Introduction

Since the issuance of process analytical technology (PAT) guidelines by the FDA in 2004, the widespread usage of PAT has been permeating every corner of process development, encompassing small and large molecules, as well as drug substance and drug product processes. Among the many successful case studies, the small molecule powder blending process has been one of the most commonly reported areas of PAT applications in drug product process development. Literature surveys suggest that the use of PAT to determine blending end-point has been the primary objective (1,2,3,4,5,6), but with relatively little attention paid onto the use of PAT tools to characterize blend heterogeneity at steady state as a means to support formulation and process development.

Among many reported PAT applications, a three-pronged purpose can be summarized in pharmaceutical process development: (1) providing real-time process monitoring to improve process understanding, (2) enabling in-process control via real-time process monitoring, and (3) fulfilling real-time release testing. These applications become even more relevant when a clinical asset is transitioning to drug product manufacturing, especially during late-stage process development. By contrast, the use of PAT is relatively limited in early R&D (early CMC or pre-CMC), as the manufacturing scale and the number of clinical campaigns are relatively small. For example, although the endpoint determination of the blending process is a common PAT practice at the pilot scale, a powder blend can possess inherent heterogeneity after the mixing endpoint is reached (7,8,9). Such information may not be readily apparent from a routine process monitoring on blending end-point via NIRS. It is therefore desirable to explore the employment of PAT tools to assist the formulation and process understanding at the early stage of the product development and to potentially forecast content uniformity downstream and thus inform late-stage process development.

As companies are gradually engaged in the continuous direct compression (CDC) process for drug product development and manufacturing, characterization of inherent blend heterogeneity becomes increasingly critical. The inherent blend heterogeneity refers to the variability of the constituent concentration in the mixture at state of control. It could have a direct impact on the outcome of product quality for CDC processes, and it is critical to understand the potential of inherent blend heterogeneity and demixing potential of CDC powder blends, preferably during the early stage of formulation development and at small scales. This is because the final drug products arising from the DC or CDC process could be prone to sub-optimal content uniformity, largely due to the lack of the granulation step serving to immobilize API with other constituent particles (10). Additionally, for low-dose drug products, employment of an API with coarse particle sizes or a broad size distribution, combined with a powder blend with overall good flowability, could also lead to content uniformity issues. For this reason, it is desirable that proper measures are put in place to mitigate blend heterogeneity and demixing potential at the early stage of the formulation development, such that the quality of content uniformity is designed into the formulation prior to at-scale continuous process.

In this paper, a near-infrared spectroscopy (NIRS)–based ranking approach was developed for DC formulations with respect to inherent blending variability after apparent blend homogeneity is reached. A series of three DC powder blends, using fumaric acid as the surrogate API, was designed with different extents of inherent blend heterogeneity and demixing potentials. Blending variabilities were studied using NIRS in two-folded efforts. First, time-based triggering was used to collect spectral data continuously in order to explore the blend inherent variability with respect to rotation angles within each revolution and between revolutions. Second, multiple experimental conditions under the angle-based triggering, including blending speeds and co-adds were studied. Varied experimental conditions were intended to mimic different estimated sample sizes by NIRS, which was further used to link with off-line HPLC analysis on the powder blends with corresponding sample sizes. Despite many assumptions taken to estimate sample size (such as depth of penetration and dynamic blending density), it is important to use the estimated sample size as the linchpin between HPLC and NIR to illustrate how early-phased NIR data can be used to characterize inherent blend variability to support formulation and process development. Even though absolute values on blend variability could be different between NIR and HPLC across blends, the relative trend across different blends given a specific sample size between NIR and HPLC is the key here to rank different formulations regarding their inherent blend variability. Such a characterization approach is expected to support batch process development in small batch scale (50–100 g) at early CMC in decision-making about excipient selection. This is based upon an improved understanding of the relationship between blending variability and blending process parameters. Moreover, such an approach is also expected to play an important role in enabling transferring the process understanding of blending variabilities to both batch and continuous manufacturing processes at commercial scale.

