Anal. Chem. 2000, 72, 2099-2108
Quantitative Analysis of Film Coating in a Fluidized Bed Process by In-Line NIR Spectrometry and Multivariate Batch Calibration Martin Andersson,† Staffan Folestad,*,‡ Johan Gottfries,‡ Mats O. Johansson,§ Mats Josefson,‡ and Karl-Gustav Wahlund†
Department of Technical Analytical Chemistry, Center for Chemistry and Chemical Engineering, Lund University, P.O.B. 124, S-221 00 Lund, Sweden, Analytical Chemistry, Pharmaceutical R&D, AstraZeneca R&D Mo¨lndal, S-431 83 Mo¨lndal, Sweden, and Pharmaceutical Technology, Pharmaceutical R&D, AstraZeneca R&D Mo¨lndal, S-431 83 Mo¨lndal, Sweden
A method is described which enables real-time analysis of film coating on pharmaceutical pellets during an industrial manufacturing process. Measurements were conducted on the solid particulate material by nearinfrared (NIR) spectrometry utilizing a diffuse reflectance fiber-optic probe positioned inside a fluidized bed process vessel. Time series of NIR spectra from 11 batches generated a three-way data matrix that was unfolded and modeled by partial least squares (PLS) in a multivariate batch calibration. The process conditions were deliberately varied according to an experimental design. This yielded good predictability of the coating thickness with a best model fit, R2 ) 0.97, for one PLS-projection, and a root-mean-square error of calibration ) 2.2 µm (range tested 0-50 µm). The regression vector was shown to be highly influenced by responses that are both direct (aliphatic C-H stretch overtones) and indirect (aromatic C-H stretch overtones), from film component and core material, respectively. The impact of different data pretreatment methods on the normalization of the regression vector is reported. Justification of the process calibration approach is emphasized by good correlation between values predicted from NIR data and reference image analysis data on dissected pellets and a theoretical nonlinear coating thickness growth model. General aspects of in-line NIR on solids and multivariate batch calibration are discussed.
INTRODUCTION The possibility of monitoring chemical reactions in industrial processes by vibrational spectrometry such as near-infrared (NIR) and Raman has lately received increased attention.1-3 Through the use of fiber-optic probes, noninvasive measurements can be * Corresponding author. Fax: +46-31-7763727. E-mail: staffan.folestad@ astrazeneca.com. † Lund University. ‡ Analytical Chemistry, Pharmaceutical R&D, AstraZeneca R&D Mo ¨lndal. § Pharmaceutical Technology, Pharmaceutical R&D, AstraZeneca R&D Mo ¨lndal. (1) Callis, J. B.; Illman, D. L.; Kowalski, B. R. Anal. Chem. 1987, 59, 624A635A. 10.1021/ac990256r CCC: $19.00 Published on Web 03/28/2000
© 2000 American Chemical Society
carried out in process streams by these techniques. Referring to the conceptual framework by Kowalski and co-workers,1 this can be classified as “process analytical chemistry” (PAC) conducted “in-line.” That is, when the sample interface is located inside the process vessel, the chemical analysis is done in situ, thus omitting the need for transport of sample out of the process vessel. Still, analysis on process samples can also be carried out “at-line” on a spectrometer located in the manufacturing area after manual sampling. The general benefit of vibrational spectrometry for PAC is that fast analysis in the order of seconds can be provided, including instrumental measurement and any sampling. Several examples of process monitoring and process analytical chemistry using NIR spectrometry have been demonstrated including both qualitative4-7 and quantitative8-19 analysis. Process applications have, for example, been reported for a wide range of samples including solids,4-13 polymer melts,14-18 and liquids.19 (2) Blaser, W. W.; Bredeweg, R. A.; Harner, R. S.; LaPack, M. A.; Leugers, A.; Martin, D. P.; Pell, R. J.; Workman, J., Jr.; Wright, L. G. Anal. Chem. 1995, 67, 47R-70R. (3) Hassell, D. C.; Bowman, E. M. Appl. Spectrosc. 