In Situ Monitoring of Polymorphic Transformations Using a Composite

Article pubs.acs.org/OPRD

In Situ Monitoring of Polymorphic Transformations Using a Composite Sensor Array of Raman, NIR, and ATR-UV/vis Spectroscopy, FBRM, and PVM for an Intelligent Decision Support System E. Simone,† A. N. Saleemi,‡ and Z. K. Nagy*,‡,§ †

Department of Chemical Engineering, Loughborough University, Loughborough LE11 3TU, U.K. EPSRC Centre for Innovative Manufacturing in Continuous Manufacturing and Crystallization, Department of Chemical Engineering, Loughborough University, Loughborough LE11 3TU, U.K. § School of Chemical Engineering, Purdue University, West Lafayette, Indiana 47907-2100, United States ‡

ABSTRACT: Raman spectroscopy, particle vision and measurement (PVM), and infrared spectroscopy have already been used singularly to detect nucleation and polymorphic transformations during crystallization processes as well as to analyze powder mixtures of multiple polymorphic forms. However, a comprehensive study and comparison of these techniques used simultaneously with integration of the signals during polymorphic transformation has not been performed to date. The aim of the present work is to compare the effectiveness of Raman spectroscopy, near-infrared spectroscopy, and PVM to detect nucleation and polymorphic transformations during crystallization processes. The information from focused beam reflectance measurement (FBRM) and attenuated total reflection ultraviolet/visible (ATR-UV/vis) probes were also used simultaneously in the same system to validate and integrate the results obtained by the other probes. This paper illustrates for the first time of the use of a composite sensor array (CSA) that truly integrates the signals from a variety of process analytical tools for better information extraction and more robust detection of various crystallization mechanisms. The proposed principal component analysis-based signal integration and the CSA can be used as the bases for an intelligent decision support system for crystallization processes, which can automatically detect various mechanisms occurring during the process, as well as a framework for the automated selection of the optimal minimal sensor configurations for the robust detection of these events.

1. INTRODUCTION The term “process analytical technology” (PAT) became widely popular in 2004 when the Food and Drug Administration published the “Guidance for Industry”,1 which describes PAT as a tool that enables process understanding for scientific, riskmanaged pharmaceutical development, manufacture, and quality assurance. These tools can provide information to facilitate process understanding, continuous improvement, and development of risk-mitigation strategies. Process analyzers and instrumentation include on- and in-line equipment, which allow instant measurements to be obtained, avoiding the time delay typical of off-line analysis. The measured properties can be univariate (scalar) or multivariate (vector or matrix) quantities. In the last decades, advanced spectroscopic instrumentation has progressed tremendously: ultraviolet (UV), visible, nearinfrared (NIR), mid-infrared (mid-IR), and Raman spectroscopy have been studied and implemented in manufacturing processes.2 For the crystallization process, PATs have been widely used recently, in particular focused beam reflectance measurement (FBRM),3 particle vision and measurement (PVM), and nuclear magnetic resonance (NMR), attenuated total reflectance (ATR)-UV/vis, ATR-FTIR, and Raman spectroscopy. NIR and mid-IR spectroscopy have been used both with ATR probes for solute concentration measurements4,5and without ATR probes for analysis of solid powders and tablets.6 © XXXX American Chemical Society

For polymorphic transformation detection, NIR and Raman spectroscopy are the preferred tools. A brief summary of the most common PAT tools used in polymorph screening and transformation monitoring in recent years is shown in Table 1. The aim of the present work is to evaluate the capability of different PAT tools to monitor polymorphic transformations during crystallization, to provide a systematic analysis of the signals, and to investigate which combinations of tools are the best for the detection of various mechanisms in a particular system. Additionally, the framework of a composite sensor array (CSA) is introduced, which is based on the integration of signals from multiple sensors using principal component analysis (PCA). The idea of combining multiple sensors in one measurement hardware cluster has been developed by Zeton B.V. (Enschede, The Netherlands). The company manufactured a sensor skid containing a microcamera, an ATR-FTIR spectrophotometer, an ultrasound device, and a refractometer.37 A device composed of multiple probes was used by Helmdach et al.38 Nevertheless, while these systems physically combined multiple pieces of hardware, they represent decentralized PAT arrays since none of the proposed systems provided any mechanism for the integration of signals Special Issue: Process Analytical Technologies (PAT) 14 Received: January 13, 2014

A

dx.doi.org/10.1021/op5000122 | Org. Process Res. Dev. XXXX, XXX, XXX−XXX

Organic Process Research & Development

Article

data fusion, and introduces a methodology for the determination of the optimal sensor configuration for the automated detection of various crystallization events.

Table 1. Summary of recent studies conducted using PAT tools for polymorph screening and transformation monitoring technique

process studied

references

mid-IR ATR-UV/ vis

discrimination between polymorphic forms of solids solvent-mediated polymorphic transformation monitoring in batch crystallization (changes in solute concentration) determination of the polymorphic ratio of solid mixtures solvent-mediated polymorphic transformation monitoring determination of the polymorphic ratio of solid mixtures comparison of the two techniques in determining the polymorphic ratio of solid mixtures

7, 8 9−12

monitoring of carbamazepine and nicotinamide cocrystallization solvent-mediated polymorphic transformation monitoring study of the impact of operating parameters on polymorphic transformation of D-mannitol

24

NIR NIR Raman FTRaman, FTIR Raman Raman Raman, FBRM, PVM Raman Raman FBRM Raman, NIR, UV−vis

measurement of the total solid concentration in a polymorphic system measurement of solute concentration in a polymorphic system solvent-mediated polymorphic transformation monitoring solvent-mediated polymorphic transformation monitoring

2. MATERIALS AND METHODS 2.1. Materials and Experimental Setup. The model compound used for the experiments is o-aminobenzoic acid (anthranilic acid, OABA), which has three known polymorphic forms. The transformation studied is the one from metastable form II (Figure 1b) to stable form I (Figure 1a) in a solution of 90% water and 10% isopropyl alcohol (IPA). The system is enantiotropic.