Materials and Methods

Materials

The following pharmaceutical excipients were employed in this study: fumaric acid (Merck KGaA), microcrystalline cellulose (Avicel PH 102 and 200, FMC Biopolymer), dicalcium phosphate dihydrate (Di-TAB, Innophos), dicalcium phosphate anhydrous (Di-Cafos A60, Budenheim), croscarmellose sodium (Disolcel GF, Mingtai Chemical), and magnesium stearate (Ligamed MF-2-V, Peter Greven). All materials are of USP grade. Powders were screened through a US standard mesh #18 (1 mm) prior to use.

Spectral Collection on Three Blends with Different Intrinsic Blending Variabilities

Three direct compression blends, using fumaric acid as the surrogate API, were produced for the blending experiments in order to characterize blending variability (i.e., blend uniformity and demixing potential) at steady state via NIRS. All three powder blends possessed 15% (wt%) of the fumaric acid (a.k.a. API), two types of diluents (microcrystalline cellulose and dicalcium phosphate), a disintegrant (croscarmellose sodium), and a lubricant (magnesium stearate). The extent of blend heterogeneity was controlled by adjusting the grade and composition of the constituents, such that the level of blend heterogeneity, or lack thereof, may arise from excipients’ deviation from fumaric acid with respect to the particle size distribution and bulk density, as well as the flowability of the powder blends. The formulation compositions of the three powder blends, as well as the particle size distribution (d10, d50, d90) of constituents, are given in Table I.

Table I Formulation Compositions of the Three Fumaric Acid Powder Blends, and the Particle Size Distribution of Individual Constituents
Full size table

A Servolift bin-blender (12 L, Servolift GmBH, Offenburg, Germany) was selected for the blending process. A SentroPAT BU II analyzer (Sentronic GmBH, Dresden, Germany) was mounted on the lid of bin-blender. The lid was fabricated to contain a sapphire window insert (4″ radius) (Fig. 1), such that NIR spectra were collected through the sapphire window. Blending speeds of 5 and 25 rpm were investigated. An integration time around 150 ms was determined before daily operation of the analyzer. A spectra collection frequency of 1 Hz was used for all the spectral collection regardless of the use of co-adds. Both time- and angle-based triggering modes were adopted for the measurements. The time-based triggering was intended to investigate the blending variability with respect to the rotation angle of the blender. No co-add (averaging of collected spectra) was used for the time-based trigger under both 5 and 25 rpm. The rotation speed of 25 rpm was found to generate only one proper spectrum per revolution under the time-based data collection mode and 1-Hz-spectral-collection frequency. Here a proper spectrum meant that a spectrum was collected when the powder blend was in direct contact with the spectrometer through the sapphire lens. Thus, the data collected under 25 rpm was not used for blending variability characterization with respect to rotation angle. By contrast, the rotation speed of 5 rpm and the 1-Hz-spectral-collection frequency were found to generate roughly five proper spectra within each revolution. Such repeated measurements within each revolution enabled our investigation into blending variability characterization at different rotation angles.

Fig. 1
figure 1

The bin-blender/SentroPAT BU set-up (left) and the close-view of the fabricated lid of the bin, which contains a sapphire lens in the center for NIRS data collection (right)

Full size image

For the angle-based triggering, the following four conditions were tested: 5 rpm/1 co-add, 5 rpm/4 co-adds, 25 rpm/1 co-add, and 25 rpm/4 co-adds. The triggering angle for all measurements was set at 30° rising edge, which was the first 30° after the spectrometer passed the bottom of the rotation. Under time-based triggering, it was discovered that there was only one proper spectrum collected within a single revolution under 25 rpm. Such phenomena were not found to impact the spectra quality collected under the angle-based triggering and the condition of 25 rpm/4 co-adds, since the necessary co-add taking place within the spectrometer smoothed out any potential impacts of noisy spectra.

