Direct Tableting on a Continuous Manufacturing Line—Impact of Mixing Parameters, Material Densities, and Drug Load on Subsequent Process Parameters and Tablet Quality

Abstract

The continuous manufacturing (CM) of solid oral dosage forms has received increased attention in recent years and has become a leading technology in the pharmaceutical industry. A model has been developed based on process data from two design of experiments (DoEs), where the impact of the mixer process parameters, throughput (THR), hold up mass (HUM), impeller speed (IMP), and the input raw material bulk density (BD_i), on the continuous process and the resulting drug product has been investigated. These statistical models revealed equations, describing process parameter interactions for optimization purposes. For the exit valve opening width (EV) at the bottom of the continuous mixer (CMT), the combination of high throughput (30 kg/h) and low impeller speed (300 rpm) resulted in optimal process conditions. Apparent bulk density of the blend (BD) within the process, fill depth (FD), and tensile strength (TS) were mainly impacted by input bulk density (BD_i) of the tableting mixture, emphasizing the role of material attributes on the continuous manufacturing process. The apparent bulk density itself was, other than from the input bulk density, equally dependent from THR and IMP in opposite deflections. However, process parameters (THR and IMP) revealed a minor impact on the apparent BD compared to the input bulk density. FD was impacted mainly by THR ahead of IMP and the TS by IMP and THR to a similar extend, in opposite deflections. A simplified linear model to estimate the input bulk density revealed satisfactory prediction quality when included in the derived statistical model equations.

Introduction

Continuous manufacturing of solid oral dosage forms has received more and more attention in recent years. The benefits of a continuous manufacturing line are widely mentioned in the literature, where advantages like smaller equipment footprint, fewer process steps, easy scale-up, optimized process control, faster manufacturing, and flexibility and cost reduction are mentioned [1,2,3,4,5,6].

To exploit these benefits, a broad understanding of the process is required. That includes the function of the machine itself, the dependencies and relationships between the processing parameters, and the impact of the material attributes on the process. For this study, the PCMM (portable, continuous, miniature, and modular) commercial production line was used, which consists of six GEA Compact Feeders, a vertical continuous blender (CMT—continuous mixing technology), and a MODUL™ P tablet press, equipped with a NIR Probe located in the feed frame [7].

The vertical blender allows for independent adjustment of the hold up mass (HUM) and impeller speed (IMP) process parameters, as well as the overall line mass throughput. These three parameters are CMT process parameters studied in this work and were used as input factors. Based on these parameters, the mean residence time (MRT, Eq. (1)) and the total blade passes (TBP, Eq. (2)) of the CMT can be calculated. MRT reflects the average time a particle spends within the CMT, and the TBP indicates how many blade passes are seen by the average particle, thus indicating the amount of shear exerted on the powder [8].

$$\mathrm{MRT }\left[\mathrm{min}\right]=\frac{\mathrm{HUM }[\mathrm{kg}]}{\mathrm{ THR }\left[\frac{\mathrm{kg}}{\mathrm{h}}\right]}*60\frac{\mathrm{min}}{\mathrm{h}}$$

(1)

$$\mathrm{TBP}=\mathrm{MRT }\left[\mathrm{min}\right]*\mathrm{IMP}[\mathrm{rpm}]$$

(2)

A basic process understanding is required if commercial products are manufactured by CM. From time to time, it is possible that expected process parameter values like fill depth (FD) or tensile strength (TS) can differ from one batch to another, even if the upstream setting parameters remain constant. These differences could be caused by variability in ingoing material properties such as density. That is why it is important to understand the impact of the raw material properties [9] and to consider lot-to-lot variability [10] during the development stages [11] in order to define a robust design space for the process. Stauffer et al. described the influence of active pharmaceutical ingredient (API) on the drug product, implicating an influence of API attributes on the process [9]. Consequently, a possible reason for density variabilities could be an adapted synthesis route of API [9] or moisture uptake of the raw materials [9, 10].

The variability in density could be seen in PCMM by fluctuating fill depths, where lower blend densities require higher fill depths to meet the target weight of the tablet.

To combine all raw material densities, a theoretical input density was used, which is explained in the “Input Bulk Density” section. Therefore, this study considered the bulk density of the raw materials as another input factor of the model to examine the fundamental impact of the raw material densities on the process behavior regarding apparent bulk density of the blend (BD), exit valve opening width (EV), fill depth (FD), and tensile strength (TS).

