the best calibration models were obtained over the nm range, and Table 5- describes the figures of merit for calibration models using this range. The different spectral pretreatments were carefully evaluated due to the changes in the baseline of the spectra acquired. This initial evaluation led to the selection of the nm range, and to narrowing down the pretreatment to only two possibilities: the standard normal variate (SNV) and SNV followed by first derivative transformation. The number of PLS factors was also evaluated and selected based on the minimum error of prediction for crossvalidation and external prediction samples. The spectra from the three runs with the calibration blends described in Table 5- were then combined into one calibration model with a total of 156 calibration spectra, with the objective of obtaining a more robust calibration model. The new model provided a root mean square error of cross validation (RMSECV) of.41% (w/w) for the samples predicted in a leave-class-out cross validation. Each concentration was defined as a class, so that when a sample of a specific concentration was predicted all samples of that concentration were left out of the calibration model. The calibration model was also challenged using a Pirouette routine that randomly selected 6% of the samples, which were then used to develop the calibration model. The remaining 4% of the samples were not included in the calibration model, and were predicted with a RMSEP of.34% (w/w). The lower RMSEP obtained with random sample selection reflects that leave-class-out cross validation is a very challenging method for evaluation of the calibration model, since the concentration of the predicted samples is not included in the calibration set. The calibration model was then established

109 99 with the 156 samples from the spectra of the three experiments described in Table 5-, and this model was used for all predictions of the drug content of spectra obtained in the continuous mixing experiments. Leave-class-out cross validation and random sample selection did not generate different calibration models, these were methods to evaluate the model and one model based on the 156 samples was developed. Figure 5-4 demonstrates the linearity of the obtained method. The scatter plot shows the predicted APAP concentration using the NIR method (leave-class-out cross-validation and external prediction values) versus the reference method (analytical balance). The external prediction results shown in Figure 5-4 are those obtained for the randomly selected validation samples that provided an RMSEP of.34% (w/w). Statistical t-testing (95% significance level) was evaluated and there is no significant difference between the slope and intercept with 1 and, respectively. The RMSEP value of.41% (w/w) indicates the method's accuracy over all the concentration ranges evaluated, providing a summary of the method's performance throughout the different concentration levels evaluated. Because the calibration model was developed to evaluate the performance of the continuous mixer for blends where APAP varied from to 1% (w/w), the model's accuracy and precision should be estimated at each of the concentration levels presented in Table 5-3. The use of leaveclass-out cross validation facilitated the calculation of the RMSEP (accuracy) and precision (standard deviation) at each of the concentrations. The model's accuracy was considered adequate at all concentrations except for the 1.46% (w/w) blend. A separate calibration model was developed using the spectra from to 6% (w/w) blends, based on previous studies where the determination of low concentration samples improved when a

110 1 narrower range of concentrations was used [11]. The 6% (w/w) was considered adequate in terms of accuracy with a RMSEP of.34% for the 1.46% (w/w) blend, and precision (standard deviation of.6% (w/w)). This model included two PLS factors for the nm spectral range and was used only for the % (w/w) continuous mixing run. The method's precision was also evaluated. In this application, blend homogeneity was assessed in terms of the precision of the drug concentrations determined for the blend flowing through the chute. The variation depends on the distribution of the drug in the blend and also on the unavoidable random error of the analytical method. Thus, the precision of the calibration model was also evaluated at the different concentration levels. Table 5-3 presents the standard deviation which ranged from. to.4% (w/w) at the different concentration levels, with a pooled standard deviation of.6. These calculated standard deviations encompass the variation of the analytical method as well as variations related to the chemical composition of the calibration blends. The precision of the NIR method was further evaluated through the development of a calibration model for spectra obtained off-line in a laboratory setting, where the powder blend was not flowing as in the spectra collected in the chute following the continuous mixer. The RMSEP for this model was now lower and the standard deviation ranged from.7 to.19% (w/w), with a pooled standard deviation of.11% (w/w). These results indicate that the off-line analysis of the calibration blends provides a variation (pooled standard deviation) of about.11%. When the same blends are analyzed as they flow through the chute, the method's variation is about.6% (pooled standard deviation). These results also indicate that the minimum variation that could be expected

111 11 from continuous mixing is about.6%, based on the random error of the analytical method Evaluation of continuous mixing runs Figure 5-5 shows the predicted concentrations for the continuous mixing experiments providing APAP concentrations every.5 s. A typical start-up process for the continuous mixer can be described as follows. Once the mixing process is started, the hold-up (mass) of the powder in the continuous blender increases with time and reaches a constant value. When the mixer is operating under constant hold-up, the incoming and outgoing feed rates are equal and the process is considered under steady state operation. Close evaluation of these experiments showed that 6 8 s were required for the steady state to occur, with the materials and conditions used. Spectral data were also collected during the start-up (before steady state), and the concentration profiles showed some deviations from the theoretical mean value. These deviations occur due to the different start-up times for APAP and Avicel feeders and also due to the transient hold-up in the mixer. The steady state is shown in Figure 5-5 by the vertical line at 8 s across all the concentration profiles. The standard deviation was used as the homogeneity index to determine the uniformity of the blend. The standard deviation was calculated for each run using the predictions under steady state (for all concentrations predicted after 8 s) as presented in Table 5-4. The highest standard deviation obtained was.64% for the 6% (w/w) run, and the second highest was.5% for the 8% (w/w) blend. A one-sided F-test indicated that these standard deviations were greater than the pooled standard deviation of.6 obtained for the analytical method. The standard deviation of the, 3, and 1% (w/w) blends was