Spectral Analysis for Time-based Triggering

An iterative principal component analysis (PCA) was first conducted on individual blends to reject outlier spectra and ensure the remaining proper spectra are representative of the powder blends. For each experiment conducted under the condition 5 rpm/1 co-add, each blend was mixed for 10 min, which resulted in ~ 600 spectra. After outlier rejection, about 45% of total spectra collected were found to be proper spectra across all three blends. Second, a formulation-dependent classical least squares (CLS) model was built to translate spectra collected at different rotation angles to concentrations for comparison across the three blends. The training set for the CLS model included the last 20 proper spectra collected under the condition of 5 rpm/1 co-add (when the steady state blending was reached), and the 20 spectra of pure microcrystalline cellulose (Avicel PH 102) collected under the same experimental conditions. The nominal blend compositions were used as the reference values. Such a formulation-dependent training set (11) allowed estimating pure spectra for all the formulation components, which further enabled concentration tracking for all the components during the blending process and thereby the characterization of the inherent blend variability. Additionally, proper spectra collected within a single revolution were considered as a sample group to characterize the blending variability with respect to rotation angles. The relative standard deviation (RSD) of predicted fumaric acid concentration calculated within a single revolution, as well as the distribution of the fumaric acid concentration across multiple revolutions, were used to compare the inherent blend heterogeneity of the three powder blends.

Spectral Analysis for Angle-based Triggering

Four aforementioned experimental conditions (including rpm and co-adds) were tested. Each blend was continuously mixed for 5 min under each experimental condition. Three different algorithms were used to characterize blending variability under the angle-based triggering, including moving window block standard deviation (MBSD) (12), principal component analysis (PCA) (1) and classical least squares (CLS) (13, 14). The first two approaches were qualitative in nature, while CLS was quantitative. The spectral signals across the entire wavelength range (i.e., 1350–1800 nm), followed by standard normal variate (SNV) preprocessing was used by the three algorithms. Details of the data analysis are provided as follows:

  1. 1.

    A window of three spectra was used for MBSD calculation.

  2. 2.

    For the training of PCA, the model was trained on the data collected under 5 rpm/1 co-add. Here, formulation dependent PCA models were built, followed by applying the model(s) to the data collected on the same blend, but under the other three experimental conditions.

  3. 3.

    For the training of CLS, the 20 spectra of each blend were combined with 20 spectra of neat microcrystalline cellulose (Avicel PH 102), both collected under 5 rpm/1 co-add, to form a lean training set. Similar to the PCA approach, formulation-dependent CLS models were built. The model(s) were then applied to the data collected on the same blend under the other three experimental conditions. Prediction-augmented classical least squares (PACLS) (13, 14) were performed to address the impact of rotation speed, when the formulation-dependent CLS model was applied to data collected at 25 rpm.

Estimated Sample Size Calculation

Given what we learned in Section “Blending Variability with Respect to Rotation Angles” regarding the static powder at the bottom layer of powder blend repeatedly scanned by NIRS, an estimated sample size for static powder was chosen here. The early reported approach in a feed frame application (15) was adjusted to remove the dependence of estimated sample size on blending speed and the number of co-adds. The estimated sample size calculation was conducted using the formula as below.

$$estimated\; sample\; size=A\times A\times d\times 0.785\times \rho$$
(1)

where A stands for the spot size; d stands for the depth of penetration; and \(\rho\) stands for the moving powder density. The spot size of SentroPAT BU II was 16 mm. The depth of penetration was assumed to be 1 mm (16,17,18) per literature survey. The moving powder density was assumed to be 0.48 g/mL (15) for all three powder blends. The factor 0.785 was reported to be the conversion factor of the interrogated surface area between a square and a circle (15).