To evaluate if the combination of CMT parameters is compatible during the process, the exit valve could be identified as the first response to be monitored. The exit valve opening width is controlled by a PID (proportional–integral–derivative) controller to ensure that mass rate_in equals mass rate_out while the hold up mass remains at or near target [7, 12]. Smaller exit valve opening widths are recommended so that newly entering raw materials can be properly mixed with the blend that is already present in the blender. Otherwise, unmixed or poorly mixed material can pass by and leave the CMT without being blended, causing content-uniformity variability [8, 12].

Since the combination of CMT settings impacts the extent of lubrication, it could be seen that higher lubrication is associated with a higher apparent density of the blend, lower fill depths, and lower tensile strength [12, 13]. Furthermore, Bos et al. published that higher raw material density increased the lubrication sensitivity and therefore the impact on tensile strength [14].

This study combines both aspects, where the impact of mixer parameters and the variation of raw material densities are considered governing input parameters for the lubrication and the continuous process.

In the literature, continuous mixing is mostly associated with horizontal mixers, where the hold up mass is considered a function of flow rate and impeller speed and cannot be set individually [15, 16]. Regarding lubrication, it is reported that higher impeller speed resulted in a shorter residence time of the powder and, hence, a higher amount of total blade passes appear at moderate impeller speeds [17]. Using the vertical blender used in this study, the selection of throughput, hold up mass, and impeller speed needs to be approached differently. Since the setting parameters are now independent of each other, the impact on the lubrication and subsequent process parameters needs to be reevaluated.

The scope of the study was to examine the impact of varying CMT settings in a vertical blender, alternating proportion of the raw materials in the formulation and differences in raw material properties of the excipients (i.e., density) on continuous process parameters. Thus, this study investigates the influence of CMT parameters in a vertical blender in combination with the theoretical input density of each blend on downstream process stability (exit valve opening width and fill depth) and the resulting material attributes (apparent bulk density of the blends and tensile strength).

The benefit of obtaining such models is that the reaction time to changes regarding the input parameters can be significantly reduced and the optimum process conditioned can be reached faster indicating a gain of yield during a production.

Materials and Methods

Materials

For the trials, saccharin sodium monohydrate (JMC, South Korea), microcrystalline cellulose (MCC; Avicel PH 102, FMC, Cork, Ireland), dicalcium phosphate (DCP; A-Tab, Innophos, Chicago Heights, USA), sodium starch glycolate (SSG; Roquette, France), and magnesium stearate V (MgSt; Mallinckrodt, St. Louis, USA) were used. As a model formulation, saccharin monohydrate was used as an API surrogate.

Test Settings and Model

Database

For this work, data of two separately conducted DoEs were combined and used to develop a new model by means of MODDE Pro 12.1 (Sartorius Stedim Data Analytics AB, Umea, Sweden).

The first DoE, hereinafter referred to as DoE 1, included 17 runs and was a central composite face design with star points at the face of each side defined by a 2-level factorial design (see Table I). As input factors, throughput (THR), hold up mass (HUM), and impeller speed (IMP) were set. Furthermore, a compression pressure/tensile strength profile was conducted using CPs of 118, 157, 169, 236, and 275 MPa.

Table I Settings of Both DoEs Where THR, HUM, and IMP and THR, IMP, and DL Were Used as the Input Factors

The second DoE, referred to as DoE 2, included 18 runs and was a full factorial DoE, where the saccharin concentrations, referred to as drug load (DL), THR, and IMP, were used as input factors (see Table I). In this trial, HUM and compression pressure (CP) were kept constant at 800 g and 275 MPa.

The following table presents the settings of both DoEs, where for the first experimental design, batches 7, 9, and 11 and, for the second DoE, the batches 8, 10 and 12 were the replicates of the center points. Additionally, DoE 2 was extended by 3 more batches (numbers 16, 17, and 18) with 5%, 25%, and 30% DL.

Input factors for each DoE and the composition of each raw material and therefore each drug load based on the saccharin monohydrate composition are shown in Table II and Table III.

Table II Input Factors of the Initial DoEs

Table III Used Materials and Corresponding Compositions in Percent of Both DoEs

The ranges regarding throughput were set between 10 and 30 kg/h, since lower throughputs than 10 kg/h would contradict the actual goal of an efficient continuous manufacturing and 30 kg/h is the maximum throughput of the CMT of this manufacturing line. Regarding hold up mass, less than 400 g can result in insufficient mixing due to a higher potential of powder fluidization especially at higher impeller speeds. More than 800 g leads to a volume constraint resulting in contact between the blend and the sieve in the upper part of the CMT causing weighing disturbances. The maximum impeller speed was set to 650 rpm in DoE 1, but due to high exit valve opening widths, it was reduced to 550 rpm for DoE 2 (see also the “Statistical Model Equations Resulting from DoE 1 and DoE 2” section). The medium impeller speed was adjusted, accordingly. To evaluate the blend uniformity, low drug loads were required, and therefore, 2% was set as minimum. It was decided to increase the drug load by 8% to keep the distance between the levels relatively small. To gain a better overview, especially for smaller drug loads, 5% was additionally set; 25% and 30% were used as reference batches.