112 1 much lower and considered similar to the variation expected from the analytical method. These results do not guarantee that the blend uniformity will be reduced to standard deviation of.3.6% (w/w) as segregation can occur after blending. Additional studies are planned where the drug concentration will be evaluated after tableting Conclusions The NIR spectroscopy and multivariate data analysis was demonstrated as an effective tool for the real-time determination of active ingredient in the output blend. The CDI spectrometer was found to be feasible for measuring blend uniformity in continuous flow systems. The sample size analyzed in the on-line NIR measurements was approximately.6 g which is close to the typical unit dose used in the pharmaceutical industry. In three of the five blending runs, the on-line blend uniformity measurements performed were close the analytical method error. This indicates that continuous mixer is capable of producing homogeneous blends as close as to the calibration standards. However, a further investigation is suggested in estimating the error in the measurement and relating that to the sample size. In the next session, a methodology is presented to estimate error in the in-line measurements. 5. Continuous monitoring using VTT Spectrometer 5..1 Equipment and experimental set-up Single-point spectrometers are used in most of the present-day PAT applications. However, the use of multipoint NIR systems has some advantages in continuous pharmaceutical manufacturing. One advantage is instrument or process failure diagnostics: one can compare the results of multiple measurement points at the same

113 13 measurement position and diagnose the process or instrument failure. Another advantage is that one instrument can serve the whole continuous manufacturing line, because one just needs to add probes at the desired measurement positions. In this work, a multipoint NIR measurement system was developed for the continuous mixing process application. It consisted of the following main parts: 1. Fiber-optic light source. 5-point fiber-optic probes 3. Spectral camera (Specim Spectral camera NIR) with fiber-optic inputs 4. Measurement software for acquisition of spectra, predicting the concentrations in real time and sending data to the process control system (custom program written with Labview, National Instruments, Austin, TX) Of these, the 5-point fiber-optic probes were specifically tailored for the process by VTT, Finland. Figure 5-6 shows a schematic of the multipoint NIR system and how it was used in the continuous blending process. The fiber-optic light source (VTT, Finland, see Figure 5-7 (a)) provides the illumination for the probes, and the spectral camera (Spectral camera NIR, Specim, Finland, see Figure 5-7 (b)) is used to collect the spectra from all of the probes simultaneously. The spectral camera is able to collect up to 5 spectra per second from each of the fiber-optic channels simultaneously. The maximum number of probes that can be attached to the spectral camera is 16. The useable wavelength range of the spectral camera is nm.

114 14 The main objective in using the multipoint NIR system to monitor the continuous mixing process is to use the NIR predictions of mean API concentration and homogeneity (RSD) to control the process. The best possible measurement position was found to be just at the discharge of the continuous blender (see Figure 5-8 (a) and (b)). Two configurations for the probe were tested: the above-the-chute configuration (Figure 5-8 (a)) and the below-the-chute configuration (Figure 5-8 (b)). The benefit of the above-the-chute configuration is that no window is needed in the chute, and therefore there are no problems of powder accumulation on the window. The benefit of the below-the-chute configuration is that the probe is shielded against dust, which is formed when the powder falls from the blender outlet onto the chute. The results presented in this paper correspond to the above-the-chute configuration, which works well with powders with a relatively large particle size which do not tend to produce dust. 5.. Methods UV Assay The formulation used in this study consisted of 3% Granulated Acetaminophen and 96% Avicel PH-11 and 1% Magnesium Stearate. Micronized acetaminophen (d5~µm) was used to build a calibration curve of UV absorption. The following procedure was followed to prepare calibration standards. Samples were weighed as 1, 9, 8, 7, 6, and 5 mg of pure acetaminophen. They were added in a 5 ml volumetric flask. ml of methanol (MeOH) was added and samples were stirred until the acetaminophen was completely dissolved. Distilled water was added in the volumetric flask to make the total volume of 5 ml. A 5 ml aliquot was taken out from the solution into a 1 ml