HPLC Reference Measurements

After the spectral data collection via angle-based triggering (i.e., 30-degree rising edge), powder blends A and C (representative of the lowest and highest blend heterogeneity) were subjected to blend uniformity characterization by HPLC. The powder blends were transferred to a V-shaped blender and samples from five spatial locations within the V-shaped blender (i.e., top left, top right, middle left, middle right, and bottom center) were collected using a sample thief. At each spatial location, samples with three different target weights (100, 500, and 2000 mg) were withdrawn. For each sample size, three replicate units were withdrawn and two of them were tested for the fumaric acid content by HPLC analysis.

An Agilent 1290 Infinity HPLC equipped with a Kinetex XB C18 column (2.1 mm by 150 mm, 1.7 µm) was used for the analysis. The mobile phase was 0.05% trifluoroacetic acid in water:acetonitrile (99:1). The HPLC was operated at a flow rate of 0.3 mL/min and a column temperature of 30°C. HPLC samples were prepared by suspending powder blends in water, followed by sonication for 30 min. The resulting suspensions were filtered through 0.45-μm-syringe filter for HPLC injection. The UV detection wavelength was set at 210 nm.

Particle Size Distribution Measurement

The particle size distribution of all formulation constituents were measured using a laser diffraction-based particle size analyzer (Mastersizer 3000, Malvern Panalytical, Malvern, UK), equipped with a wet dispersion unit. Samples were suspended in heptane with 0.1% Span 85 and subjected to stirring at 2000 rpm prior to measurement. The particle size distribution was obtained using the Mie scattering model. Measurements were performed in triplicate for each material.

Software

All data analysis calculations were executed using MATLAB 2020b (The Mathworks, Natick, Massachusetts) with the PLSToolbox 8.9.1 (Eigenvector Research, Inc., Manson, Washington), as well as multiple MATLAB routines written in-house to support this work, such as MBSD and CLS.

Results and Discussion

Spectral Outlier Rejection in Time-based Triggering

Under the time-based triggering mode, both unrepresentative spectra (i.e., spectra collected when the powder blend was not in full contact with the spectrometer through the sapphire window) and proper spectra were collected. Among those unrepresentative spectra, some displayed a discontinuity around 1550 nm, which was speculated to be caused by the two optic modules in the spectrometer not being able to stitch the spectra properly. As shown in Fig. 2A–C, spectral outlier rejection was necessary, so that meaningful blend heterogeneity characterization can be carried out on the proper spectra. For this reason, an experiment-specific principal component analysis (PCA) was conducted iteratively on the spectral data collected under 5 rpm/1 co-add for individual blends. The iterative PCA successfully removed these unrepresentative spectra while keeping proper spectra intact for blend heterogeneity characterization. As exhibited in Fig. 2D –I, the proper spectra were found to possess score values on the 1st PC close to zero when an initial PCA was conducted, while the spectral data showing extreme score values were projected and later confirmed to be outlier spectra. The outlier rejection led to “cleaned-up” score plots (Fig. 2J–L), which enabled robust characterization of the blending process and the blend heterogeneity. The blend “endpoints” illustrated by the vertical lines in Fig. 2J–L were obtained when the score values started to hover around zero. This only represented one of many endpoint criteria that could be applied onto the PCA score data. Since the endpoint determination was not the primary focus here, the vertical lines were drawn to call for attention to the significantly larger inherent blend variability of powder blend C compared to that of powder blends A and B, despite its surprisingly early “endpoint” being reached.

Fig. 2
figure 2

Spectral rejection for time-based triggering measurements. Each column represents the data flow of spectra collected under 5 rpm/1 co-add for a blend, where rows 1–4 represent the raw spectra (plots A–C), initial PCA score plot (plots D–F), proper spectra after spectral rejection (plots G–I) and final PCA score plot after spectra rejection (plots J–L), respectively. The red circles inside panels D–F represent where proper spectra, obtained after the outlier rejection, were projected in the initial PCA. The y-axis ranges of all plots in the same row were kept the same for meaningful visual comparison. The vertical lines in plots J–L indicate the blend “endpoint” (i.e., a steady state was reached when score values started to hover around zero)