Continuous Manufacturing of Each Batch Within Both DoEs

Each raw material was fed using an individual feeder. After one initial batch start-up, the process was running continuously. For both DoEs, a transition phase (3 × mean residence time) was initiated after new CMT settings were adjusted until a steady state was reached, which was then maintained for at least 10 min. A MODUL™ P tablet press (GEA Pharma Systems, Courtoy™, Halle, Belgium) was implemented at the end of the continuous manufacturing line. Mode 2 (Courtoy dual control force method) was selected, where the tablet weight control is based on pre-compression displacement measurements, adjusting the fill depth accordingly [18].

For both DoEs, convex, round tablets with 11 mm diameter and 1.12 mm cup height were manufactured. During steady state, a target compression pressure of 275 MPa and a pre-compression pressure of 2.1 MPa were set.

The target tablet weight of 600 mg, tablet crushing strength, and tablet thickness were tested periodically in the middle of each steady state using the at-line automated tablet testing system UTS IP65i (Kraemer Elektronik GmbH, Germany).

The feed chute level was controlled to a constant level at 40%, and the paddle speed remained constant at 45 and 40 rpm at each conducted setting in both DoEs. Turret speed set points and speed tolerances of the tablet press were adapted to the respective mass throughput (11 rpm ± 2.2 rpm; 21 rpm ± 4.2 rpm, and 32 rpm ± 6.4 rpm).

After each steady state, a powder sample of the blend was withdrawn by opening the sampling port underneath the feed frame and collecting approximately 300 g of powder. Subsequently, the samples were measured by the FT4 powder rheometer (Freeman Technology Inc., Worcestershire, UK).

The tensile strength of the convex, round tablets was calculated based on the following equation [19]:

$$\mathrm{Tensile}\;\mathrm{Strength}=\frac{10P_s}{\pi D^2}\left(2.84\frac tD-0.126\frac tW+3.15\frac WD+0.01\right)^{-1}$$

(3)

where P_s is the tablet core crushing strength, D is the tablet core diameter, t is the tablet core thickness, and W is the cylinder length. Tablet crushing strength was measured using the automated tablet testing system, which is directly connected to the continuous manufacturing line.

Normalization of the Raw Material Bulk Density

To normalize the different proportions of mixture components in the formulation and to evaluate the influence of variations in bulk density of the raw materials, a theoretical value, the input bulk density (BD_i) was introduced, which is calculated as shown in Eq. (4). This value represents the theoretical density of the blend based on the composition and bulk density of each raw material. For that, a powder sample of each raw material was taken at the beginning of the process and analyzed by the FT4 Freeman powder rheometer using the stability and variable flow rate method. The obtained conditioned bulk density was used to calculate the input bulk density. Further explanations are shown in the “Input Bulk Density” section.

$${\mathrm{CBD}}_{i}=\frac{\sum {\mathrm{Composition}}_{n}*{\mathrm{CBD}}_{n}}{100}$$

(4)

In this equation, i stands for input and n represents each individual raw material. Furthermore, the input factors and responses considered in this work are shown in Fig. 1. A linear model was used to evaluate the significant terms and included 115 tensile strength values and 35 values for each bulk density, exit valve opening width and fill depth. In order to assess the quality of the models, Q² (estimate of the future prediction precision) and R² (model fit) obtained by MODDE were considered. In addition, a k-fold crossvalidation was carried out where R² and RMSE were calculated.

Fig. 1

Visualization of the input factors and responses considered in the model of this study

k-Fold Crossvalidation

In general, a k-fold crossvalidation can be used to evaluate the quality of a predictive model. Therefore, the dataset is split into k-folds of approximately equal size. To train the model, k-1 datasets are used to build the model, and the remaining one is used as a testing set to validate it. This procedure was repeated until each dataset was used as either a training or testing set [20].

In this work, a fivefold crossvalidation was used to evaluate the quality of the obtained models. Therefore, for each response (EV, BD, FD, or TS), the dataset was randomly partitioned into five equally sized subsets.

Using MODDE, for each fold, a model was built with the four training sets. The obtained models were then applied to the respective testing sets. To evaluate each model, R² and RMSE of the predicted and the observed values were used. R² was obtained by GraphPad Prism 9 (GraphPad Software, Inc., USA), and RMSE was calculated according to the following equation [21]), where P_i stands for the predicted values from the model, O_i for the observed values in the dataset, n for the sample size of the dataset, and i for the fold:

$$\mathrm{RMSE}=\sqrt{\frac{\sum_{i=1}^{n}{({P}_{i}-{O}_{i})}^{2}}{n}}$$

(5)

To evaluate the performance of the model, average values of R² and RMSE values of the training and testing sets were calculated.