115 15 flask and diluted further with distilled water. An aliquot from the 1 ml vol. flask was taken further into a quartz cuvette for UV absorption measurement. UV absorption at 44 nm was recorded as triplicate. This procedure was repeated for all the calibration standards. A calibration curve was fit using linear regression with R >.99. The UV reader used was Ocean Optics USB4 Miniature UV/VIS Fiber Optic Spectrometer. Since the formulation used in the mixing experiments contain granulated acetaminophen (which also contained PVP), it was necessary to check for the interference of PVP at 44 nm. The following procedure was followed to check for PVP interference. Since the formulation contained 3% COMPAP, the total PVP in a 5 mg tablet will be 1.5 mg. Twice of the theoretical amount (3 mg PVP) was added in a 5 ml vol. flask. Distilled water and MeOH was added in the flask following the same dilution procedure as that of other tablets. The solution was centrifuged and an aliquot from that solution was subjected to UV analysis. No significant interference was observed, as the UV absorption reading was very close to a case with just distilled water. The UV assay was validated by performing recovery experiments. 194 mg of MCC and 6 mg of micronized APAP was weighed and added in a L flask. Subsequently, 6 ml of MeOH was added, and the flask was stirred to give the methanol ample opportunity to dissolve the APAP. Then, 1 ml of water was added and the flask was stirred further for another minutes before completing to volume in a. L volumetric flask. The solution was centrifuged at 3 rpm for 5 minutes before testing it in the UV spectrometer at 44 nm. The drug concentration in the solution was calculated using the calibration curve. This procedure was repeated three times to calculate the average drug recovery. This study was repeated using 8% and 1% of the drug content. Recovery

116 16 was found to be between 98 1% of the drug, and an RSD of less than.5% (w/w) was obtained PLS (Partial Least Square) model development for NIR Spectroscopy Calibration samples with APAP concentrations ranging from % to 6.5 % (by mass) were prepared manually and mixed in a vortex shaker. The magnesium stearate concentration in the calibration set was randomized around a mean value of 1%. The list of off-line calibration samples is presented in Table 5-5. Two 5-point probes were used in measuring the calibration set. To allow a simultaneous measurement with both 5-point probes, each calibration sample was divided in two rectangular aluminum containers. The procedure for measurement was as follows: 5. Divide the sample into the two containers. 6. Mix the powders manually in the container to avoid segregation effects. 7. Place the containers in the measurement position of the two 5-point probes and take one measurement. 8. Mix the samples manually again to get a different representation of the sample surface. 9. Go back to item 3 and repeat ten times. 1. In the middle of the ten repeats, exchange the containers so that both probes will see each of the containers. Measuring in this fashion gave 14 samples x 1 repetitions x 1 probes = 14 measurements altogether.

117 17 Figure 5-9 shows the unprocessed spectra measured from the calibration set together with the baseline corrected spectra. The APAP absorption bands are located at approximately 113and 165 nm. They are hardly visible in the spectra by the naked eye because of the low APAP concentration range. Still, it is possible to use all the 14 spectra in the calibration model. However, this leads to high levels of sampling noise since the measurement spot size was only about 3 mm in diameter. A better way is to average over the subsamples (i.e. over the ten repetitions / sample), leaving behind only 14 spectra (14 samples x 1 probes). In this way, the probe-to-probe variation is preserved in the calibration model, but the effect of sampling noise is minimized. In the following description of the calibration model and analysis of in-line data, the spectra used were always averaged in this manner unless otherwise stated. The standard PLS approach was selected as the calibration model. The first step is to choose a suitable preprocessing method. The methods tested included baseline correction, multiple scattering correction (MSC), standard normal variate correction (SNV), first and second derivative, and combinations therewith. The most notable interference in the raw spectra is the shift of the baseline, which is nearly completely removed by the baseline correction (cf. Figure 5-9). Moreover, there does not seem to be any substantial scaling variation in the spectra. Therefore, methods that try to cope with the scaling variations, such as MSC and SNV, tend to remove meaningful spectral information, thus resulting in suboptimal calibration performance. Hence it does not come as a surprise that the baseline correction alone worked the best out of the methods tested. The specific baseline correction method used in this work was the projection of a straight line and linear tilt variations out of the spectra, i.e. a

118 18 detrending-type approach. If we arrange a set of spectra row by row in the matrix, the baseline correction can be expressed in matrix notation [113]: Here, denotes the baseline corrected spectra, is the projection operator, and is the unit matrix. The denotes a column vector of ones, and denotes a column vector of the wavelength scale. Figure 5-1 shows the scatter plot resulting from using the PLS model for the calibration set (left side) and the scatter plot after cross-validation (right side). Leave-one-out crossvalidation was used, where one sample at a time was removed from the calibration set. The PLS model was used to predict the API concentration of in-line measurements. In the analysis of this work, the average predicted API concentration over the five points of a 5- point probe was used. To give some perspective on the calibration performance for the probe-averaged results, the scatter plots averaged over the five points of each probe are shown in Figure As expected, the calibration performance is considerably better after averaging, especially for probe number two Results Methodology for estimating error in the NIR measurement Total ( s) Mixing ( s ) (5-1) s represents the sample size being analyzed. Equation (5-1) shows that the total variability ( Total ( s)) in concentration measured at the blender discharge can be