Full size image

Blending Variability with Respect to Rotation Angles

Only the proper spectra representative of steady state were used for the characterization of blending variability with respect to rotation angles. The steady state here refers to the status in which fluctuation of scores remained stable over time (Fig. 2J–L). As shown in the last 20 good quality spectra collected for each blend (shown in Fig. 3A–C), powder blend C possessed the largest variability at the steady state (especially around 1500 nm), followed by powder blends B and A. This difference was further corroborated by the predicted fumaric acid concentration variability at the steady state (Fig. 3D–F). The difference between powder blends A and B was small, while the difference from powder blend C was substantial. This observation was generally in agreement with the original formulation design (Table I). In fact, powder blend C almost never reached an acceptable level of blend homogeneity, despite the prolonged blending. Additionally, the range of predicted fumaric acid concentration across revolutions in Fig. 3G–I serves as another good indication of the ranking of blend heterogeneity. This finding was further corroborated by the 50 and 95 percentile values of the calculated RSD based on the NIR data within each revolution between blend C and the other two blends, as shown in Table II.

Fig. 3
figure 3

CLS characterization on blending variability with respect to rotation angle. Each column represents results from individual fumaric acid blends. Rows 1–3 represent raw spectra used for CLS calibration (plots A–C), predicted fumaric acid concentration with respect to blending time (plots D–F), and predicted fumaric acid concentration with respect to rotation angles (plots G–I), respectively. The consecutive spectra # in plots G–I reflects different rotation angles within the same revolution. Each circled line represents predicted fumaric acid concentration within the same revolution, but at different rotation angles. The y-axis ranges of all plots in the same row were kept the same for meaningful visual comparison

Full size image
Table II Relative Standard Deviation (RSD) Comparison Across Three Fumaric Acid Blends per Revolution Under the Experimental Condition of 5 rpm/1 co-add
Full size table

More information can be derived from Fig. 3G–I. These plots clearly show that the greatest variability in predicted fumaric acid concentration occurred across revolutions (different line segments), but not within the same revolution (the same line segment). This agreed with earlier literature finding that in tumble blending, the bottom layer of the powder mixture tends to move with the blender within each revolution due to powder-wall friction (19). For the fumaric acid powder blends the most effective blending typically takes place at the completion of a revolution, at which time the powder residing at the bottom layer is carried to the top and begins to fall off due to gravity. Because the NIR lens is in contact with the bottom layer of the powder blend, the powder specimen exposed to the NIR lens likely remains largely static within one revolution. A new powder specimen is acquired by the NIR lens only at the inception of a new revolution. This also indicates that the co-add measurements commonly used during batch blending process under angle-based triggering essentially scan the same sample repeatedly within a revolution. Moreover, the common notion of operating powder mixture under different blending speed in a feed frame application to improve sampling volume no longer applies here given the static powder scanned by NIRS within each revolution. In other words, the use of co-adds and blending speed to adjust the estimated sample size inside a tumble bin blender is no longer representative of the underlying blend heterogeneity.

Moreover, the aforementioned results in Fig. 3G–I are also found to agree with the visual observation of the powder movement inside the bin, obtained from a GoPro camera mounted on the sapphire window. Since the GoPro camera was mounted on the sapphire window, what was observed by the camera was at the opposite end of what SentroPAT BU measured. That being said, the blending patterns from revolution to revolution stayed the same. The camera footage revealed that at 5 rpm, powder blends did not remain entirely static inside the bin. However, there was a velocity gradient within the powder bed. The top surface of the powder bed was almost always in motion, whereas the powder at the bottom layer did not move significantly until it was carried to a position with a very steep repose angle.