13−15 14, 16 17−21 22, 23

25, 26 27 28, 29 30, 31 32−36 37, 38

and data fusion. An example of data processing with the commercial softwares PEAXACT (S-PACT GmbH) and OPUS was performed on mid-IR data by Helmdach et al.,39 but the concept of a centralized CSA was first proposed by Nagy and Braatz40 and is based on the integration of information from multiple PAT tools for automated detection of various events that may occur during crystallization. Often several PAT tools are used for development and scale-up of crystallization processes, and while crystallization experts typically analyze and interpret the complementary information to understand the process, no framework is available for automated signal integration and interpretation using modelbased approaches or chemometrics techniques (such as PCA). According to the concept of CSA, combining the signals from all of the techniques is the same as having a single global sensor that can monitor all of the different phenomena detectable by different individual tools. The signals can be analyzed simultaneously using chemometrics techniques and combined in principal components that can be more easily interpreted and eventually used in control of the process.40,41 The Crystallization Process Informatics System (CryPRINS) is a software platform that consists of the communication interface required to integrate various PAT hardwares from the CSA, the automated data fusion mechanism (e.g., based on chemometrics or model-based approaches), and intelligent algorithms that can be used to coordinate the various components of the CSA and exploit all of the signals from all of the measurement devices for automated decision support and feedback control.40 This paper provides a systematic comparative investigation of the potential of different PAT tools in monitoring of polymorphic transformations, presents the first proof of concept of the CSA framework using chemometrics-based

Figure 1. (a) Prismatic form I and (b) needlelike form II of OABA.

OABA (>98% form I, Sigma-Aldrich), IPA (purity of 99.97%, Fisher Scientific), and ultrapure water obtained via a Millipore ultrapure water system were used. An RXN1 Raman analyser with an immersion probe and 785 nm laser (Kaiser with iC Raman 4.1 software) was used, together with an IN268 NIR reflection probe (Bruker with OPUS 7.0129), an IN271P NIR transflection probe (Bruker with OPUS 7.0129), a D600L Lasentec FBRM probe (Mettler Toledo with FBRM software version 6.7.0), an MCS 621 Carl Zeiss ATR-UV/vis spectrometer (in-house LabView software) with a Hellma ATR probe (type 661.822-UV) and a V819 PVM probe (Mettler Toledo with PVM online image acquisition software version 8.3) was employed. The pre- and postprocessing of the data was done with MATLAB R2012b. A 500 mL stirred, jacketed vessel fitted with a retreat curve impeller was used for the experiments. The temperature of the jacket was controlled by a PT-100 temperature probe connected to a Huber Ministat CC3 pump. The data from the FBRM, the ATR-UV/vis, and the Huber are transmitted in real time to the CryPRINS software. Raman and NIR data were collected with the software packages of the respective instruments. CryPRINS provides a communication interface for a variety of PAT tools via file B



Article

Figure 2. Schematic of the rig used during experiments with the composite PAT array.

The solubility S was then obtained using the following equation:

transfer, object linking and embedding for process control (OPC), RS232, or dynamic data exchange (DDE), etc., and allows real-time monitoring of these signals as well as setting a desired temperature profile and performing different types of feedback control. Experiments were performed using all of the probes at the same time in order to compare the information under identical experimental conditions and integrate the signals. The probes were put close to the stirrer and at the same height in order to have the best quality of measurement. A simple schematic representation of the experimental system with CryPRINS and the CSA is shown in Figure 2. This work is mainly focused on polymorphic transformations that happen in the solid state, and that is why most of the tools used are related to solid-state analysis (NIR and Raman spectroscopy). ATR-UV/vis spectroscopy was also used because of its capacity to detect the different solubilities of the two polymorphs. 2.2. Experimental Methodology. To determine the solubility of each polymorphic form in the solution, slurries of OABA with excess solids were kept at constant temperatures in the range of 10−50 °C in increments of 5 °C. The slurries were kept at different temperatures long enough to reach equilibrium. The solvent used was 10% IPA and 90% water by weight. Form I solubility measurements lasted 3 h; form II measurements lasted 1 h to avoid transformation, as form II is the less stable form at low temperatures (samples were periodically taken and analyzed with differential scanning calorimetry to check the polymorphic form). As the temperature neared 50 °C, the equilibrium time for form I solubility measurements was reduced to 1 h, again in the interest of preventing a polymorphic change. For each experiment, three separate samples were taken through a filter to ensure that only the OABA dissolved in the solution was present, and the samples were left to dry at 40 °C for 24 h. At high temperatures, the pipet and containing dishes were heated to prevent undesirable rapid nucleation of the sample. The containing dishes were weighed before the sample was taken (the empty weight w0), before the sample was dried (the wet weight ww), and after the sample was dried (the dry weight wd).