1998, 52, 18A-29A. (4) Wargo, D. J.; Drennen, J. K. J. Pharm. Biomed. Anal. 1996, 14, 14151423. (5) Sekulic, S. S.; Ward, H. W.; Brannegan, D. R.; Stanley, E. D.; Evans, C. L.; Sciavolino, S. T.; Hailey, P. A.; Aldridge, P. K. Anal. Chem. 1996, 68, 509513. (6) Sa´nchez, F. C.; Toft, J.; Bogaert, B.; Massart, D. L.; Dive, S. S.; Hailey, P. Fresenius’ J. Anal. Chem. 1995, 352, 771-778. (7) Hailey, P. A.; Doherty, P.; Tapsell, P.; Oliver, T.; Aldridge, P. K. J. Pharm. Biomed. Anal. 1996, 14, 551-559. (8) Kirsh, J. D.; Drennen, J. K. J. Pharm. Biomed. Anal. 1995, 13, 1273-1281. (9) Kirsh, J. D.; Drennen, J. K. Pharm. Res. 1996, 13, 234-237. (10) Han, S. M.; Faulkner, P. G. J. Pharm. Biomed. Anal. 1996, 14, 1681-1689. (11) Buchanan, B. R.; Baxter, M. A.; Chen, T.-S.; Xue, X.-Z.; Robinson, P. A. Pharm. Res. 1996, 13, 616-621. (12) Rantanen, J.; Lehtola, S.; Ramet, P.; Mannermaa, J. P.; Yliruusi, J. Powder Technol. 1998, 99, 163-170. (13) Frake, P.; Greenshalgh, D.; Grierson, S. M.; Hempenstall, J. M.; Rudd, D. R. Int. J. Pharm. 1997, 151, 75-80. (14) Aldridge, P. K.; Kelly, J. J.; Callis, J. B. Anal. Chem. 1993, 65, 3581-3585. (15) Fischer, D.; Bayer, T.; Eichhorn, K.-J.; Otto, M. Fresnius’ J. Anal. Chem. 1997, 359, 74-77. (16) Mockel, W. D.; Thomas, M. P. Proc. SPIE-Int. Soc. Opt. Eng. 1992, 1681(Opt. Based Methods Process Anal.), 220-230. (17) Khettry, A.; Batra, J.; Stewart, D. A.; Hansen, M. G. Annu. Tech. Conf.sSoc. Plast. Eng. 1992, 50, 2674-2676. (18) van den Berg, F. W. J.; Osenbruggen, W. A.; Smilde, A. K. Process Control Qual. 1997, 9, 51-57. (19) Hansen, M. G.; Khettry, A. Annu. Tech. Conf.sSoc. Plast. Eng. 1995, 53, 2820-2823.
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000 2099
In the pharmaceutical industry, high-quality demands on drug products necessitate high quality not only in raw materials but also in manufacturing processes and processing. Spectrometric techniques in this context thus offer attractive possibilities for novel process control. Quantitative analysis of coating on tablets was presented by Kirsh and Drennen8,9 who conducted the measurements at-line (tablets withdrawn from the process). Andersson et al.20 developed an at-line method for determination of the amount of coating on tablets in which whole diffuse reflectance NIR spectra were used in a multivariate calibration. They indicated that imprecision in the analysis may be due to an uneven distribution of the amount of tablet coating between tablets which is why the method should be useful for studying coating homogeneity. Despite these successful examples, there is so far a limited number of reports on in-line analysis by NIR in fluidized bed processes. These processes are commonly utilized for drying, granulation, or coating.21 In-line analysis of the moisture content in 0.05-0.7 mm pellets during spray granulation in a fluid bed has been demonstrated by Frake et al.,13 who carried out the calibration in a univariate way. Rantanen et al.12 described a similar moisture content measurement using ratioing of 3-4 selected wavelengths. To the best of our knowledge, there are so far no reports on quantitative in-line NIR measurements on pellets during coating in a fluidized bed process. This process is used to encapsulate pellets containing the active drug substance with a membrane, e.g., a polymer coating. Ultimately, the purpose of such a membrane may be to obtain a controlled release rate of the active drug substance, resulting in therapeutically optimized plasma levels. A method to achieve the membrane is to apply the coating in a fluidized bed process in which the coating thickness of an individual pellet is increased every time the pellet is hit by spray droplets. Important requirements for the process are that it should result in a predetermined average coating thickness and, implicitly, a specified surface area.22,23 Other properties such as degree of agglomeration of pellets and the density24 and porosity of the coating are also important but are not further discussed here. The objective of this study was to evaluate the possibilities for quantitative calibration of NIR spectra acquired in-line for estimation of the coating thickness on pharmaceutical pellets and to enable the process end point to be determined. The approach taken was to systematically alter the process conditions in a multivariate way by means of an experimental design, inside and outside of the range constituting normal processing. Multivariate calibration25 was used to enable monitoring of the growth of the film coating via an NIR diffuse reflectance fiber probe. It was of particular importance to stabilize the calibration toward variations in the process conditions that may affect the NIR response but, in parallel, may not be related to the film growth. (20) Andersson, M.; Josefson, M.; Langkilde, F. W.; Wahlund, K.