119 19 expressed as the sum of the variability due to mixing ( Mixing ( s)) and analytical method error ( ). In equation (5-1), total variance and variance due to mixing are both functions of the sample size, whereas method error is independent of the sample size. The RSD is defined in equation (5-). In equation (5-3), all the terms are normalized by the square of the mean concentration. RSD C N n 1 ( C i N C C ) 1 (5-) RSD Total ( s) RSDMixing ( s) RSD (5-3) RSD Total ( s) RSD RSD Mixing ( s) (5-4) RSD Mixing ( s). s (5-5) The normalized variance (or ( s) ) can be expressed by a power law relationship RSD Mixing as shown in equation (5-5). After substituting this relationship in equation (5-4), it can be re-written as equation (5-6). RSD Total ( s) RSD. s For a random blend, the exponent (5-6) in equation (5-5) is -1. In this case study, our objective is to determine this relationship for a set of data from UV absorption spectroscopy and NIR spectroscopy. In order to obtain the values of the coefficients in the powder law and the method error, the following procedure was followed: Equation (5-6) was logarithmically linearized as equation (5-7).

120 11 Ln( RSDTotal ( s) RSD ). Ln( s) Ln( ) (5-7) In equation (5-7), y Ln( RSDTotal ( s) RSD ) is expressed as a linear function of x Ln(s), where and Ln( ) are the slope and the intercept respectively. The optimal values of RSD, and were obtained by maximizing the regression coefficient coefficient R for this linear equation. The optimization problem is shown in the following equations. Objective function: N N ( x y ) N x i i i i i 1 i 1 i 1 R (5-8) N N N N N ( xi ) xi N ( yi ) yi i 1 i 1 i 1 i 1 N y RSD Constraints: [ RSD Total ( i 1,,.., N ) ( s) RSD ] Values of RSD were selected iteratively by using an optimization program in MATLB Data collection of RSD as a function of sample size RSD data as a function of sample size was acquired using in-line NIR spectroscopy and off-line UV absorption assays In-line measurements and data fitting In-line measurements were performed by mounting the fiber optic probe above the chute, and projecting the NIR radiation onto the flowing powder. Measurement was performed at five spots arranged in the transverse direction of the flow. The experimental set-up is shown in Figure 5-7. In order to capture the effect of sample size on RSD, sample size was varied by averaging different numbers of scans. Sample size could also be varied by

121 111 increasing the cross sectional area of the NIR beam. However, this imposes limits on the lowest and the highest sample size that can possibly be analyzed, and it also affects the intensity of NIR radiation, which is higher if the spot size is smaller. In our measurement set-up, a constant spot diameter ( d 3mm) was used. With an increasing number of scans, the area exposed to the NIR radiation increases. The scanning process is discrete; it requires 5 milliseconds to acquire each spectrum (.5sec), and milliseconds to post-process it. The area being analyzed per scan can be approximated as A v.. d d / 4, where v is the powder velocity. The velocity of the powder on the chute was measured by performing impulse tests. Two probes were installed on the chute, the first one at the beginning of the chute and the other at the end of the chute. A tracer was inserted in the flowing stream of powder on the chute, and the time required to travel the distance between two probes was measured. This test was repeated for about 1-15 times to obtain a good estimate of the velocity of the powder on the chute. The total area analyzed for one unit sample size (g) can be calculated just by multiplying the area for one scan ( A v.. d d / 4) and the number scans (n) being averaged. Penetration depth of the NIR radiation (l) for similar pharmaceutical powders was obtained from literature [11,111] and found to be 1 mm. It is expected that the penetration depth of the NIR radiation will vary as a function of particle size and the intensity of light being used, but for this particular case study 1mm penetration depth was assumed in order to develop the basic methodology. Thus the sample size (g) being analyzed in the NIR measurement can be expressed as s n. A. l., where (. 36 g/cm 3 ) is the bulk density of the powder. Calculation of RSD for the in-line measurements was performed using the following procedure. For the length of the steady state run, N consecutive measurements were

122 11 predicted using the previously developed PLS model. By increasing the averaging window (m), the number of averaged measurements ( n N / m) decreased. When the highest number of scans were averaged, n was as low as 1. In the interest of capturing the effect of sample size (and not confounding it with different number of samples) on RSD, n was kept constant to be 1. Thus RSD was calculated for only 1 samples (selected randomly from a pool data at each sample size) under varying sample sizes. The relationship between RSD and sample size is shown in Figure 5-1. RSD decreases with increase in sample size, and eventually becomes a steady after a sample size of.3 g at approximately Off-line measurements (UV Absorption Assay) Off-line analysis of blend uniformity was performed by sampling tablets instead of sampling powder blends due to the fact that it was found difficult to extract samples of considerably smaller size from the blending process. Tablet samples provide an easy way to reproduce the sample size. Once the powder is compacted in the tablet form, samples of smaller size can be made simply by cutting pieces from the tablet. The cutting procedure leads to some variability in the sample size, but this contribution was found to be negligible. In addition, using this method samples do not segregate, which they would happen if a large powder sample was sub-divided. A total of samples were analyzed at each sample size. In order to measure blend uniformity at a sample size greater than one tablet, two tablets were dissolved together, which gives a sample size twice that of a single tablet. Raw data consisting of sample size, mean concentration, RSD and the confidence intervals for RSD is presented in Table 5-6. The relationship between RSD and sample size for the off-line data is shown in Figure RSD decreases with