S=

(wd − w0) (ww − wd)

(1)

The solubilities of forms I and II were found using thermogravimetric measurements, and the experimental data were interpolated using the second-degree polynomial curves shown in eqs 2 and 3: SII = 4.76 × 10−6T 2 + 9.06 × 10−5T + 2.23 × 10−3

(2)

SI = 5.20 × 10−6T 2 + 7.31 × 10−5T + 1.41 × 10−3

(3)

where SI is the solubility of form I and SII the solubility of form II in grams per gram of solvent and T is the temperature in degrees Celsius. A graph showing the experimental solubility points and the interpolation curves is reported in Figure 3. From the polynomial interpolation of the experimental data the transition temperature in solution seems to be around 68 °C, which is higher than the experimental value (around 55 °C). However, experimental data were not taken above 45 °C.

Figure 3. Solubilities of OABA forms I and II in 90:10 (w/w) water/ IPA solution. C



Article

Solutions of OABA pure form I containing 0.017 g/g of solvent (saturation temperature of about 40 °C) were used for the experiments. The solutions were heated to 50 °C, kept for 30 min in order to dissolve all of the solid, and then cooled at a rate of −1 °C/min. Data recording was started at 50 °C after the complete dissolution of the starting material. Because of the high concentration of water in the solution, which favors sticking of the crystals on the probes, the FBRM signal was disturbed (the total counts and chord length distribution present spikes and peaks). However, it was still possible to distinguish nucleation and polymorphic changes from the FBRM in both of the experiments. The NIR and Raman signals were apparently less influenced by sticking, and also, the PVM could still give very good and useful information. Four experiments were conducted with the arrangements of probes shown in Table 2.

Table 3. Summary of the preprocessing for each probe measurement before principal component analysis was performed; in the right column, the indexes of the total merged matrix of signals are associated with each probe

probes used

1 2 3 4

NIR reflectance (NIR ref.), Raman, FBRM NIR transflectance (NIR trans.), Raman, FBRM ATR-UV/vis, Raman, FBRM ATR-UV/vis, Raman, FBRM

merged matrix index

smoothing in the region from 5000 to 6500 cm−1 (higher smoothing for the reflectance probe) time averaging (10 s)

1−1286 1287−1526

1287−1525 (NIR refl.) 1287−1575 (NIR trans.) 1527 (ATR-UV/vis) 1526 (NIR refl.) 1576 (NIR trans.)

scores plotted versus time, the corresponding loadings plotted in scatter plots, as well as plotting the first and second principle components (PC1 versus PC2) and identifying clusters. Score plots are generally used to detect similarities, differences, or other interesting relationships among samples. Also, the loadings are often presented in two-dimensional scatter plots, but they can be plotted against the wavelength too. The loadings are used to interpret the relationship between the scores and the original variables and to select the optimal sensor configuration given the set of available PAT tools. In order to interpret the loadings in an easier way, varimax rotation was performed on the CSA loadings using the MATLAB function rotatefactors. This operation maximizes the sum of the variances of the square loadings; in this way, every component is associated with a smaller number of variables and the determination of the most influential probes in the set that compose the CSA is greatly simplified.42 For the matrix of loadings P = [pj,i] (with dimensions m × k), performing a varimax rotation means maximizing the varimax criterion among all of the orthogonal rotations of P. The varimax criterion is defined by the the following formula:

2.3. Methodology for Signal Integration Using Principal Component Analysis. PCA is a method of data analysis that can model all of the variations in the data set using orthogonal basis vectors (eigenvectors) that are called principal components (PCs). One of the main goals of PCA is to eliminate the PCs associated with noise.42 If A is the n × m matrix of input data (e.g., n rows of spectra recorded at m wavelengths), PCA can decompose it in the following linear system: A = CPT + ε

preprocessing second derivative smoothing in the region from 400 to 1700 cm−1 first derivative in the region from 220 to 725 nm

FBRM data

Table 2. Probes used during the performed experiments experiment

data Raman spectra ATRUV/vis spectra NIR spectra

(4)

where C is the n × k matrix of k PC scores and P is the m × k matrix of the eigenvectors of A. The eigenvectors are called “loadings” in PCA. The variable ε is a matrix containing the unexplained variance. A simple computational method to obtain simultaneously the loading, score, and percentage of variance associated with each PC is singular value decomposition (SVD). In this work, PCA was performed (using the svd function in MATLAB) on the CSA matrix formed by the signals taken from all the probes at the same time intervals. PCA can be effectively used to isolate the effects of different crystallization phenomena (such as nucleation, polymorphic transformation, agglomeration, etc.) and obtain useful real-time information about the process. This work represents the first step towards the realization of the CSA concept. PCA was performed on a matrix formed by multiple signals from different probes, and the results were analyzed in order to understand the potential of this approach to automatically detect polymorphic transformations and differentiate them from other potential mechanisms that may lead to changes in the signals. Raman signals, total counts, and NIR or ATR-UV/vis data were arranged in a single matrix. The data were preprocessed separately as shown in Table 3. Before merging the data in a unique matrix autoscaling was performed individually for every type of signal. In this way the same weight is given to the signals from each probe. The crystallization process can be analyzed and interesting information may be inferred from the

⎡ ⎢1 V (P) = ∑ ⎢ m i ⎢ ⎣

⎛ 1 ∑ pj ,i −⎜⎜ ⎝m j 4

⎞2 ⎤ ∑ pj ,i ⎟ ⎥⎥ ⎠ ⎥⎦ j 2⎟

(5)