-G. J. Pharm. Biomed. Anal. 1999, 20, 27-37. (21) Arwidsson, H. G.; Rude´n, M. In Industrial Aspects of Pharmaceutics; Sandell, E., Ed.; Swedish Pharmaceutical Press: So ¨derta¨lje, 1993; pp 212-226. (22) Ragnarsson, G.; Johansson, M. O. Drug Dev. Ind. Pharm. 1988, 14, 22852297. (23) Ragnarsson, G.; Sandberg, A.; Johansson, M. O.; Lindstedt, B.; Sjo ¨gren, J. Int. J. Pharm. 1992, 79, 223-232. (24) Senjekovic’, R.; Jalsenjac, I. Pharm. Acta Helv. 1982, 57, 16-19. (25) Martens, H.; Næs, T. Multivariate Calibration; John Wiley & Sons: Guildford, 1989.
2100
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
Notably, pioneering work on multivariate statistical procedures for monitoring the progress of batch processes was recently reported by Nomikos and MacGregor.27,28 They used partial least squares (PLS) to extract information from process measurement trajectories in which time series of nine process parameters were used to model the resulting product properties (five variables). The generated three-way array was unfolded such that a normal PLS could be performed. This methodology has been extended by Wold et al.,29 who conducted the PLS analysis on an unfolded three-way data matrix using “local process time” instead of product properties as the response. In the present study, the PLS analysis is carried out in a similar way, albeit the primary objective here is to calibrate a spectrometric sensor for the purpose of making it robust to varying process conditions. EXPERIMENTAL SECTION Apparatus. Coating of pellets was performed in a fluidized stainless steel bed apparatus constructed in-house. The process vessel was a laboratory-scale version of production-scale manufacturing equipment of the Wurster type.21 Briefly, the spray gun through which the coating solution is pumped is placed at the bottom of the vessel, see Figure 1A. Located above the gun is the Wurster column that serves to direct the pellets while they cycle in and out of the spraying zone. To minimize the effect of static electricity, and in this way to avoid pellets being stuck on the walls, the process vessel was grounded. NIR diffuse reflectance spectra were acquired with a dispersive single beam instrument (NIRSystems 6500, NIRSystems Inc., Silver Spring, MD) using a 0.5-in o. d. fiber-optic probe (NIRSystems Smartprobe, NIRSystems Inc.)30 that was mounted in the process vessel as shown in Figure 1A. To secure a representative sampling during processing, a sample collector that was emptied by compressed air was used inside the vessel. The sapphire window at the probe tip constituted the interface of the probe to samples in the process vessel. The wavelength range scanned by the spectrometer was 1100-2500 nm (2-nm steps, spectral bandwidth of 10 nm). Individual nonaveraged spectra were stored for subsequent multivariate analysis. The spectrometer used a polystyrene standard for internal wavelength calibration. A white ceramic standard (70% reflectivity) was used as the external reference to check for instrumental drift. The difference in the reference spectrum before and after each batch run appeared mainly as a baseline shift, corresponding to less than 0.002 log 1/R units (peak-to-peak). The total time for scanning, internal operations within the spectrometer, and storage of spectra was estimated to be 7.8-7.9 s. It should be noted that the scanning time of the sample was only a minor part thereof. It was determined to be ∼0.25 s by means of a stroboscope light source flashing at 100 Hz (see Figure 1B). Materials. Uncoated pellets containing the active drug substance were sieved to a diameter of ∼400-500 µm. Their shape (26) Zhou, F.; Vervaet, C.; Massart, D. L.; Massart, B.; Remon, J. P. Drug Dev. Ind. Pharm. 1998, 24, 353-358. (27) Nomikos, P.; MacGregor, J. F. Technometrics 1995, 37, 41-59. (28) Nomikos, P.; MacGregor, J. F. Chemom. Intell. Lab. Syst. 1995, 30, 97108. (29) Wold, S.; Kettaneh, N.; Fride´n, H.; Holmberg, A. Chemom. Intell. Lab. Syst 1998, 44, 331-340. (30) NIRSystems Intrument Performance Test Guide; NIRSystems Inc., Silver Spring, MD, 1995.