Freeman Powder Rheometer FT4

Stability and Variable Flow Rate

Using the Freeman powder rheometer FT4, a cylindrical 25 mm × 25 ml split vessel was used. After an initial condition cycle, the powder was split to obtain a defined amount of powder to ensure reproducible measurements. The actual testing consists of seven alternating conditioning and test cycles where the blade is inserted in the powder bed and moved downwards with a rotational blade tip speed of 100 mm/s to remove history and operator influence. Subsequently, 4 cycles with decreasing blade tip speed (100 mm/s, 70 mm/s, 40 mm/s, and 10 m/s) were performed.

The stability and variable flow rate method was used in this study to evaluate the bulk density of the raw materials and the blends.

The bulk density (BD) of the raw materials and the blends is measured after the initial conditioning cycle and the powder split, where agglomerates and air inclusions can be evened to ensure reproducible measurements [22].

Results

Statistical Model Equations Resulting from DoE 1 and DoE 2

DoE 1 focused on the impact of throughput, impeller speed, and hold up mass on the continuous process, where the model Eqs. (6)–(9) and corresponding ANOVA results (supplementary data, Tab. S 1.) could be obtained. DoE 2 included the drug load additionally to THR and IMP. For this model, Eqs. (10)–(13) could be obtained. The corresponding ANOVA results are shown in the supplementary data, Tab. S 1.

$${\mathrm{Log}}_{10}\left(\mathrm{EV}\right)=0.125741+0.020304*\mathrm{THR}-0.00164767*\mathrm{IMP}+{4.37955*10}^{-6}*{\mathrm{IMP}}^{2}$$

(6)

$$\mathrm{BD}=0.558553-0.00113*\mathrm{THR}+{3.54999*10}^{-5}*\mathrm{HUM}+{5.2*10}^{-5}*\mathrm{IMP}$$

(7)

$$\mathrm{FD}=9.63996+0.0701824*\mathrm{THR}+0.000317503*\mathrm{HUM}-0.000564445*\mathrm{IMP}-0.000811434*{\mathrm{THR}}^{2}-{2.18752*10}^{-5}*\mathrm{THR}*\mathrm{HUM}$$

(8)

$${\mathrm{Log}}_{10}(\mathrm{TS})=0.364822+0.017798*\mathrm{THR}-0.000162112*\mathrm{HUM}-0.00030635*\mathrm{IMP}-0.000204041*{\mathrm{THR}}^{2}-{3.32993*10}^{-6}*\mathrm{THR}*\mathrm{HUM}-{9.37246*10}^{-6}*\mathrm{THR}*\mathrm{IMP}+{3.0309*10}^{-7}*\mathrm{HUM}*\mathrm{IMP}$$

(9)

$${\mathrm{Log}}_{10}\left(\mathrm{EV}\right)=-0.00288867+0.00810256*\mathrm{THR}+0.000567753*\mathrm{IMP}+0.023019*\mathrm{DL}+{3.38677*10}^{-5}*\mathrm{THR}*\mathrm{IMP}-{2.08147*10}^{-5}*\mathrm{IMP}*\mathrm{DL}$$

(10)

$$\mathrm{BD}=0.49869-0.00109062*\mathrm{THR}+{6.90476*10}^{-5}*\mathrm{IMP}+0.00309795*\mathrm{DL}-{2.34376*10}^{-5}*\mathrm{THR}*\mathrm{DL}$$

(11)

$$\mathrm{FD}=11.3245+0.0379166*\mathrm{THR}-0.00107143*\mathrm{IMP}-0.0633723*\mathrm{DL}$$

(12)

$$\mathrm{TS}=3.40964+0.0625729*\mathrm{THR}-0.00142559*\mathrm{IMP}-0.0252941*\mathrm{DL}-{5.00001*10}^{-5}*\mathrm{THR}*\mathrm{IMP}-0.00114062*\mathrm{THR}*\mathrm{DL}+{4.01786*10}^{-5}*\mathrm{IMP}*\mathrm{DL}$$

(13)

Only for DoE1–exit valve opening width, the ANOVA revealed a significant lack of fit. This can be traced back to the distribution of the exit valve data, where the majority of obtained opening widths were below 5 mm and only a few deviated with values > 30 mm vastly. Therefore, for DoE 2, impeller speed was reduced to avoid such high EV values and, hence, to improve the model.