3. RESULTS AND DISCUSSION 3.1. Qualitative Analysis of the Single Signals (Decentralised CSA). The Raman and NIR spectra of the solid forms of OABA are shown in Figure 4a,b, respectively. In the Raman spectra, form I can be unequivocally distinguished from form II by the peak at 950 cm−1, which is present only for form I. It was also observed that both the form I and II solids have two characteristic peaks in the regions 817−786 cm−1 (peak 1) and 786−755 cm−1 (peak 2). The more intense peak for form I is peak 1, while for form II peak 2 is more intense. Upon examination of mixtures with different ratios of forms I and II, it can be observed that when the fraction of form II increases, the intensity of peak 2 increases. At the same time, there is a decrease in the intensity of peak 1. Thus, it is possible to differentiate whether a solid sample is pure form I, pure form II, or a mixture.43−45 This characteristic trend is due to the fact that vibrations of different chemical bonds are affected by the crystalline arrangement. In the case of OABA, the crystal structure of form I allows easier vibrations in the 817−786 cm−1 region, while vibrations in the 786−755 cm−1 region are D



Article

Figure 4. (a) Raman spectra and (b) NIR reflection spectra of solid samples of OABA: pure form I (blue), pure form II (red), and a mixture of the two forms (75% form II) (black).

Raman, FBRM, and NIR data obtained during experiments 1 and 2 are shown in Figure 5a,b, respectively. Experiment 1 was conducted using an NIR reflectance probe. The absorbance signal from the reflectance probe in solution is noisy and shows high intensities because there are no solids that can reflect the radiation back to the detector. The NIR transflectance probe can also work with transmission, so the signal is specular: it is higher in slurries and lower in solution. The FBRM probe is the one with the largest measuring area and therefore the one most affected by fouling. The presence of particles on the probe made it difficult to interpret the results, since the repeated measurements of the sticking crystals generated meaningless peaks in the total counts and the chord length distribution (CLD). In experiment 1 (Figure 5a) in particular, the total counts trend has two peaks around 170 min; those peaks are just an artefact of the FBRM since none of the other probes detected any changes in the process. Nevertheless, the PCA applied on the CSA showed good results in detecting the real kinetic events in the process for all the performed experiments, demonstrating the robustness of this approach. Tables 4 and 5 show the detection times for nucleation and polymorphic transformation start and end for all of the probes used in the first two experiments, while Table 6 reports the characteristic times for experiment 3. To extract information from the PVM images, the start of the transformation was considered to occur when the first recorded image presented prismatic crystals (typical of form I), as shown in Figure 6. The end of the transformation was considered to occur when long needle crystals completely disappeared from the recorded images; this empirical way is not very accurate and justifies the different detection times for PVM compared with the other probes. The PVM, Raman, and NIR techniques are practically equally efficient in detecting nucleation and polymorphic transformation. The FBRM signal is less easy to interpret, but a significant variation in the total counts is present at the end of the polymorphic change. The increase in intensity in NIR transflectance before nucleation is related to the temperature change, but nucleation is still distinguishable since it generates a more sudden increase in the signal. The end of the polymorphic transformation was detected at the same time by the Raman, FBRM, and NIR probes in experiment 1. In experiment 2, the Raman probe recorded the end of polymorphic transformation a few minutes before the FBRM and NIR probes did. The beginning of the transformation was detected first by the Raman and PVM methods. The

not favored. The structure of form II instead favors vibrations at 786−755 cm−1 and partially blocks the ones in 817−786 cm−1. It can be noted that the two studied peaks both correspond to ring vibrations of benzene derivatives. The two mentioned Raman peaks (see Figure 4a) are still visible in solution, and they were tracked during the experiment together with an additional peak for form I. In particular, the exact peak positions chosen are 770 cm−1 for form II and 784 for form I. Second derivative and smoothing were used in order to partially eliminate the effect of fluorescence that was observed in the system even with clear solutions, and peak values were multiplied by −1 in order to have the same sign as the raw spectrum. In the case of the NIR spectrum in slurry, the absorption of water overlaps in the region that presents more differences between the two polymorphic forms in the solid state (6500−7500 cm−1). However, it is still possible to identify a peak for form I around 5950 cm−1 and also to identify the transformation due to a baseline effect in both reflectance and transflectance mode. In fact, the amount of diffuse radiation going to the detector is affected by the total amount of solids, their shape and size.46 Change in one or more of these parameters due to dissolution, nucleation or polymorphic transition, will generate a shift of the entire spectrum to lower or higher absorbance intensities. In particular, form I of OABA is characterized by larger crystals than form II, which results in a shift to higher absorbance values in the case of the reflectance probe and to lower values for the transflectance probe during the polymorphic transformation. For this reason, every significant peak in the NIR spectrum (apart from the extreme zone at 4000−4500 cm−1, where there is absorption by the optical fiber) could be followed. Differences in the spectra of the two polymorphic forms were visible even in slurries of 2 mg of solid/g of solvent (the peak at 5950 cm−1 was still present and was tracked during the experiments), but it was found that the baseline effect was stronger than the changes in the specific regions of the spectrum. The ATR-UV/vis probe can monitor the liquid phase and detect changes in solute concentration during the experiment. In experiment 3, a calibration function developed in the group47 was used to calculate the solute concentration from the UV data, as shown in Figure 5c. FBRM can give important information about the amount of solid in solution (counts or counts per second) and also about the crystal size distribution and, indirectly, about the shape of the crystals. E



Article

Figure 5. (a) Evolution of temperature, NIR reflectance signal, and Raman intensities for specific peaks of forms I and II in experiment 1. (b) Evolution of temperature, NIR transflectance signal, and Raman intensities for specific peaks of forms I and II in experiment 2. (c) Evolution of temperature, ATR-UV/vis signal, and Raman intensities for specific peaks of forms I and II in experiment 3.

