The samples consisted of a white powder contained in small vials and consisted of varying
percentages of the following constituents:
- PMC-50U (Cellulose ether)
- PVA (polyvinylalcohol)
- Casucol (starch ether)
There were 10 total pre-prepared samples. 7 samples had actual values were and the remaining 3 samples were classified as unknowns. There were also 100 grams each of the pure form of the three constituents. Additional samples were prepared using the pure form of each constituent. A precision balance accurate to 0.1 mg was used to weigh the powder. Blending was done by vigorously shaking the mixture for approximately a minute.
All samples were scanned using the AOTF-NIR spectrometer with a full bundle transflectance probe. The probe was placed in each sample such that its weight was the only force to compact the sample. This minimized error in collecting data from the samples because the varying degree of compaction of particles plays a part in the scatter and reflection of light. The spectral range was 1200nm to 2300nm with each spectrum consisting of an average of 200 scans. The data were collected directly in the absorption mode. The spectral data were entered into The Unscramblerä and PLS 1 regression analysis was performed on each of the constituents.
- Sample weights
|Samples Prepared by Brimrose|
|SAMPLE||Wt. PMC||Wt. PVA||Wt. Casucol||% PMC||% PVA||% Casucol|
Table 1. Weights and percent values of all samples.
Figure 2. Absorbance spectra of pre-prepared samples and samples prepared by Brimrose.
3. Regressions and Modeling
Figure 3. PLS 1 regression model for % PMC.
Figure 4. PLS 1 regression model for % PVA
Figure 5. PLS 1 regression model for % Casucol.
The results for these regression models were excellent and showed good correlation between the calibration and validation sets. The regression for PMC had a SEC of 0.55 and an SEP of 0.93 with two outliers removed. The regression for PVS had a SEC of 0.35 and an SEP of 0.61 with one outlier removed. The regression for Casucol had a SEC of 0.66 and an SEP of 0.98 with one outlier removed.
3 data points were removed from the data set and models were created using the remaining
data points. The models were then used to predict the values for the points taken out and these values were compared to the known values. The results were excellent for such a small calibration set. Predictions were then done for the 3 unknown samples.
|CALCULATED PERCENT||PREDICTED PERCENT|
|SAMPLE||% PMC||%PVA||%CAS.||% PMC||%PVA||% CAS.|
Table 2. Prediction of % values using models created with these samples removed.
|SAMPLE||% PMC||% PVA||% CASUCOL|
Table 3. Prediction of % values for the 3 unknown samples.
- Conclusions and Recommendations
It is concluded that it is feasible to use the Brimrose AOTF-NIR spectrometer to determine percent values for PMC, PVA, and Casucol. Both the regression models and values for the predictions show good correlation between the known values and the values determined from the spectral data. It is noted that the PVA and Casucol had larger error values than the PMC. This is because the percent values for PVA and Casucol are much smaller than those for PMC. The small number of data points used for this model definitely contributed to the error. The results were excellent considering there were only 24 data points used in each model and past experience has shown that using 100 or more samples will certainly create a more robust model. It is recommended that a purchase order be placed for a Brimrose AOTF-NIR spectrometer to allow for testing of a larger amount of samples which can be used to create a more robust model that can predict percent values for PMC, PVA, and Casucol using spectral data.