Additionally, we found that the powder blend did not follow a cascading motion as literature suggested (20). Rather, it underwent full rotation along with the bin (aka. similar to cataracting motion), with the most intensive mixing taking place at the conclusion of the revolution, at which time the entire powder blend, including the bottom layer, cascaded down from the top position (Fig. 4). This movement pattern is in agreement with Fig. 3G–I, which exhibited greater inter-revolution variability than intra-revolution variability. It is worth noting that the camera cannot make observations when the sapphire window was fully covered, so that the exact velocity, or the lack thereof, of the powder movement against the window, at which NIR data was acquired, cannot be determined. Nevertheless, the visual observation made by the camera clearly indicated that the moving pattern of the powder blends, as probed by the NIR lens, is not uniform, and that such information may be required when the NIR spectroscopic data is interpreted.

Fig. 4
figure 4

Snapshots from a video file, generated by a camera mounted on the sapphire window during the blending operation of powder blend C. The video showed that at 270° (top panel), the sapphire window was fully covered by the powder; toward the conclusion of one revolution (bottom panel), powder were cascading down from above the scope of the camera

Full size image

Blend Heterogeneity at Steady State Using Angle -based Triggering Measurement

The three aforementioned algorithms — moving block standard deviation (MBSD), principal component analysis (PCA), and classical least squares (CLS) — were used to characterize blend heterogeneity at steady state using the angle-based triggering measurement. The angle-based method is valid because only one acquisition was made per revolution. For every acquisition, the NIR probe captured a new sample specimen arising from adequate powder mixing. Thus, the inherent blend heterogeneity can be faithfully reflected accordingly.

Being qualitative, MBSD does not require any training data, making it well-suited for early-phase R&D process characterization. The disadvantage is that MBSD values may not directly correlate to concentration of API, especially when API does not possess dominantly unique absorption bands compared to major diluents. Under certain circumstances, the variabilities on MBSD may be impacted more by physical properties than chemical properties of powder blends. As a comparison, PCA and CLS do outperform MBSD in this regard. Both PCA and CLS require a training dataset. For a batch blending process, the training dataset can be conveniently acquired after blending homogeneity is achieved by attaching an NIR spectrometer for the blending operation. A two-leveled concentration training set has been proven useful in this case (11, 21). This enables the translation of spectral data to concentration information, essentially starting from the very first blending experiment at small scale. For PCA, the first PC is expected to describe the concentration evolution of API within a batch blending process. While for CLS, given the estimated pure spectra for each formulation components as the outcome of the CLS training, the predicted concentration of individual components in the blend can be directly used to inform the blend heterogeneity. Moreover, the lean requirement of the necessary training data makes CLS a very attractive method for early phase R&D, especially considering the limited API supply and the fact that a fully quantitative method calibration via partial least squares (PLS) is often not feasible (11, 21).

As expected, regardless of the algorithm employed, the outcome of the blend heterogeneity is the same (i.e., powder blends C > B > A, as shown in Fig. 5). Similar blending profiles between PCA and CLS also confirmed that the 1st PC was capturing the concentration change of fumaric acid in the blending process. Moreover, we believe that CLS is uniquely advantageous over other two methods. CLS provided predicted constituent concentrations, which were otherwise not attainable by MBSD or PCA methods. The predicted concentrations of constituents offer additional insights into the blend heterogeneity. For example, although high variability in predicted fumaric acid concentration was observed in powder blend C (Fig. 4I), the average concentration over all data points remains close to the nominal value (15%). This observation suggests that despite the high inherent heterogeneity of the blend, there is no pronounced demixing potential. Namely, fumaric acid particles do not appear to preferentially reside in preferred locations within the bin, which would lead to strong deviations from the nominal concentration when observed at certain rotation angles. A further advantage of CLS is that the constituent concentrations from CLS can directly inform the content uniformity of the downstream processes, so that in-process decisions can be readily made based upon predetermined control limit.