Table 4. Detection times of nucleation and polymorphic transformation from the different PAT tools in experiment 1 technique NIR reflectance Raman FBRM PVM

nucleation (min)

polymorphic trans. start (min)

polymorphic trans. end (min)

26

200

281

26 26 27

120 268 105

281 283 261

Table 5. Detection times of nucleation and polymorphic transformation from the different PAT tools in experiment 2 technique NIR transflectance Raman FBRM PVM

transformation from form II to form I is known to be solventmediated,43−45and from the trend in the Raman signal it is

nucleation (min)



31

280

330

31 31 31

150 289 220

320 328 307

possible to infer that this transformation is also proceeds in two steps: a slower one and a faster one. The Raman signal is F



Article

the different shapes of the two forms (larger surface area and larger crystals of form I compared with form II). For FBRM, the decrease in the total counts is associated with the change in the shape of the crystals. A similar behavior in the total counts per second was observed for the same substance (but a different solvent) by Saleemi and Nagy,48 who also compared the trend in total counts during polymorphic transformation with the aspect ratio of the crystals. Although the conditions of the three experiments were the same, the third experiment presented a different transformation time. Many parameters that cannot easily be controlled influence the time of the transformation (mixing conditions, presence of dust or small amount of impurities), and this might explain the different behavior of the system in experiment 3. 3.2. Centralized Composite PAT Array with PCABased Information Fusion. The data from the CSA were combined in a single data matrix, and PCA was applied to it. The first four principal components of the global CSA matrix resulting from combining the information from the three PAT tools (Raman, FBRM, and NIR) were analyzed. The chosen components can detect nucleation and the polymorphic transformation, which are the phenomena of interest in this study, but the entire matrix contains additional information (e.g., agglomeration, secondary nucleation, and also fluorescence or simple noise). The scores and loadings were analyzed together with the amount of variance covered by the first four principal components, and the results are shown in Table 7. All of the principal components (among the first four) that can detect nucleation and/or the polymorphic transformation are also listed in the same table. It can be observed that nucleation was always detected by the first three principal components in every combination of PAT tools, while the polymorphic transformation (in particular the beginning) was observed in fewer components and in most cases was not distinguishable in the first principal component. In experiment 3 in particular, PC1 and PC2 start changing after nucleation and stabilize to a constant value after the completion of the transformation. This trend is similar to the UV signal during the experiment, demonstrating how different techniques can affect the PCA in different ways. In PCA the components are ordered according to the variance covered; because nucleation generates a more significant change in signal than the transformation, it appears in more components and in particular in the first one. For all of the experiments the fourth principal component shows only noise, while the other three can detect the phenomena of interest: nucleation of form II, the start of the polymorphic transformation, and the end of the transformation. The variances covered by the first three principal components (the ones that do not contain noise) change significantly from one experiment to the other: this is due to the fact that reflectance spectroscopy is considerably noisier than trans-

Table 6. Detection times of nucleation and polymorphic transformation from the different PAT tools in experiment 3 technique ATR-UV/ vis Raman FBRM PVM

nucleation (min)



30

72

190

30 30 30

70 70 65

190 192 190

Figure 6. PVM images from experiment 2 at (a) 220 min and (b) 280 min.

directly proportional to the ratio of the two polymorphic forms, and therefore, Raman spectroscopy is the only tool that can distinguish between the two steps of the solvent-mediated transformation. The NIR signal changed only at the beginning of the second step of the transformation, and this is related to the way that the two probes can detect the polymorphic transformation. Besides, the Raman probe can detect new peaks from the stable form just nucleated, but those crystals might be too small to be picked up and differentiated by the FBRM or NIR probes, which mainly look at the different shapes and sizes of the two polymorphic forms This signal from the reflectance probe increased at the end of the transformation, while that from the transflectance probe decreased. This behavior can be explained considering that the NIR signal is strongly influenced by crystal size and shape. When form I appears, the NIR intensity increases as a result of

Table 7. Results of the PCA performed on data from combinations of three probes and on the total data matrixes in experiments 1, 2, and 3: the second column reports the variance covered by the first three principal components, and the following columns report which phenomena are detected by each PC score (N = nucleation; FST = first step of the transformation; SST = second step of the transformation; ET = end of the transformation) experiment

variance (first three PCs)

PC1

PC2

PC3

PC4

1 (FBRM + Raman + NIR reflectance) 2 (FBRM + Raman + NIR transflectance) 3 (FBRM + Raman + ATR-UV/vis)

59.64% 77.52% 82.87%

N, FST, SST, ET N, SST, ET N, ET

N, FST, SST, ET N, FST, SST, ET N, ET

N, SST, ET N N, FST, SST, ET

noise noise noise

G



Article

flectance, which means that a large amount of variance in that technique is related to noise. If a signal is intrinsically noisy (e.g., NIR reflection and FBRM in the first experiment), most of the variance, which represents the change in the signal itself, will be associated with noise. In the case of NIR reflectance, 55% of the change in the signal is due to events during the process, while the rest is considered noise. PCA is able to include the entire significant signal in the first components and separate this from the noise. Furthermore, the presence of a disturbed signal does not compromise the PCA analysis. Figure 7a shows the scores of the first three principal components in experiment 1, which used NIR reflectance +

the signal keeps increasing for PC2 and decreasing for PC1 until they both reach a stable value at the end of the polymorphic transformation. PC3 is sensitive to both nucleation and the polymorphic transformation; this score has a trend very similar to that for the total counts. The interpretation of the scores of the composite sensor array is easier when two or more scores are plotted on the same graph and clusters and patterns that can be correlated with the various events in the process are found. It is worth noticing that the noisy features present in the FRBM signals in all the experiments are not present in any of the scores shown in Figure 7, demonstrating the advantage of performing PCA on the composite sensor array matrix. Figure 8 shows plots of PC1 versus PC2 for the three experiments: only points for pure form I, pure form II and the

Figure 7. Scores of the first three principal components in (a) experiment 1, (b) experiment 2, and (c) experiment 3.