Input Bulk Density

To combine the findings of both DoEs and extend the input information by the raw material density, a normalization in terms of the input bulk density (BD_i) was added as additional model factor.

This study used the BD_i as a calculated input factor to show the fundamental influence of density on the process, to evaluate to what extent lot-to-lot variations regarding density of the used raw materials or changes in the composition of the formulation impact the downstream process in a routine production. As shown in Eq. (4), each raw material and the corresponding density, as well as the fraction within the formulation, were included. That means, for each used drug load (DL), a different BD_i could be calculated (Table IV). The bulk density of the raw materials is shown in Table IV.

Table IV Bulk Density of the Raw Materials and the Theoretical Density (BD_i) Depending on the Drug Load Settings and BD of Each Individual Raw Material, Calculated According to Eq. (4)

As observed in Fig. 2 and Table V, there is a linear relationship between the measured apparent BD of the blends and the calculated BD_i of the blends. For demonstration, blends mixed at high (2650 rev), medium (~ 920 rev), and low total blade passes (TBP) (320 rev) are depicted in this figure. The dataset of ~ 920 TBP includes blends mixed at 880–960 rev. It confirmed the linear relationship between the additively calculated BD_i and the apparent BD of the blends even if the weight fraction of raw materials with varying particle sizes are included and hence volume fractions of the constituents of the mixture are likely to deviate from their respective weight fraction. So, the use of BD_i is a suitable approach to estimate the tableting mixture’s apparent BD.

Fig. 2

Apparent BD samples taken from the feed frame and analyzed by the FT4 powder rheometer in dependence of the calculated input BD. Green, round data points show the BD of the blends mixed at 2650 rev; the blue, squared data points show the data of blends mixed between 880 and 960 rev; the red, triangle data points represent the BD obtained by 320 rev

Table V Fit Statistics of the Linear Regression of the Datasets Shown in Fig. 2

So, a simplified model could be built that explains the fundamental impact of raw material densities on continuous process parameters, as further explained in the following chapter.

Impact of Input Variables on Bulk Density of the Blend, Exit Valve Opening Width, Fill Depth, and Tensile Strength

The impact on the process stability (EV and FD) as well as the resulting material attributes (apparent BD of the blends and TS) was evaluated against a change in input process parameters (throughput, impeller speed, and hold up mass) in combination with varying composition of the formulations (incorporated as material attribute BD_i).

The data were fitted using an MLR model, where significant model terms are identifiable when error bars (= 95% confidence interval) do not cross the zero line otherwise; these model terms are non-significant. The obtained high values of Q² and R² indicate a significant correlation between input variables and responses. Furthermore, a k-fold crossvalidation was carried out to provide additional information on the applicability of the models.

The model terms for bulk density of the blend, exit valve opening width and fill depth are shown in Fig. 3. For the bulk density, it is revealed that high input bulk density (BD_i), high impeller speed (IMP), low throughput (THR), and high hold up mass (HUM) resulted in high apparent densities of the blend. Furthermore, the EV was significantly reduced by low IMP and low THR, and the FD was impacted by BD_i, THR, and IMP.

Fig. 3

Impact of the input density (BD_i) and the CMT parameters on the apparent BD of the blend measured by the FT4, EV, FD, and on the TS of the tablets. The 95% confidence interval is displayed as an error bar

For tensile strength, the model terms revealed that lower BD_i, higher compression pressure (CP), lower IMP, higher THR, and lower HUM increased the tablet tensile strength.

To predict the process parameters, the following equations could be used, where only significant parameters were included. Corresponding fit statistics and p values are shown in the supplementary data, Tab. S1 .

$$\mathrm{BD}=0.574402-0.0011625*\mathrm{THR}+0.0119951*\mathrm{IMP}+0.00769962*\mathrm{HUM}+0.117658*{\mathrm{BD}}_{\mathrm{i}}$$

(14)

$${\mathrm{log}}_{10}\left(\mathrm{EV}\right)=0.739238+0.203871*\mathrm{THR}+0.446307*\mathrm{IMP}$$

(15)

$$\mathrm{Fill Depth}=10.4146+0.268194*\mathrm{THR}-0.268194*\mathrm{IMP}-0.0458841*\mathrm{HUM}-1.51*{\mathrm{BD}}_{\mathrm{i}}$$

(16)

$$\mathrm{log}(\mathrm{TS})=0.174086+0.042137*\mathrm{THR}-0.0583895*\mathrm{IMP}-0.0230503*\mathrm{HUM}-0.351427*{\mathrm{BD}}_{\mathrm{i}}+0.115393*\mathrm{CP}$$