Fig. 5
figure 5

Blend heterogeneity characterization under 5 rpm/1 co-add by three different algorithms. Each column represents the results for individual blends. Rows 1–3 represent the results from different algorithms, including MBSD (plots A–C), PCA (plots D–F), and CLS (plots G–I), respectively. The y-axis ranges of all plots in the same row were kept the same for meaningful visual comparison

Full size image

A detailed comparison on CLS characterization on individual blends using four different experimental conditions is shown in Fig. 6 and Table III. The orthogonal results by HPLC are also shown accordingly. Here, we focused on powder blends A and C, which represent the highest and lowest blend heterogeneities. Overall, no noticeable variability reduction was observed when the co-adds were increased from 1 to 4. This corroborated the early findings that the bottom layer of the powder mixture moved with the blender (and hence the NIR lens), which resulted in similar data variability observed within each revolution regardless the choice of co-adds, due to the fact that different replicate scans essentially measured the same sample again and again under the angle-based trigger (Fig. 3G–I). In other words, the estimated sample size is no longer a function of co-adds and blending speed in the tumble-bin blending operation. It is worthy to point out that both parameters were in fact routinely employed to adjust sampling volume and data variability when considering the deployment of PAT tools for feed frame applications (15). Our study indicates that probe-based sample sizing is largely operation dependent, and one needs to take into account the powder movement pattern, especially for the portion presented to the lens, while altering sample volume via co-add and blending speed given the acquisition time.

Fig. 6
figure 6

PACLS characterization of blending variability at steady state for powder blends A and C under different experimental conditions

Full size image
Table III Estimated Blending Variability (i.e., RSD, Relative Standard Deviation) and Sample Size Measured by NIRS and HPLC Under Different Experimental Conditions
Full size table

Meantime, it is also worthy to point out that a slight reduction of the RSD was observed when the blending speed was increased from 5 to 25 rpm (Fig. 6). This can be attributed to the two following observations. First, video images collected by the GoPro camera did indicate more intense overall powder movement when the blender was operated under 25 rpm vs. 5 rpm. There is a possibility that the powder immediately adjacent to the sapphire window is more mobile when operated at 25 rpm, due to the turbulent flow pattern. Second, given a single scan time of 150 ms, a 22.5-degree of circular arc was traveled under 25 rpm compared to a 5-degree of circular arc at 5 rpm. Considering the angle-triggered measurement was conducted at 30-degree of rising edge, the powder blend is much closer to the equatorial line under 25 rpm at the conclusion of the data acquisition, compared to that under 5 rpm. This may indicate that powder could start to move due to gravity towards the end of the scan collected under 25 rpm.

As shown in Fig. 6 and Table III, the HPLC data from the blend uniformity samples agreed with NIRS regarding the overall blend variability between blends A and C, though the absolute RSD values did not always match. The agreement suggested the possibility of leveraging NIR spectra collected within a bin-blender to forecast content uniformity on tablets, given common oral solid dosage forms varies 50–500 mg. Moreover, if NIRS data can be collected at small scale (i.e., 50–100 g) across multiple formulations during early-phase formulation development, one may be able to assess the risk of formulation blend heterogeneity by leveraging the RSD data calculated by NIRS. Furthermore, when projects advance to later development stages, continuous NIRS data collection on either a larger scale batch blender or a continuous blender is also expected to inform and forecast downstream uniformity across scales and different modes of manufacturing.