Raman + FBRM. Nucleation is detected by all of them, as are the beginning and the end of the polymorphic transformation. The clearest principal component is PC2: the signal is not too noisy, and there are considerable changes corresponding to both nucleation and the end of the polymorphic transformation. Furthermore, it is possible to distinguish the two steps of the transformation on the basis of the contribution given by the Raman probe. Figure 7b shows the scores of the first three principal components in experiment 2, which used NIR transflectance + Raman + FBRM. Nucleation is indicated by all of the scores, but the transformation is clear only in PC2. A significant change in the score of PC1 is present corresponding to the transformation, but then the score keeps changing, so the interpretation of this component becomes difficult. However, in PC2 the two steps of the polymorphic transformation are recognizable. Figure 7c shows the scores versus time in experiment 3 which used FBRM + Raman + ATR-UV/vis. In the first two principal components a sudden change is present corresponding to nucleation, and then

Figure 8. Plots of PC1 vs PC2 for the clear solution, pure form I, and pure form II in (a) experiment 1, (b) experiment 2, and (c) experiment 3.

clear solution are presented in order to simplify the identification of the main clusters. The three species are clearly distinguishable in all three plots. In particular, in Figure 8a (NIR reflectance + Raman + FBRM) it can be noticed that form I has a negative PC2 while form II has a positive one. The nucleation and transformation of one form could be easily identified in this system by looking at the sign of PC2. For the CSA formed by NIR transflectance + Raman + FBRM, a similar behavior of the PCs can be noticed, as shown in Figure 8b. In H



Article

for each experiment. The clustering analysis and representation were found to be more effective than the scores plotted versus time and were therefore used as the basis for the loading study. For experiment 1, it is immediately clear that the information from FBRM is less influential than those from Raman and NIR. In panels 1 and 2 in Figure 10, the loading for FBRM is always lower than those for the other probes (although still significant). This tool does not strongly affect any of the rotated loadings for the most important PCs; therefore, it could be excluded from the optimal subset of PAT tools for the combination FBRM + NIR ref + Raman. In the case of experiment 2, PC1 is strongly affected by the Raman signal (see panel 3 in Figure 10), while PC2 is affected mainly by FBRM (panel 4); therefore, those two probes are the most important for understanding the process. Finally, for experiment 3, only the UV/vis and FBRM data strongly influence the first two PCs. The methodology presented here can provide a generic framework for the automated selection of optimal sensor combinations for the detection of particular crystallization events. While for a particular crystallization system and phenomenon of interest one can often recommend the suitable tool, in the context of an automated decision support system, the proposed approach can automate this selection. Such a system could also automatically change the signals collected from multiple sensors and use different sensor configurations present in the crystallization vessels during different phases of the crystallization depending on the phenomenon of interest (e.g., a certain signal combination may provide the highest sensitivity during nucleation, whereas a different set of signals may be best during the polymorphic transformation).

this case, PC1 is negative for form I but positive for form II. In experiment 3 (ATR-UV/vis + Raman + FBRM), PC2 is positive for form I and negative for form II (Figure 8c). Plotting PC1 and PC2 together allows the immediate identification of the form nucleated or transformed. PCA on the same CSA as in experiment 3 (ATR-UV/vis + Raman + FBRM) was performed in the case of nucleation only of the stable form I (experiment 4). Plotting the first two principal components allows the identification of only two clusters: one for the solution and one for the solid (Figure 9). This correlates with the lack of polymorphic transformation in this experiment.

Figure 9. Plot of PC1 vs PC2 for the clear solution and pure form I during a cooling crystallization experiment with no polymorphic transformation.

Discrimination between a clear solution and a slurry is also possible since the cluster for the clear solution is in a different position than those for the slurries of form I, form II, or a mixture. Using more tools is useful to avoid misinterpretation of signals due to operating problems with a single probe (e.g., fouling and particle sticking in the case of FBRM and NIR or fluorescence in the case of Raman). In this way, the use of the CSA enables not only a better understanding of the process but also the identification of sensor failures or measurement problems. However, instead of always using a full CSA, identifying the best subset of PAT tools that should be included in the CSA in order to monitor and identify particular combinations of crystallization mechanisms can save significant costs. The interpretation of the rotated loadings for the first three principal components (the fourth was not included as it showed noise for all of the performed experiments) can help to identify the sensors that most affect the PCs and help in the choice of an appropriate subset. Table 8 shows the most influential probes (the ones that have higher loading values) for the three performed experiments, while Figure 10 shows plots of the rotated loadings. Only the loadings of the PCs used in the clustering analysis (see Figure 8) were analyzed to identify the best subset of tools