covariance between a linear combination of X variables (NIR spectra) and a linear combination of Y variables (coating thickness) is maximized in a series of linear models using the NIPALS algorithm.25 Important constraints32 are, e.g., orthogonality between the columns in the score matrix, T. The resulting model
{
Figure 1. (A) Schematic of the fluidized bed process vessel used for coating of pellets. P denotes the position of the fiber optic probe; S, the spray gun; A, the air flow for fluidizing the pellet bed; W, the Wurster column. (B) NIR spectrum obtained by illuminating the probe with 100 Hz stroboscope light flashes. (C) Conceptual illustration of the sampling intervals.
was approximately spherical with a ratio of maximum to minimum radii in the range of 1.1-1.8 for 95% of the pellets. The coating solution was composed of ethyl cellulose dissolved in ethanol with a water content between 2 and 8%. Approximately 1/2 kg of pellets was used for each batch experiment, and dry air ( 1, the target coating volume is exceeded, which can be of interest when calibrating NIR spectra to the coating thickness to obtain a working range for the NIR predictions also above the target coating thickness. lcoat,pellet is obtained as a function of z by substituting Vcoat,pellet in eq 4 for zqVfeed,target
((
lcoat,pellet ) rcore 1 +
qVfeed,target z Vcore
) ) 1/3
-1
(7)
(2)
The total volume of the pellet after application of a layer of coating can be described by another sphere having the volume
Vcore + Vcoat,pellet )
During the processing, Vcoat,pellet increases. It can be written as a fraction, z, of Vcoat,target
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
(5)
where q is assumed to be constant within a batch. Vcore and rcore are characteristic parameters for the pellets. The mean values of these parameters were experimentally determined by image analysis to be rcore ) 2.3 × 102 µm (n ) 312 pellets), which corresponds to Vcore ) 5.4 × 107 µm3. Dissected pellets were studied at three different levels of z (0.75, 1, and 1.25), and the coating thicknesses, lcoat,pellet, were determined accordingly to be 37, 45, and 52 µm, respectively, with the number of pellets studied being 60, 60, and 192, respectively. Nonlinear curve fitting by the least-squares method gave qVfeed,target ) 3.6 × 107 µm3. Using eq 7, lcoat,pellet can now be plotted as a function of z (Figure 4A, solid line). The fit seems reasonable, taking into account that the pellets are not perfect spheres and that the growth model may represent a simplification of the real situation. The fitted curve should therefore be useful in predicting the coating thickness. Assuming that the density of the coating is constant for all of the batch experiments, z and q can be determined as mass fractions, which may be more useful in practice. Preprocessing of NIR Data. Spectra collected while running the coating process were observed to have baseline levels irregularly spaced, see upper traces in Figure 3. This could be explained by instrumental baseline shifts or differences in the packing of the pellets, or by differences in particle size within the batch. To handle different baseline levels, different spectral data pretreatments were tested. These were multiplicative scatter correction (MSC), standard normal variate (SNV),25,40 and second (40) Dhaona, M. S.; Lister, S. J.; Sanderson, R.; Barnes, R. J. J. Near Infrared Spetrosc. 1994, 2, 43-47.
Figure 5. Regression vectors from PLS models used for prediction of coating thickness for mean-centered raw NIR data. Large absolute values indicate important spectral regions. Upper trace: secondderivative pretreatment model (no. 5), offset by 0.2 log(1/R) units for display purposes. Lower traces: dotted line, no pretreatment model (no. 2); solid line, MSC pretreatment model (no. 3, solid line); dashed line, SNV-pretreated model (no. 4). See Table 2 for further details on PLS models.
