UNIDAD ELECTRÓNICA DE CONTROL (ECU) [20]

Before calibrations were attempted for both materials, the data of each was split into two sets: a calibration set of mainly sieved fractions (microcrystalline cellulose {n = 24) and lactose {n = 15)) and a validation set of bulk samples (microcrystalline cellulose {n

= 33) and lactose {n = 18)). With each calibration set, highly significant correlation {p <

0.005) was obtained between NIR predicted InJso and ln(FALLS d^d) values (Table 2.1). However, the validation sets for each material exhibited more scatter (Table 2.1). With the microcrystalline cellulose prediction set of bulk samples (Avicel grades PH 101 and PH 102), significant correlation { p < 0.005) between NIR predicted \nd50 and ln(FALLS

dso) was obtained with a SEP greater than the SEC (Table 2.1). The high SEP is possibly accounted for by the FALLS and scanning electron microscopy (SEM) results

Table 2.1. Microcrystalline cellulose and lactose MLR calibration (sieve fraction data) and validation (bulk sample data) results.

M a t e r i a l M i c r o c r y s t a l l i n e c e l l u l o s e L a c t o s e m o n o b y d r a t e b o ' - 3 . 9 9 4 . 5 9 b , ' 5 9 . 6 5 - 1 7 2 . 3 b : - 5 7 . 9 9 1 6 7 . 8 W a v e n u m b e r 1 ( c m ') 8 2 4 4 6 0 1 2 W a v e n u m b e r 2 ( c m " ') 5 9 6 4 5 9 4 0 S E C ( l n ( * n / f t m ) ) 0 . 0 6 7 0 . 0 9 7 S E P ( l n ( ( / ; , ) / |a m ) ) 0 . 1 7 0 . 1 8 In P = c + m l n ( F A L L S d s o ) C a l i b r a t i o n s e t r 0 . 9 9 0 . 9 9 m 0 . 9 9 0 9 8 c 0 . 0 3 5 & % 8 n 2 4 ( P H l O l , P H 1 0 2 , P H 2 0 0 s i e v e d ) 1 5 ( S i e v e d a n d 1 1 0 n V a l i d a t i o n s e t r 0 . 8 4 0 . 0 1 4 m 0 . 9 6 0 . 0 1 8 c 0 . 1 7 4 . 3 8 n 3 3 ( P H l O l & P H 1 0 2 , b u l k ) 1 8 ( F a s t f l o s a m p l e s ) * M L R c o e f f i c i e n t s : h o - i n t e r c e p t , h , - w a v e n u m b e r 1, a n d l?2 - w a v e n u m b e r 2 . r i s c o r r e l a t i o n c o e f f i c i e n t ; m a n d c a r e s l o p e a n d i n t e r c e p t o f p l o t s o f N I R p r e d i c t e d In rA o v s . F A L L S m e a s u r e d I n J s o ; ti i s t h e n u m b e r o f s a m p l e s i n e a c h d a t a s e t . ____________ _________ ____________________________ __________________ ____________ _____________________________

Fig. 2.4. SEM photographs. (A) Avicel PHlOl bulk sample, (B) Avicel PH200 > 200 pm sieve fraction, (C) lactose monobydrate < 31 pm sieve fraction and (D) lactose monobydrate > 150 pm sieve fraction.

(Fig 2.4A & 2.4B) which showed that these bulk samples had broad distributions and comprised a mixture of irregularly shaped fines and large spherical particles. Previous work (Kortum, 1969) has shown that this can produce more variable results than the use of narrow or uniform-size distributions as the NIR scattering and absorbing properties of these particles will be different to that of median sized particles.

Validation of the lactose calibration used bulk samples of Fastflo from 18 different batches. This spray dried material generally has a relatively uniform and spherical particle size (Pearce, 1986). A narrow range of <^50 was confirmed by FALLS (Range dsQ. 81.1 - 115.7 |im ) . Poor correlation was obtained between NIR predicted InJso and ln(FALLS dso) with these samples and is probably due to the narrow range of particle size in the prediction set as the SEP is not significantly different from that of the microcrystalline cellulose prediction set of bulk samples (Table 2.1).

SEM results of the lactose sieve fractions used in the calibration set showed small, irregularly shaped fines in the smallest sieve fractions and large spherical particles in the largest sieve fractions (Fig 2.4C & 2.4D), much the same as with the

microcrystalline cellulose calibration set.

2,8.5.2 Particle Size Calibration Using Randomised Sieve Fraction A n d Bulk Sample Data

To produce working calibrations with the microcrystalline cellulose and lactose data sets, the sieve fraction and bulk sample data were randomly assigned to either the calibration set (67% of spectra) or validation set (33% of spectra). This procedure was repeated three times to test the robustness of the method, giving three different

calibration and validation sets for the two materials. Both sieve fractions and bulk samples were used in calibrations as preliminary investigation showed that this produced more robust calibrations.