123 113 increasing sample size and eventually shows a plateau around., which is similar to the case of in-line measurements. This result indicates that. is the minimum possible RSD that can be reached for this particular powder mixture. Since the UV absorption calibration model was found to be very accurate (R >.99), the similarity between the two cases also indicates that the base-line RSD of. for the in-line measurements is primarily due to the inherent non-uniformity in the powder blend, and not analytical method error Fitting of the experimental data in the mathematical model In order to quantify the method error, the fitting procedure described in the section 4.1 was applied for the in-line as well as off-line RSD data. Table 5-7 summarizes the model fitting results. The model fitting procedure was exercised for different ranges of the RSD vs. sample size data. This step is necessary to determine the optimal range of RSD vs. sample size data, which minimizes the number of experiments required to arrive at the same conclusion. A p-value of <.1 was obtained for the in-line data for all possible selections of sample size ranges, which indicates good fitting of the experimental data. The fitting parameter ( RSD ) was found to be zero in all the ranges except for the case of -.1g. Off-line data fitting was relatively poor because fewer data points were available. Model fitting up to a range of -. g was good (p-value <.1); model fitting was not good for the case of -.1g range (p-value =.31). In this case also, method error ( RSD ) was found to be zero. Model fitting results concluded that RSD. is purely because of the inherent nonuniformity in the powder mixture, and that the variance contribution of the analytical

124 114 method error is negligible. Another observation from these results is that a range of -.43 is enough to determine the method error, and more information in the asymptotic region of the curve (RSD vs. sample size) is not very useful. These results also provide an important guideline to select an optimal unit dose size based on blend uniformity criteria. The number of scans to be averaged to measure one unit dose (.43g) was found to be 6. This number changes depending on the NIR penetration depth, powder bulk density and the scanning area used in the measurement set-up. The experimental data and model predictions are shown in Figure 5-14 (In-line data) and Figure 5-15 (Off-line data) respectively. The values of and Ln( ) were very close for both in-line and off-line data (Table 5-7) which shows that the same mixing model applies to both cases Conclusions A case study of an application PAT based on NIR spectroscopy to continuous blending was demonstrated. In-line measurement at the blender discharge was made possible by utilizing a chute on which powder was allowed to flow by gravity. Sample size of the powder being analyzed in the in-line measurements was determined by measuring the velocity of the powder on the chute. Measurement of velocity made it possible to determine the number scans required to be averaged to measure RSD relevant to one unit dose sample size. A methodology was presented which provides a way to quantify the analytical method error in the in-line blend uniformity measurements, and compare that with the off-line measurements. For the case study presented in this paper, the method error in the in-line as well as off-line measurements was found to be negligible. The base-line RSD of. was attributed to the inherent non-uniformity of the powder blend. This methodology

125 115 provides a direct relationship between PAT and laboratory data, which is very important for reducing the analytical testing time in the pharmaceutical industry. This study is very useful for allow real time release (RTR) of drug products based on blend uniformity criteria, and to determine the optimal size of a unit dose.

126 Figures for Chapter 5 Figure 5-1: (a) CDI Spectrometer installed on a powder conveying chute at the mixer discharge (b) Chute (a) (b) Spectrometer mounting Chute

127 117 Figure 5-: NIR spectra for acetaminophen and for % and 15% (w/w) powder blends Figure 5-3: Scores plot from principal component of analysis of calibration set spectra in nm spectral range.

128 118 Figure 5-4: API content predicted by NIR for cross validation and external validation samples Figure 5-5: NIR predictions from monitoring the continuous mixing process for three representative blends

129 119 Figure 5-6: Schematic of the multipoint NIR measurement equipment, it consists of a fiber-optic light source, fiber-optic probes and a fiber-optic spectral camera Mixer Chute Powder Illumination fiber bundles Light source Spectral camera Probe 5 measurement spots Collection fiber bundles Real-time calculation module To process control Figure 5-7: (a) Schematic picture (left) and photograph (right) of the multipoint fiber-optic light source. It has 4 output fibers with a ST connector (b) The fiberoptic spectral camera (Spectral camera NIR, Specim Ltd., Oulu, Finland) (a) Output fibres Fibre bundle Chopper blade Lamp (b) Collection mirror Imaging mirror