4. CONCLUSIONS Nucleation and the polymorphic transformation of OABA from metastable form II to stable form I have been studied using different PAT tools, including FBRM as well as ATR-UV/vis, Raman, and reflectance and transflectance NIR spectroscopy. PVM images were also used to analyze and explain the results obtained using the other probes. It was found that Raman spectra can give the most information about nucleation, crystal growth, and the polymorphic transformation. NIR can also be a powerful tool, but the presence of water in the system limits its usage, although from its baseline effect it was possible to detect nucleation, crystal growth, and the polymorphic transformation. FBRM also provides very useful information about the polymorphic transformation even if it shows lower sensitivity compared with the other tools. The advantages of using more than one PAT tool in a composite sensor array (CSA) configuration to monitor complex processes and capture different phenomena have been demonstrated. The application of the proposed PCA-based data fusion and information extraction approach represents the first centralized CSA structure reported. A real benefit of this approach can be visualized in a processing plant environment where more signals are monitored at the same time. Once a robust process has been established during the development stage and the output of the PCA is well-understood, the monitoring of the signals can be shifted from a decentralized array to a centralized composite array. The major advantage of this method is the reduced amount of signals to monitor, which will greatly assist the plant operators in process monitoring. The signals can be analyzed all together in the CSA or in various subsets of configurations (e.g., two or more probes can be combined) in order to capture the crystallization phenomena of interest. The

Table 8. Most influential probes in the rotated loadings of the first two principal components (used for the clustering) for experiments 1, 2, and 3 experiment 1 (FBRM + Raman + NIR reflectance) 2 (FBRM + Raman + NIR transflectance) 3 (FBRM + Raman + ATR-UV/vis)

rotated loading 1

rotated loading 2

Raman Raman

NIR FBRM

UV

FBRM I



Article

Figure 10. (1) Rotated loading of PC1 (L1) in experiment 1; (2) rotated loading of PC2 (L2) in experiment 1; (3) rotated loading of PC1 (L1) in experiment 2; (4) rotated loading of PC2 (L2) in experiment 2; (5) rotated loading of PC1 (L1) in experiment 3; (6) rotated loading of PC2 (L2) in experiment 3.

■

loadings and the principal component scores can be used to determine the best PAT structure. The approach proposed in this paper can be used to determine the optimal sensor configuration as well as for more efficient and automated information extraction, which are key elements for the implementation of automated intelligent decision support and control systems for crystallization processes. The CSA concept proposed in this paper is very general and can include all types of PAT signals: the choice of the best component is instead system-dependent and varies as a function of the phenomena of interest and the available set of tools.

■

■

ABBREVIATIONS NIR: near-infrared FBRM: focused beam reflectance measurement PVM: particle vision and measurement PCA: principal component analysis CSA: composite sensor array REFERENCES

(1) Guidance for Industry: PATA Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance; U.S. Food and Drug Administration: Rockville, MD, 2004. (2) Chew, W.; Sharrat, P. Anal. Methods 2010, 2, 1412−1438. (3) Yang, Y.; Nagy, Z. K. Cryst. Growth Des. 2014, 14, 687−698. (4) Freitas, M. P.; Sabadin, A.; Silva, L. M.; Giannotti, F. M.; do Couto, D. A.; Tonhi, E.; Medeiros, R. S.; Coco, G. L.; Russo, V. F. T.; Martins, J. A. J. Pharm. Biomed. Anal. 2005, 39, 17−21. (5) Kadam, S.; van der Windt, E.; Daudey, P. J.; Kramer, H. J. M. Cryst. Growth Des. 2010, 10, 2629−2640. (6) Berntsson, O.; Danielsson, L.; Johansson, M. O.; Folestad, S. Anal. Chim. Acta 2000, 419, 45−54. (7) Van Eerdenbrugh, B.; Taylor, L. S. Int. J. Pharm. 2011, 417, 3−16. (8) Wartewig, S.; Neubert, R. H. H. Adv. Drug Delivery Rev. 2005, 57, 1144−1170.

AUTHOR INFORMATION

Corresponding Author

*Tel: (765) 494-0734. E-mail: [email protected]. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS Financial support was provided by the European Research Council (Grant 280106-CrySys). Bruker and Thermo Fisher Scientific are acknowledged for the NIR probes. J