Figure 4. (A) Theoretical coating growth model (solid line, eq 7) fitted to experimental data (B) determined by image analysis. z is the fraction of the target coating volume that has been applied as coating onto the pellets; lcoat,pellet is the coating thickness. The error bars show the 95% confidence interval of the average coating thickness, respectively. (B) Calibration curve for prediction of film thickness obtained by the theoretical growth model and NIR data. PLS model no. 5 in Table 2.
derivative.41,42 The results are compiled in Table 2. For both MSC and SNV, the whole spectrum (1100-2500 nm) was used to adjust the baseline level, and in Figure 5 it is shown that the respective regression vectors are almost identical. Calibrations using the second derivative (model 5, Table 2) showed the best results among the pretreatment methods tested. It was calculated using the method of Savitzky-Golay43,44 with a 15-points (30 nm) wide window regressing to a quadratic polynomial. The calibration curve corresponding to model 5 is shown in Figure 4B. The second derivative appeared to suppress scattering more effectively than SNV and MSC. This is indicated in the regression vector, in which three distinct wavelength regions can be found (11001250, 1300-1450, and 1600-1800 nm). These regions have strong influence on the predictions of the coating and correlate well with absorption from the cellulose backbone. Still, a major drawback (41) Talsky, G. Derivative Spectroscopy; VCH: Weinheim, Germany, 1994. (42) Blanco, M.; Coello, J.; Iturriaga, H.; Maspoch, S.; Pezuela, C. Appl. Spectrosc. 1997, 51, 240-246. (43) Savitzky, A.; Golay, M. Anal. Chem. 1964, 26, 1627-1639. (44) Steiner, J.; Termonia, Y.; Deltour, J. Anal. Chem. 1972, 44, 1906-1909.
of using second-derivative calculations is that the model becomes more sensitive to chemical and instrumental wavelength shifts, as is seen in the wavy regression vector in Figure 5 (upper trace). Thus, pretreatments using MSC, SNV, or no pretreatment might be regarded as more robust compared with the second derivative. Note that the regression vector for the second derivative shown in Figure 5 is used for raw NIR data and not for second-derivative NIR spectra. This emanates from the fact that a derivative spectrum, xD, can be determined by a nonisometric rotation of the raw spectrum x with a rotation matrix D, such that XD ) xD. Projection of the derivative spectrum onto the determined regression vector b yields a predicted value, yˆ ) xDb ) xDb ) xbD. The objective of this operation is to move the derivation from the spectrum to the regression vector. In this way, the contribution from different regions in unmodified raw spectra could be compared for all models. NIR Band Assignment. By inspecting the regression vector in detail, different spectral regions contributing to the predictions of the ethyl cellulose coating thickness can be judged. The regions that have a strong influence on the predictions vary, depending on what type of pretreatment method is employed, see Table 3. For all models it could be seen that indirect absorption measurements, i.e., a vanishing signal from the first overtone of the aromatic C-H-stretch of the active drug substance, was important in the multivariate calibration. The “baseline” between 1250 and 1300 nm was found to contribute only weakly in all models. Other regions, such as overtones and combination bands of aliphatic C-H-stretches also contributed but with quite varying results. It is therefore apparent that different pretreatment methods emphasize different spectral regions to varying extents. Since these correspond to different physicochemical properties of the sample, the different pretreatment methods yield different normalizations of the absorption bands. Evidently, scattering responses due to scattering are not fully normalized for, but are solely suppressed to a varying degree. This is, however, not surprising since scattering coefficients for this type of particulate material (coarse powder) are affected at regions of strong absorption.38,45 Test Batches. Spectra from a number of different batches were collected and used to test the multivariate batch calibration. Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
2105
Figure 6. Application of the multivariate batch process calibration on two test batches. An alarm limit for the residuals has been arbitrarily set to three standard deviations (SD). (A) Normal evolvement of the process (run no. 12, Table 1). Predicted values and spectral residuals are smoothed by a 5-point moving average. (B) Possibilities of detecting deviations from normal process behavior through real-time NIR measurements of film thickness (nonsmoothed data). The dashed path, indicating normal behavior, was calculated as (3 SD using eq 8. The experimentally determined film thickness for batch run no. 13 (see Table 1) is shown in the upper trace (solid line). This batch was deliberately run with half of the amount of starting material. Three tentative points for terminating the process are indicated: (i) >5 consecutive measurements outside the (3 SD path; (ii) measured thickness equal to target amount of coating; (iii) normal end point of coating as determined only from processing time.