(17)

Discussion

Apparent Bulk Density of the Blend

As expected, the BD_i revealed the highest impact on the apparent BD of the blend (Fig. 3). That means that with higher initial densities of the raw materials, the density of the blend was higher at the same CMT parameter settings. This should be considered if a different grade of raw material is used [9], lot to lot variabilities are to be expected [10], or if the drug load or the composition is changed in early development stages [11]. Furthermore, the knowledge about the impact of BD_i is increasingly important if sequential addition of API is considered, where two different API lots are used in the same production run. In this case, the obtained model may help to estimate if the lubrication sensitivity of the blend may be impacted, as described by Bos et al. or how the fill depth needs to be adjusted after the second API is introduced into the system (see also the “Fill Depth” section) [14].

The total blade passes are the combination of impeller speed and mean residence time (Eq. (2)), governing shear and mixing intensity of the lubricant into the blend.

Higher total blade passes represent more contact between impeller and powder particles, and hence, it is implied that more impeller rotations can result in a potential risk of film formation and over-lubrication. That blend lubrication reduces particle–particle friction, and the powder can arrange more compactly [23, 24]. Regarding the individual model terms, the impact of impeller speed and throughput were deflected in opposite directions but shared almost the same extent of impact, whereas the influence of HUM was less pronounced.

For visualization, a response surface plot is shown in Fig. 4a, where both significant model terms with the highest effect on the bulk density are included. The input bulk density and impeller speed are displayed on the x- and z-axes, and the bulk density of the blend is shown on the y-axis. To illustrate the mentioned findings, the surface area presents the linearly increasing bulk density of the blend with rising input density. Opposed, the impeller speed revealed only minor input where the apparent BD increased less than 0.05 g/ml with increasing rotational speed between 200 and 650 rpm.

Fig. 4

Response surface plots of the apparent bulk density of the blend, tensile strength, exit valve opening width, and fill depth

Regarding the mixing parameters, the model is primarily applicable for vertical blenders, where the lubrication is influenced by the total blade passes as calculated by Eq. (2).

For horizontal blenders, the higher number of blade passes are at medium impeller speeds since hold up mass and impeller rotational speed are opposed and not proportional as in a vertical blender [15, 16]. Furthermore, Furukawa et al. found that increasing apparent BD of the blend increased the mean residence time of powder in the feed frame and the residence time distribution (RTD), which is beneficial to improve the blend uniformity if potential fluctuations regarding API concentration occurred in the continuous process. This model follows the advice to assess variations in bulk density, since it is necessary for a robust control strategy, especially, when NIR is in place [25].

To evaluate the predictive model, a crossvalidation was carried out, where the 35 settings of the DoEs were partitioned into 5 datasets of 7 settings each. As described earlier, 4 datasets were used as training sets, and the remaining one was used to validate the model. Therefore, for each fold, R² and RMSE were calculated. The averaged R² and RMSE of the 5 folds are shown in Table VI and present a good model to predict the apparent bulk density of the blend based on the CMT parameters and the input density of the raw materials.

Table VI Average Values of R² and RMSE Obtained by the k-Fold Crossvalidation (k = 5)

Figure 5a shows the predicted and the observed values of the bulk density of the blends based on the obtained model where all 35 settings were included. The dotted line theoretically represents R² = 1.

Fig. 5

a The predicted apparent BD values in comparison to the observed values of the dataset, where R² = 0.926 and RMSE = 0.008 could be calculated. b The predicted TS values in comparison to the observed values of the dataset, where a R² = 0.942 and RMSE = 0.171 could be calculated. c The predicted EV values in comparison to the observed values of the dataset, where R² = 0.800 and RMSE = 5.011 could be calculated. d The predicted FD values in comparison to the observed values of the dataset, where R² = 0.756 and RMSE = 0.270 could be calculated. A reference line (dotted line) was inserted in each graph to demonstrate a theoretical perfect fit

Tensile Strength

Figure 3 reveals the impact on tensile strength by the combination of process parameters (throughput, hold up mass, and impeller speed), material attributes (BD_i), and the compression pressure (CP).

By far, the highest effect was obtained by the BD_i. As already described in the previous section, in a routine production in CM, the necessity to combine two different API lots through a sequential API addition may occur. If these API lots differ in density, it impacts the tensile strength significantly. The potential reason for this comparably high impact may be again the increased lubrication sensitivity by the higher density [14].

As expected, another significant input parameter is the compression pressure, where higher pressure values resulted in higher tensile strength of the tablets, which is also seen in the literature [26, 27].

Furthermore, an increase in impeller speed and a reduction in throughput led to higher total blade passes, which indicates a higher extend of lubrication within the blend and resulted in lower tensile strengths, which is also well-documented in the literature [12, 28,29,30].