In each case, all three calibrations employed slightly different combinations of wavenumbers. The selected wavenumbers were found to occur on the slopes of overtone peaks; the selection of each wavenumber is therefore likely to have been influenced by the random noise in each data set. With both materials, each of the three MLR calibrations showed a good fit between NIR spectral and FALLS data

(microcrystalline cellulose: SEC (ln(J5o/|Lim)) = 0.10 - 0.11^ and lactose monohydrate:

SEC (ln(J5o/|im)) = 0.12 - 0.13^). This was confirmed by plots of NIR predicted InJso

versus ln(FALLS d s o ) which showed significant linear association (microcrystalline

cellulose: r = 0.98 (n = 38 for each set) and lactose monohydrate: r = 0.97 - 0.98^ (n = 22 for each set); in each case with p < 0.005) (Figs 2.5A & 2.6A). The three validation sets for each material showed similar results (Figs 2.5B & 2.6B) with highly significant correlation between NIR predicted InJso and ln(FALLS d s o ) (microcrystalline cellulose (« = 19): r = 0.98, lactose monohydrate (n = 11): r = 0.93 - 0.97;^ in each case p <

0.005), and SEP comparable to SEC, microcrystalline cellulose: SEP (ln(J5o/jim)) =

0.12 - 0.14,^ and lactose monohydrate: SEP (ln(6f5o/|im)) = 0.15 - 0.21^).

2.9 Measurement of The Cumulative Percentage Frequency Particle Size

Distribution

In this Section, NIR measurements of powdered microcrystalline cellulose are calibrated to measure the cumulative percentage frequency particle size distribution. Two different chemometric methods are compared: 3-wavenumber MLR and 3 principal components regression (PCR).

^ The range gives the minimum and maximum values observed for the three randomly selected calibration and validation sets.

10 Q . ,1 10 ₁ .2 ,3 10 10 10 10 g ,2 10 Q . .1 10 ,1 .2 .3 10 10 10 FALLS FALLS

Fig. 2.5. Results of microcrystalline cellulose MLR calibration with randomised sieve fraction and bulk sample data. NIR measured median particle size, dso, versus FALLS dso. (A) Calibration set and (B)

validation set. 10 s • a ,2 10 a. ,1 10 ,1 ,2 ,3 10 10' 10 10 ■o ,2 10 Q . ,1 10 ,2 .1 .3 10 10' 10

FALLS dgg/|im FALLS

Fig. 2.6. Results of lactose monobydrate MLR calibration with randomised sieve fraction and bulk sample data. NIR measured

median particle size, dso, versus FALLS dso. (A) Calibration set and (B) validation set.

2.9.1 Preliminary Investigation

In Section 2.7, it was shown that useful calibrations for median particle size can be obtained by using NIR reflectance data with a logarithmic transform of the FALLS particle-size data, hence these data have been used in this section.

With MLR calibrations, preliminary work showed that a 3 wavelength linear regression at any of the FALLS quantités produced calibrations more robust than a two wavelength fit. It was therefore decided that three wavelength MLR calibrations would be employed subsequently. With PCR models, three principal components were required to produce satisfactory calibrations and this number was used for all subsequent calibrations.

2.9.2 Spectral Characteristics

The spectra of each powdered sample exhibited the effects of multiple scatter, as described in Section 2.8.2 (Fig. 2.1 A).

2.9.3 Model Generation

The FALLS instrument gives values of the cumulative percentage frequency particle- size distribution at 64 particle sizes (range: 564 to 5.8 pm), at intervals which follow a geometric progression. For each sample, linear interpolation of the measured FALLS values was used to calculate the particle size values corresponding to the 5,10, 20, 30, 40, 50, 60, 70, 80, 90 and 95% quantiles (Appendix D, CD-ROM inside back cover). The samples exhibited a wide range of particle sizes at each quantile (Table 2.2) and a wide variety of distributional shapes. Of the 57 samples available, 34 were chosen at random for the calibration set; the remaining 23 samples were used as an independent validation set. To aid comparison of the two calibration methods, the same calibration and validation data were used for each method.

Table 2.2. Particle size ranges at each quantile for the calibration and validation

sets as determined by FALLS.

Quantile Particle size/|im

Calibration set (n= 34) Validation set (n= 23)

Minimum Median Maximum M inimum Median M aximum

5% 6.45 25.72 216.52 7.21 23.06 167.11 10% 9.92 37.14 268.91 11.44 32.34 187.67 20% 14.48 52.92 311.96 18.05 45.40 219.33 30% 18.39 67.10 345.62 22.55 56.36 251.13 40% 21.40 81.41 376.44 26.27 67.35 283.66 50% 23.99 96.59 406.07 29.82 78.98 319.67 60% 26.47 112.82 436.21 33.71 91.57 359.67 70% 29.25 131.29 466.55 38.30 105.95 402.81 80% 33.11 154.78 497.51 44.94 124.03 451.54 90% 40.62 197.11 529.54 57.16 152.54 504.66 95% 48.47 240.07 546.76 70.34 184.74 533.21