These were designed to evaluate the calibration within and outside of the calibration domain. An example of normal process behavior is shown in Figure 6A, run no. 12 in Table 1. The predicted values (and the spectral residuals) are smoothed along the time axis using a 5-point moving average to obtain a better precision in predicting the overall process. This inevitably yields a predicted value that is somewhat below the expected in a growth trend curve, see Figure 6A. Still, the trend with respect to coating growth is followed. In contrast to the spectra used for calibration, no spectra were removed for the test batches in order to be able to detect any deviations from normal behavior. As expected, some of the spectra in the process initiation phase gave larger spectral residuals after projection onto the PLS model. This is exemplified in Figure 6B. In test batch run no. 13 (Table 1), the amount of uncoated pellets was only half of the normal amount. When looking at the (45) Burger, T.; Kuhn, J.; Caps, R.; Fricke, J. Appl. Spectrosc. 1997, 51, 309317.
2106
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
coating thickness estimated by the NIR model, it became obvious that the growth rate of the coating was double that of the normal coating rate, Figure 6B (nonsmoothed data). This was also a test for the ability to extrapolate the PLS model outside of the calibration domain. The accuracy was found to be good (within 10%) even at double the amount of coating applied. In Figure 6B the encircled time regions (i, ii, and iii) denote three different stages of the process. At time i the process was observed to systematically move outside three standard deviations, corresponding to outside of the 99% confidence interval of normal process behavior. From the spectral residuals it is obvious that below a thickness of 5 µm the measurements are not reliable, and thus, NIR is only applicable after the process initiation. This emphasizes that spectral residuals alone are not sufficient to identify altering process conditions. The control action here would be to abort the process at this early stage because the process did not behave normally. However, since the pellet spectra looked normal, an alternative action would be to continue processing to time ii at which the specified end-point is reached. This points to the benefits of a spectrometric process analyzer since running the process in a common way, i.e., a specified time, would result in a coating thickness corresponding to point iii. The test of the batch calibration was then continued with another test batch (run no. 14) of pellets having diameters different from those in the calibration set. At first, it appeared to be monitorable using the calibration model without modifications. However, a bias in the predictions could be noted, which has also been observed in other multivariate calibrations on process samples measured at-line or in-line.16,18,48 The bias can be partially attributed to a change in the spectral response that appears as a wavelength shift. If wavelength shifts are at hand, instrumentally or physicochemically caused, an accuracy problem will occur, which will show up as a bias error. Wavelength shifts were simulated on the calibration data set. In this way, the effect of a wavelength shift in raw spectral data and second-derivative data could be tested. From the results, shown in Figure 7, it is clear that the wavy regression vector for the second-derivative model will introduce a larger bias than that introduced by a model using raw spectral data. Models for MSC and SNV pretreated spectra yielded a bias similar to that for the raw spectra since the degree of bias is directly correlated to the wavy character of the regression vector. Shortcut to Calibration. At this stage of the study, an alternative approach to obtain a shortcut to calibration was tested. By varying all process conditions in one single batch (run no. 15) instead of 11 separate batches and thereby conducting a single batch calibration, time could be saved. That is, the process conditions previously used for run no. 1 were used the first 10 min of run no. 15. Thereafter, the process conditions were changed to those previously used in run no. 2. These conditions were then maintained for ∼10 min and subsequently altered accordingly in further steps during the remainder of the process. This resulted in a model with RMSEC ) 2.5 µm, in close (46) Wold, S.; Albano, C.; Dunn, W.; Edlund, U.; Esbensen, K.; Geladi, P.; Hellberg, S.; Johansson, E.; Lindberg, W.; Sjostrom, M. M. In Chemometrics: Mathematics and Statistics in Chemistry; Kowalski, B. R., Ed.; Reidel Publishing Co.: Dordrecht, Holland, 1984; pp 17-95. (47) Kresta, J. V.; MacGregor, J. F.; Marlin, T. E. Can. J. Chem. Eng. 1991, 69, 35-47. (48) Podulski, D. E. Chem. Eng. Prog. 1997, 10, 33-46.
Figure 8. Deviation in growth of film-coating thickness measured by NIR in relation to the theoretical rate calculated using the growth model eq 7. ∆yNIR and lcoat,pellet denote differences in thickness between two consecutive samples, measured experimentally and derived from the theoretical growth model, respectively.