To distinguish between the three CMT parameters, it was shown that changes in impeller speed impacted the TS the most ahead of throughput and hold up mass, i.e., in a vertical continuous blender, higher impeller speed impacted the shear and, therefore, the lubrication the most. On the other hand, with regard to horizontal continuous blender, a higher amount of total blade passes appears at moderate impeller speeds, and therefore, the obtained model is primarily valid for vertical mixers [17].

A response surface plot is shown in Fig. 4b, where impeller speed and input bulk density are displayed as input factors and TS as the response. Whereas a change of BD_i from 0.4 to 0.7 g/ml resulted in a vast rise of TS (~ 1 to 4 MPa), an increase of impeller speed from 200 to 650 rpm impacted the TS only little. Assuming a tensile strength specification between 1.5 and 2.5 MPa (turquois and green color layer), it seems that compliance with the limits is mainly dependent on the BD_i, where densities between 0.45 and 0.55 g/ml were required independent of impeller speed. Furthermore, the impact of impeller speed increases at lower BD_i values, whereas at 0.7 g/ml, the reduction of impeller speed from 650 to 200 rpm only increased the TS by approximately 0.5 MPa; the same variation in impeller speed increased the TS from 3.7 to 4.7 MPa at 0.4 g/ml.

The model facilitates the assessment of the potential impact on the tensile strength and improves the process understanding if lot-to-lot variations are to be expected or adjustments regarding impeller speed, throughput, or hold up mass are conducted.

For tensile strength, a k-fold crossvalidation was conducted, where 115 DoE settings were partitioned into 5 folds of 23 settings each. Average values of R² and RMSE were used to evaluate the model (Table VI).

Figure 5b shows the predicted and the observed tensile strengths based on the obtained model where all 115 settings were included. The dotted line theoretically represents R² = 1.

Exit Valve Opening Width

Figure 3 shows the impact of the input parameters on the exit valve opening width, where higher impeller speed and throughput both resulted in higher exit valve values, which aligns with our previous findings [12]. For vertical continuous blender, low opening widths are necessary to provide a processable mixing condition [8].

For more clarity, a response surface plot is shown in Fig. 4c, where both significant model terms throughput and impeller speed are displayed. For that, on the x-axis, the throughput, on the z-axis, the impeller speed, and on the y-axis, the corresponding exit valve opening width are shown. On the surface, 5 contour levels are illustrated to highlight the opening width of the exit valve (5 mm (blue), 10 mm (turquois), 15 mm (green), 20 mm (yellow), and above (red)). Since both model terms are positively deflected, the opening increased with higher IMP and higher THR, so a leverage effect is created when both parameters are increased. Since the effect of IMP is higher than that of THR, slight increases in throughput leveraged the impact of IMP. Therefore, at low throughputs (10 kg/h), impeller speeds up to 500 rpm were still processable to maintain an EV smaller than 5 mm, whereas with higher throughputs (30 kg/h), the process window for the impeller speeds decreased and only smaller settings than 300 rpm are possible to ensure low EV values.

To use the efficiency of the continuous manufacturing equipment, the process needs to be run at its maximum capacity regarding throughput. In consideration of maintaining low EV, the optimal process conditions are 30 kg/h throughput and low impeller speeds, preferably under 300 rpm. This model can be used to determine the optimum vertical mixer parameter settings to obtain an optimal process condition in terms of low opening widths of the valve.

To evaluate the predictive model for the exit valve opening width, a crossvalidation was conducted, where the same folds as for the bulk density validation was used, i.e., the 35 settings were split into 5 sets of 7 data settings each. Table VI shows the averaged values of R² and RMSE of the 5 folds. The high RMSE values can be explained by the large range of exit valve opening widths obtained throughout the DoEs (1.47–44.28 mm), where the model did not show accurate results at higher values (Fig. 5c).

Another approach is the normalization of the CMT parameters (CMT_Norm) hold up mass (HUM), impeller speed (IMP), and throughput (THR) to the following term:

$${\mathrm{CMT}}_{\mathrm{Norm}}=\frac{\mathrm{HUM}}{{\mathrm{IMP}}^{2}*\mathrm{THR}}$$

(18)

This term can help to assess the impact of the CMT process parameters on the exit valve opening width. Figure 6 shows the EV as a function of this term for 10, 20, and 30 kg/h. The EV data are the calculated results of the model equation, where input bulk density (BD_i) was randomly set to 0.536 g/ml, IMP varied between 200 and 700 rpm as highlighted in the legend of the graphs, and HUM was used between 400 and 800 g. For example, the black data points in Fig. 6a are the calculated EV values at 200 rpm. The data point on the right side is obtained at 400 g and the one on the left at 800 g. With increasing IMP, the values shift further to the left and therefore to smaller values of ${\mathrm{CMT}}_{\mathrm{Norm}}$. Simultaneously, the EV values increase at a certain point. In this case, the displayed data points for each IMP are a straight line, dependent on HUM. The slopes of these lines increase with higher IMP and THR. Normalizing the input parameters can help visualize the ranges in which IMP and THR can be adapted to the continuous manufacturing process maintaining similar and low EV.