130 Absorbance BL-corr. spectra 1 Figure 5-8: Experimental set-up (a) Above-the-chute configuration (b) Below-thechute configuration Probe (a) Mixer outlet Continuous mixer (b) Mixer outlet Tablet press inlet Chute Measurement window Probe Figure 5-9: The unprocessed calibration set spectra (left) and the spectra after the baseline correction (right) Absorbance spectra Baseline corrected spectra Wavelength [nm] Wavelength [nm]

131 Prediction Prediction Prediction Prediction 11 Figure 5-1: The scatter plot of using the PLS model with the calibration set (left) and after cross-validation (right) Calibration set, 5 factors used RMSEC :.3157 cc :.9879 SEC : slope : offset :.7844 bias : -7.34e-15 CV : R : #of smpl : Reference Cross-validation, 5 factors used RMSECV : cc :.9835 SECV : slope :.963 offset :.1955 bias : CV : R :.9651 #of smpl : Reference Figure 5-11: The scatter plot of cross-validation after averaging the results of each sample over the 5 measurement points of probe number 1 (left) and probe number (right) 7 RMSEC :.688 cc : SEC :.7685 slope :.918 offset : bias : CV : R : #of smpl : 14 3 Cross-validation, 5 factors used 3 Cross-validation, 5 factors used 7 RMSEC : cc : SEC : slope : 1.11 offset : bias : CV : 5.88 R : #of smpl : Reference 4 6 Reference

132 RSD RSD 1 Figure 5-1: Blend uniformity (RSD) as a function of sample size (NIR Spectroscopy) Sample size(g) Figure 5-13: Blend uniformity (RSD) as a function of sample size (UV Absorption) /4 Tablet 1 Tablet 1/ Tablet Tablets Sample size(g)

133 Ln (RSD ) Ln (RSD ) 13 Figure 5-14: (a) RSD as a function of sample size (Comparison between NIR data and mathematical model) (b) Linear regression for the best case ( RSD ) (a) RSD Sample size (g) Experimental (NIR) Model (b) y = x R² =.71 Ln (Sample size) Experimental (NIR) Linear (Experimental (NIR)) Figure 5-15: (a) RSD as a function of sample size (Comparison between UV absorption data and mathematical model) (b) Linear regression for the best case ( RSD ) (a) RSD Sample size(g) (b) y = -.66x R² =.81 Ln (Sample size) Experimental (UV Absorption) Model Experimental (UV Absorption) Linear (Experimental (UV Absorption))

134 Tables for Chapter 5 Table 5-1: Experimental conditions for continuous mixing experiments % APAP (w/w) APAP feed rate (kg/h) Avicel feed rate (kg/h) Table 5-: Development of calibration models and its initial evaluation Spectral pre-treatment Sample set RMSEP (%) 1 st Derivative Cross.6 Run.58 Run 3.65 nd Derivative Cross.68 Run.49 Run 3.71 SNV Cross.37 Run.33 Run 3.35 SNV 1 st Derivative Cross.33 Run.7 Run 3.35 SNV nd Derivative Cross.4 Run.48 Run 3.5 Table 5-3: Evaluation of model precision (standard deviation) and model accuracy (RMSEP) at each calibration concentration. % APAP (w/w) Predicted Standard RMSEP % (w/w) Average(w/w) Deviation

135 15 Table 5-4: Evaluation of continuous mixer experiments at various APAP concentrations. % APAP (w/w) Mean Concentration (%) Standard deviation Table 5-5: Off-line calibration samples Sample # APAP [%] MgSt [%] Avicel 11 [%]

136 16 Table 5-6: Off-line (UV absorption) data of sample size, mean concentration, variance, RSD and confidence intervals Confidence Interval (C.I.) Mean for RSD Sample concentration Variance ( N 1) ( N 1) RSD size (g) (C ) ( ) C /, N 1 1 /, N Table 5-7: Model fitting results for in-line and off-line data Range Ln ( ) R p value ( RSD ) (g) ( RSD ) In-line data (NIR Spectroscopy) (-.84, (-7.9, ) 7.7) (-1.,-.7) -8.3(-8.31, ) (-1.3,-.79) -8.54(-9.14, ) (-.,-1.1) -1.3(-11.49,- 9.15) E Off-line data (UV Spectroscopy) (-1.1,-.3) (-8.89, ) (-1.45,-.3) -8.14(-9.9, ) (-.,.19) -8.9(-1., ) (-4.33,3.) -7.94( ,4.3).78.31