Figure 7. Simulation of wavelength shifts in NIR raw data and their impact on PLS prediction of coating thickness. Data pretreatment: (A), none; (B), second derivative (15-points quadratic Savitzky-Golay filter). The solid line represents the ideal predicted versus measured regression line.
agreement with the model obtained in the full batch calibration (11 batches). The predictive ability of the single batch calibration was tested by using one of the experiments from the full calibration set, now as a test batch (run no. 1). This resulted in a RMSEP as good as 2.1 µm, (R2 ) 0.99, slope ) 0.93, intercept ) 3.4 µm). Still, more studies are needed to develop this simplified methodology. Coating-Thickness Variations. When looking at the coating thickness estimated from NIR spectra, the thickness appears to vary more as the amount of coating increases, see Figure 6B. To study the increase in variance of the NIR measurements, the increase between two consecutive NIR measurements was compared with the expected increase determined by the growth model during the same time interval. The relative deviation in the increase can thus be determined as:
erel )
∆yˆNIR - ∆lcoat,pellet lcoat,pellet
(8)
where ∆yˆNIR is the increase in coating thickness between two consecutive measurements determined by NIR, and ∆lcoat,pellet is the expected increase in coating thickness determined from lcoat,pellet (calculated from eq 7) for two consecutive samplings. Individual values of erel (Figure 8) show that the relative standard deviation of the film thickness is constant (thickness > 5 µm) and close to normally distributed. Indeed, this is in accordance with what has been reported by Wnukowski49 for the fluid bed coating process. Typically, a larger variation in coating thickness is obtained since the pellet material is cycled in and out of the
spraying zone and the number of cycles for different pellets are random. It should, therefore, be noted that the variation in the coating thickness predicted by NIR is not due to imprecision in the fiber-probe measurements but rather reflects the expected distribution in film thickness. At the target amount of coating (47 µm), the standard deviation of the NIR-predicted film thickness is 2.4 µm. This should be compared with the repeatability of the NIR measurements. Measuring the same sample inside the process 100 times without resampling gave a repeatability in terms of standard deviation of only 0.9 µm. In addition, this can be compared with the variance of film thickness between pellets, which corresponds to a standard deviation of 5.1 µm, as obtained from image analysis. CONCLUSIONS The objective of this study was attained, i.e., to enable realtime analysis of film coating in an industrial manufacturing process. A major benefit with the developed in-line NIR method is that the coating can be judged not only qualitatively but also quantitatively during the evolvement of the process. In particular, this enables the end-point to be determined with high precision. The concept of multivariate batch calibration was shown to be useful in stabilizing model predictions and, in addition, may also provide data for process optimization. Here, the use of an experimental design is the key to obtaining an effective set of experiments if the manufacturing conditions are also to be optimized. The novel approach presented, to arrange all the experiments in one single batch run, demonstrates the potential of a more cost-effective way to accomplish multivariate batch calibration. Despite the promising results, further studies are needed both to develop this methodology and to evaluate its general applicability. An interesting result of this study is that, aside from the possibility of determining the mean value of the coating thickness with high precision, an estimate can also be obtained of its variation. Thus, formerly unknown details of the process are revealed, which is common when in-line measurements are (49) Wnukowski, P. On the coating of particles in fluid-bed granulators. Ph.D. Thesis, Royal Institute of Technology, Stockholm, Sweden, 1989.
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
2107
introduced. Regarding this, a prerequisite is that a sufficient sampling rate is used in the measurements and that the fiberoptic probe has been optimized and characterized for this purpose. Additionally, this points out the need for a more in-depth characterization of the true variations in film thickness on the single-pellet level so that relevant reference data become available. In this context, in-line NIR may thus be further developed to provide a means for gaining information on the underlying mechanisms of the coating process, the process dynamics, and how the flow of particles in the process is distributed.
2108
Analytical Chemistry, Vol. 72, No. 9, May 1, 2000
ACKNOWLEDGMENT Prof. Jo¨rgen Vessman is gratefully acknowledged for scientific discussions on process analytical chemistry and Dr. Erik Johansson, Umetrics AB for valuable comments on the multivariate batch calibration. This work was made possible with support from Astra Ha¨ssle AB. Received for review March 4, 1999. Accepted June 20, 1999. AC990256R