Fig. 6

Exit valve opening width predictions at a 10 kg/h, b 20 kg/h, and c 30 kg/h. For better comparability, the x- and y-axes remained equal for the three throughputs

Fill Depth

To decrease the amount of time and material needed to tune the tablet press to the required tablet weight specifications, Eq. (16) can be used to predict the fill depth value to fulfil the requirements. Since FD and bulk density of the blend strongly correlate (− 0.913, p < 0.0001), the impact of BD_i must clearly be considered. As shown in Fig. 3, the input density represents the highest impact on FD. Furthermore, as explained in the “Apparent Bulk Density of the Blend” section, higher total blade passes resulted in higher powder densities of the blend, and therefore, lower required FD settings. The deflection of the displayed effects of throughput, impeller speed, and hold up mass is equivalent to the total blade passes equation, where higher throughput, lower impeller speed, and lower hold up mass resulted in low total blade passes, low apparent bulk density of the blend, and higher fill depths.

Regarding the individual CMT parameters, throughput showed the highest impact, since it is both impacted by the turret speed and the corresponding die filling time and the total blade passes expressed as the bulk density [31].

Therefore, a response surface plot was used to display the impact of throughout and input bulk density, the two significant model terms with the highest impact on the FD. Comparable to the apparent BD of the blend, the fill depth increased linearly with rising input BD. Only at low and high input bulk density values, the impact of throughput is visible, where, e.g., the FD increased by approximately 0.75 mm when throughput was increased from 10 to 30 kg/h.

Exemplarily, if a fill depth of 10.5 mm is required to meet the tablet weight specification, combinations of 0.52 g/ml and 10 kg/h or 0.57 g/ml and 30 kg/h were required.

An optimal setting regarding fill depth is required since tablet weight is directly affected by the die filling. Since throughput is beside the total blade passes, an adjustment of the turret speed as well as the information of this model can be applied to other continuous blenders and batch processes.

Again, the fivefold crossvalidation was conducted for FD, where the same folds as for BD and EV were used. Table VI shows the averaged values of R² and RMSE of the 5 folds.

Figure 5d shows the predicted and the observed values of the FD based on the obtained model where all 35 settings were included. The dotted line theoretically represents R² = 1 indicating a good model fit.

Conclusion

Using a vertical continuous blender in the continuous manufacturing process, the adjustments regarding throughput, hold up mass, and impeller speed must be carefully considered, which is why this paper evaluated the process stability (exit valve opening width of the CMT and fill depth of the tablet press) and the resulting material attributes (apparent bulk density of the blends and tensile strength of the tablets) based on throughput (THR), hold up mass (HUM), impeller speed (IMP), and input bulk density (BD_i) in a continuous direct compression line using a vertical blender (CMT).

To provide good process conditions and ensure blend homogeneity, the vertical blender specific exit valve (EV) needs to operate at low opening widths. In this model, the EV was significantly increased by high throughput and high impeller speed, where optimal process conditions were achieved at 30 kg/h and impeller speeds lower than 300 rpm.

At the same time, the apparent bulk density itself was, other than from the input bulk density, equally dependent from THR and IMP in opposite deflections. However, process parameters (THR and IMP) revealed a minor impact on the apparent BD compared to the input bulk density. Since the impact of the input bulk density is dominant, the findings can be adapted to other manufacturing processes.

It has been shown that bulk densities (BD and BD_i) had naturally the highest impact on fill depth. Compared to the apparent bulk density of the blend, differences regarding the individual CMT parameters could be revealed. Where THR und IMP were almost equally affecting the apparent BD, the THR showed much higher influence on the fill depth than the IMP. These results can be transferred to build models for horizontal blenders or batch processes.

The tensile strength (TS) was primarily affected by the input density and the compression pressure. Considering the CMT parameters, the shear introduced by the impeller speed showed the highest influence on the total blade passes and the lubrication, which reduced the TS.

A simplified linear model to estimate the input bulk density revealed satisfactory prediction quality when included in the derived statistical model equations.

Using these models, the continuous manufacturing process can be optimized.