137 17 Chapter 6 Development of integrated continuous mixing and delumping process In previous chapters, the feasibility of a continuous powder mixing process was shown for cases of APAP (silicated) blending and MgSt blending. However for cases involving blending of highly cohesive powders, continuous mixing process loses its flexibility. If the material to be blended is very cohesive, in a batch blender the necessary shear can be applied by using an intensifier bar. In the continuous blender, increase in the rotation rate leads to decrease in residence time making it difficult to apply high shear on the powder. Experiments with pure acetaminophen and Avicel were conducted in the continuous blender which showed presence of agglomerates. Feeding, which is an essential part of the continuous mixing process, also adds more constraints on the process. Feeding highly cohesive material is difficult. Although feeding can be improved using proper tooling, sometimes pre-blending with glidants is the easiest and most efficient step. However, preblending introduces additional steps, which sometimes are intrinsically batch mode. One possibility to address these issues would be to use an in-line mill in the process to break the agglomerates early, perhaps right after the feeding step. The product from the mill could be fed in the continuous mixer, providing additional mixing. Having two units as opposed to one can provide better control on the shear environment in the process. Alternatively, the mill can be used at the discharge of the mixer to improve microhomogeneity of the blend. In this chapter, an integrated process consisting of de-lumping using a Comil and mixing using the Gericke mixer is developed.

138 Mixing effects in low shear (Gericke mixer) and high shear mixing (Quadro - Comil) continuous mixing equipment A case study of mixing of cohesive and free flowing powders in a continuous system was examined. Micronized APAP and Micro-Crystalline Cellulose (MCC) were used as the model materials to represent cohesive and free flowing powders respectively. The mixing problem in this case can be decomposed as a combination of macro-mixing in the blender and micro-mixing in the mill. Macro-mixing which is governed by the bulk powder flow behavior in this mixer is required to compensate for the incoming feed rate variability (axial mixing), and to mix the initially unmixed powders (radial mixing). Micro-mixing is also required de-lump the agglomerates of acetaminophen and reduce the scale of mixing from agglomerates to primary particles. In this case study, the combined performance of Comil (Quadro) and a continuous mixer (Gericke) for micro and macro mixing was examined. Mixing experiments in individual units as well as integrated experiments were performed to optimize the overall mixing performance Equipment A conical mill (Comil) manufactured by Quadro (Model # 197) (Figure 6-1) was utilized in this study. Since the mill was used for de-lumping purposes and not size reduction, it was operated using a round impeller (Model # 161), and using screens with round holes. The screen diameter was selected such that stagnation of the material in the mill was minimal, and hence primary particle breakage was avoided. However screens with large diameter cannot be used since their efficiency for de-lumping is insufficient. Screens with hole diameter of 6 and 8 µm were used in this study which provided good delumping efficiency.

139 Results Low-shear mixing The design of the impeller used in the Gericke continuous mixer is shown in Figure -1(b). The paddle type impeller imposes axial and radial flow on the powder. The shear environment in this blender can be qualified as low-shear since the powder in the high shear zone (the region between the impeller tip and the mixer shell) is a small fraction of the overall hold-up in the mixer, and the net movement of the powder is in the axial direction. The operational range of impeller speeds for this mixer lies between -3 RPM which correspond to tip speeds in the range of -15 cm/s. Under lower impeller speeds (4-1 RPM), powder is relatively less fluidized in the mixer and its flow behavior can be described as a powder bed stirred by the impeller. Under these conditions, the axial velocities lie in the range of.6-1 cm/sec, which indicates low shear environment. Under higher impeller speeds (16-5 RPM), powder bed is completely fluidized in the mixer, which leads to fewer particle-particle contacts, which again imposes low shear on the powder. The effectiveness of the Gericke mixer for de-lumping and mixing was examined by mixing the two representative materials, micronized APAP and Avicel- under 4,16 and 5 RPM impeller speeds. While the process is operating under steady state, samples were extracted at the mixer discharge for assessing the blend uniformity. Samples were subsequently analyzed by NIR spectroscopy to determine the content of APAP. Mixing performance, measured as the RSD between concentrations of the extracted samples, as a function of impeller speed is shown in Figure 6- (b). At the lowest speed (4 RPM),

140 13 worst mixing performance (RSD=.18) was obtained. With an increase in speed, RSD decreased to as low as ~.1. As described in the previous section, the analytical method error in the NIR measurement is ~.44. For the case of 1% APAP formulation, the inherent RSD for a well-mixed powder sample can be assumed to be ~.4. Powder is considered to be well mixed if the RSD between the sample concentrations is close to the analytical method error. In this case since the minimum RSD obtained was almost twice the minimum possible RSD, the performance of the mixer was considered to be suboptimal. Agglomerates of acetaminophen, visible to naked eyes were observed in the powder samples. Incomplete de-lumping or micro-mixing was thus witnessed in this case. In order to identify the contribution of incomplete axial mixing, a simulation of axial mixing was performed using equation 5. A convolution algorithm was run in MATLAB using RTD datasets at 4,16 and 5 RPM, and the incoming feed rate datasets. Simulation results showed that the contribution of incomplete axial mixing to the variability in concentration at the mixer discharge was minimal. As described in the previous section, radial mixing capability of the mixer is maximal under the intermediate rotation rates. Given that the mixer was operated under optimal macro-mixing conditions (axial mixing and radial mixing), poor micro-mixing was identified as the primary source of mixing variability High shear mixing The conical milling section and position of the impeller in the Comil is shown in Figure 6-1(b). The impeller imposes a radial flow on the material which makes it pass through the screen. The Comil was operated at 14 and 8 RPM, which correspond to a tip speed of 8 and 19cm/s. The extent of shear in co-mill is significantly greater than the