Article

(9) Abu Bakar, M. R.; Nagy, Z. K.; Rielly, C. D. Org. Process Res. Dev. 2009, 13, 1343−1356. (10) Abu Bakar, M. R.; Nagy, Z. K.; Rielly, C. D.; Dann, S. E. Int. J. Pharm. 2011, 414, 86−103. (11) Howard, K. S.; Nagy, Z. K.; Saha, B.; Robertson, A. L.; Steele, G. Org. Process Res. Dev. 2009, 13, 590−597. (12) Howard, K. S.; Nagy, Z. K.; Saha, B.; Robertson, A. L.; Steele, G.; Martin, D. Cryst. Growth Des. 2009, 9, 3964−3975. (13) Gimet, R.; Luong, A. T. J. Pharm. Biomed. Anal. 1987, 5, 205− 211. (14) Aaltonen, J.; Rantanen, J.; Siiria, S.; Karjalainene, M.; Jorgensen, A.; Laitinen, N.; Savolainen, M.; Seitavuopio, P.; Louhi-Kultanen, M.; Yliruusi, J. Anal. Chem. 2003, 75, 5267−5273. (15) Patel, A. D.; Luner, P. E.; Kemper, M. S. Int. J. Pharm. 2000, 206, 63−74. (16) Wang, I.-C.; Lee, M.-J.; Seo, D.-Y.; Lee, H.-E.; Choi, Y.; Kim, W.-S.; Kim, C.-S.; Jeong, M.-Y.; Choi, G. J. AAPS PharmSciTech 2011, 12, 764−770. (17) Tudor, A.; Church, S.; Hendra, P.; Davies, M.; Melia, C. Pharm. Res. 1993, 10, 1772−1776. (18) Deeley, C. M.; Spragg, R. A.; Threlfall, T. L. Spectrochim. Acta, Part A 1991, 47, 1217−1223. (19) Strachan, C. J.; Rades, T.; Gordon, K. C.; Rantanen, J. J. Pharm. Pharmacol. 2007, 59, 179−192. (20) Langkilde, F. W.; Sjöblom, J.; Tekenbergs-Hjelte, L.; Mrak, J. J. Pharm. Biomed. Anal. 1997, 15, 687−696. (21) Campbell Roberts, S. N.; Williams, A. C.; Grimsey, I. M.; Booth, S. W. J. Pharm. Biomed. Anal. 2002, 28, 1135−1147. (22) Al-Zoubi, N.; Koundourellis, J. E.; Malamataris, S. J. Pharm. Biomed. Anal. 2002, 29, 459−467. (23) Hennigan, M. C.; Ryder, A. G. J. Pharm. Biomed. Anal. 2013, 72, 163−171. (24) Rodriguez-Hornedo, N.; Nehm, S. J.; Seefeldt, K. F.; PaganTorres, Y.; Falkiewicz, C. J. Mol. Pharmaceutics 2006, 3, 362−367. (25) Herman, C.; Haut, B.; Douieb, S.; Larcy, A.; Vermylen, V.; Leyssens, T. Org. Process Res. Dev. 2012, 16, 49−56. (26) Barrett, M.; Hao, H.; Maher, A.; Hodnett, K.; Glennon, B.; Croker, D. Org. Process Res. Dev. 2011, 15, 681−687. (27) Su, W.; Hao, H.; Barrett, M.; Glennon, B. Org. Process Res. Dev. 2010, 14, 1432−1437. (28) Caillet, A.; Puel, F.; Fevotte, G. Int. J. Pharm. 2006, 307, 201− 208. (29) Caillet, A.; Puel, F.; Fevotte, G. Chem. Eng. Process. 2008, 47, 377−382. (30) Cornel, J.; Lindenberg, C.; Mazzotti, M. Ind. Eng. Chem. Res. 2008, 47, 4870−4882. (31) Qu, H.; Kohonen, J.; Louhi-Kultanen, M.; Reinikainen, S. P.; Kallas, J. Ind. Eng. Chem. Res. 2008, 47, 6991−6998. (32) Barthe, S. C.; Grover, M. A.; Rousseau, R. W. Cryst. Growth Des. 2008, 8, 3316−3322. (33) Zhao, J.; Wang, M.; Dong, B.; Feng, Q.; Xu, C. Org. Process Res. Dev. 2013, 17, 375−381. (34) Sathe, D.; Sawant, K.; Mondkar, H.; Naik, T.; Deshpande, M. Org. Process Res. Dev. 2010, 14, 1373−1378. (35) Liu, X.; Sun, D.; Wang, F.; Wu, Y.; Chen, Y.; Wang, L. J. Pharm. Sci. 2011, 100, 2452−2459. (36) Simon, L. L.; Nagy, Z. N.; Hungerbuhler, K. Chem. Eng. Sci. 2009, 64, 3344−3351. (37) Kadam, S. S.; Vissers, J. A. W.; Forgione, M.; Geertman, R. M.; Daudey, P. J.; Stankiewicz, A. I.; Kramer, H. J. M. Org. Process Res. Dev. 2012, 16, 769−780. (38) Helmdach, L.; Feth, M. P.; Ulrich, J. Org. Process Res. Dev. 2013, 17, 585−598. (39) Helmdach, L.; Feth, M. P.; Minnich, C.; Ulrich, J. Chem. Eng. Process. 2013, 70, 184−197. (40) Nagy, Z. K.; Braatz, R. D. Annu. Rev. Chem. Biomol. Eng. 2012, 3, 55−75. (41) Nagy, Z. K.; Fevotte, G.; Kramer, H.; Simon, L. L. Chem. Eng. Res. Des. 2013, 91, 1903−1922.

(42) Gemperline, P. Practical Guide to Chemometrics; CRC Press: Boca Raton, FL, 2006. (43) Jiang, S.; ter Horst, J. H.; Jansens, P. J. Cryst. Growth Des. 2010, 10, 2123−2128. (44) Jiang, S.; ter Horst, J. H.; Jansens, P. J. Cryst. Growth Des. 2010, 10, 2541−2547. (45) Jiang, S.; ter Horst, J. H.; Jansens, P. J. Cryst. Growth Des. 2008, 8, 37−43. (46) Liu, J.; Lu, Y.; Liu, J.; Yang, X.; Yu, X. J. Alloys Compd. 2010, 496, 261−264. (47) Simone, E.; Saleemi, A. N.; Nagy, Z. K. Presented at the 20th International Workshop on Industrial Crystallization (BIWIC), Sept 18−20, 2013, Odense, Denmark; pp 168−175. (48) Saleemi, A. N.; Nagy, Z. K. Poster presentation at the UK PharmSci Conference, Sept 14, 2012, Nottingham, U.K.

■

NOTE ADDED AFTER ASAP PUBLICATION The authors have clarified statements in the 2nd, 3rd, and 5th paragraphs of section 3.1. The correct version reposted on September 5, 2014.

K


In Situ Monitoring of Polymorphic Transformations Using a Composite

Recommend Documents