Despite the agreement between NIR and HPLC on the RSD trend between two blends, discrepancies were observed regarding the absolute RSD values between NIRS and HPLC and absolute RSD differences between two blends. Greater discrepancies were observed under the smallest sample size of 100 mg when comparing NIR against HPLC. These discrepancies may stem from the following considerations. First, based on Eq. (1), the sample size interrogated by NIR was 96 mg (assuming 1-mm depth of penetration and 0.48 g/mL as the powder density), and the corresponding RSD ranged between 1.46 and 3.77%. In contrast, the HPLC-derived RSD at 100-mg-sample size was significantly higher (approximately 7%, as shown in Table III). Noting that the NIR-based RSD was on par with those by HPLC in the order of sample size of 500 mg, it was thus postulated that the assumed depth of penetration of 1 mm was under-estimated. For comparison purposes, the estimated sample sizes impacted by different values of assumed depth of penetration are shown in Table IV. As can be seen, in order for the RSD values estimated by NIR to be similar to that by HPLC, the depth of penetration would need to be increased to around 5 mm. Given most of the reported depths of penetration were measured on static powder or tablets, the powder bed within a blender could thus be speculated to possess a larger depth of penetration. Second, a common formula (Eq. (1)) was used to calculate the estimated sample sizes by NIR, of which the same assumptions on dynamic powder density were applied across all three blends. It is known that the powder blends possess different bulk density values (blends A, B, and C exhibit bulk density of 0.43, 0.56, and 0.64 g/mL, respectively). Although the dynamic powder density is not expected to be equivalent to the static bulk density, applying the same dynamic powder density to all three powder blends inevitably led to errors in sample size estimation. Thus, it was speculated that a considerable amount of uncertainty existed on the assumed dynamic powder density when estimating the sample size for different blends. Taken together, uncertainties are expected in sample size calculation for powders (especially at smaller estimated sample size) when they are subjected to bin blending operation.

Table IV Estimated Sample Size (mg) of the Fumaric Acid Powder Blend via Different Values of Assumed Depth of Penetration by Eq. (1)
Full size table

It is worthy to point out the different uses of CLS (Figs. 3 and 5) and PACLS (Fig. 6) were demonstrated here to translate dynamic spectra collected in batch blending process to API concentration predictions and subsequently blending variability characterization at the steady state. Since PACLS was used here to address the physical interferences caused by different blending speed, the predicted concentrations by PACLS were observed to possess less variability compared to those by CLS on the same blend under the same blending conditions, such as the results under 5 rpm and 1 co-add in Fig. 6 vs. the results in Fig. 5G and I. However, the similar variability ranking across three blends was observed regardless of the choice between CLS and PACLS. Thus, as far as a common analysis routine is used to characterize blending variability across different formulas containing the same API or across different scales of the same formula, the characterization of blending variability is expected to be valid for the intended purpose of improving process understanding and forecasting content uniformity at the tablet level.

Conclusion and Prospective

A NIRS-driven methodology for characterizing blend heterogeneity of direct-compression powder blends was developed. The method was tested on three powder blends, designed to harbor different levels of blend heterogeneity. By using a SentroPAT BU II NIRS unit, the inherent powder heterogeneity among blends was well characterized via the NIRS data and confirmed by the HPLC data, collected after the blending reached a steady state. Importantly, we found out that the adjustment of the number of co-add and the blending speed did not lead to changes in the estimated blending variability, in contrast to the regular practice undertaken in a typical feed frame operation. The insensitivity of co-add and blending speed to the estimated blending variability arises from the fact that the powder at the bottom layer essentially moved with the bin during tumble blending, such that the powder specimen captured by the NIR lens remained largely static until the completion of one revolution. Thus, it is important to consider the powder movement behavior in an operation, when the PAT interface is leveraged to interrogate the blend heterogeneity.

Additionally, we also showed that the CLS (and PACLS) had a unique advantage over other processing algorithms (MBSD and PCA). CLS gave rise to predicted constituent concentrations, thereby allowing for a more in-depth assessment of blend heterogeneity and segregation patterns and rendering the blending step a potential process control point in manufacturing. Overall, a proper set up of the experimental conditions for NIRS data collection, combined with the use of CLS in data processing, can provide a wealth of information on inherent heterogeneity of powder blends, which would be otherwise not attainable with implementation of off-line, LC-based methods.

Our future work entails a continued endeavor to demonstrate the value of such methodology from the lab scale (~ 50 g) to pilot scale (~ 1 kg), and further to the scale commensurate with continuous manufacturing. Given the utmost importance of blend homogeneity to the success of continuous direct compression processes, the methodology presented in this study holds a promise to enable robust, early-stage formulation development toward the continuous direction compression process trains.