141 131 Gericke mixer since the material has to pass through the high shear zone before it exits the mill. The schematic of the experimental set-up for conducting mixing in the Comil is illustrated in Figure 6-3 (a). A full factorial DoE with two levels of impeller speed (14 RPM, 8 RPM) and two screen sizes (6, 8 µm) was conducted. Since the main function of co-mill in this case study is de-lumping and not size reduction, the screen size was chosen such that the hole diameter is significantly greater than the d9 of the particle size distribution of the powder blend (d9= ~ 4 µm). Continuous mixing in co-mill was facilitated by feeding individual powders using the same feeding setting as used for the Gericke mixer. As shown in Figure 6-3 (b), lower impeller speeds and smaller screens lead to better mixing performance. Quantitatively, the effect of screen size was relatively less than the effect of speed. Presence of screen significantly reduced agglomerates in the powder blend. Smaller error bars in the RSD measurement is also an indication of reduced number of agglomerates in the powder blend. However RSD values obtained over the operating range of the mill (.8-.15) being greater than the analytical method error (RSD=.4), mixing performance again was considered to be sub-optimal. In this case since the de-lumping/micro-mixing behavior was better than the Gericke mixer, poor macro-mixing was identified as the potential source of mixing variability. In order to qualify the macro-mixing capability of the Comil, residence time of the powder in the mill was measured. Residence time was measured by monitoring the steady-state powder hold-up in the continuous mixer. As shown in Figure 6-4, residence time in the mill increase from 1.4 s to 4. s as the impeller speed increases. This trend indicates that with an increase in speed, stagnation of the material in the mill increases,

142 13 possibly near the closed area of the screen. The fact that an increase in residence time in the mill does not necessarily decrease RSD clearly indicates that under high speeds there is a possibility of short-circuiting of the material. Very low mill speeds (< 14 RPM) were not used since de-lumping efficiency of the mill is less under lower speeds, and there is a possibility of preferential stagnation in the mill Integrated low shear mixing de-lumping process: Mixing experiments performed in individual units led to the conclusion that it was difficult to achieve good blend uniformity using either of the individual equipment. Mixing experiments were then performed in an integrated fashion which included two scenarios, namely low-shear mixing first and high-shear mixing first. In this set of experiments, process parameters of only the first unit were varied; the second unit was operated only under the optimal (best) operating condition Low-shear mixing first The schematic of the experimental set-up for the first scenario is shown in Figure 6-5 (a). Impeller speed of the Gericke mixer was varied from 4 to 5 RPM; the Comil was operated at a constant speed of 14 RPM. As shown in Figure 6-5 (b), RSD values slightly decreased after the milling stage. However RSD values were still higher than the minimum possible RSD. These results indicate that further mixing is necessary. Comil essentially de-lumps the remaining agglomerates from the incoming powder stream which creates high concentrations of APAP locally. There needs to be another mixing stage to mix the de-lumped APAP uniformly with rest of the powder. In this scenario,

143 133 since a post-low-shear mixing stage is missing, overall mixing performance remains suboptimal High-shear mixing first The schematic of the experimental set-up for the second scenario is shown in Figure 6-6 (a). In this case, speed of the Comil was varied from 14-8 RPM, and the Gericke mixer was operated only at 16 RPM. Mixing performance showed significant improvement after the second mixing stage; RSD values being less than the analytical method error indicates that the mixing performance cannot be optimized any further with the existing analytical method. A separate mixing experiment was performed by mixing the initially de-lumped material in a batch blender. De-lumped powder was blended for 3 min in a 8 Quart V-blender operating at 1.5 shell RPM. Similar degree of mixing was achieved; RSDs in the range of.-.4 were obtained. Since RSD values obtained after the batch mixing step were close to the analytical method error, further mixing is not required Conclusions Mixing behavior in a low shear mixer (Gericke) and a high shear Comil (Quadro) was examined. Gericke mixer was found to be a poor micro-mixer/de-lumper and a good macro-mixer, whereas the Comil was found to be a good micro-mixer and poor macromixer. Short-circuiting of the material was identified as the main source of poor macromixing in the Comil. Macro-mixing capability of the Gericke mixer and micro-mixing capability of the Comil was utilized by integrating them together. Integrated system with high shear mixing first provided the best possible mixing performance.

144 Figures for Chapter 6 Figure 6-1: (a) Conical mill (Comil - Quadro Model # 197) (b) Milling chamber with conical round impeller (a) (b)