Study of near infrared chemometric models with low heterogeneity films: The role of optical sampling and spectral preprocessing on partial least squares errors

Abstract

This study is focused on understanding absorption and scattering effects and their impact on the errors observed in near infrared calibration models developed using partial least squares regression able to predict the number of polypropylene film layers stacked together. The films provided a system with reduced heterogeneity to study the sources of error due to light scattering, selection of spectral preprocessing and spectral region on near infrared calibrations. Near infrared spectra were acquired using two experimental setups with the integrating sphere module of the FT-NIR spectrometer. The first setup consisted in stacking the polymer films to determine the penetration of the near infrared radiation. The second setup was a variation using a reflective surface on the top of the films in transflection mode, increasing the pathlength and therefore transmission and absorption of the material. The estimation of the penetration of radiation was performed using talc placed over the polymers films. The narrow bands of talc, related to first and second overtones of the O-H stretching mode, were used to estimate the near infrared sampling depth into polymer film layers, which ranged from 2.95 to 3.12 mm. Partial least square calibrations developed with up to 30 film layers were accurate with bias values that were not significantly different from zero at the 95% confidence level. Statistical errors were calculated for seven near infrared regions using different spectral preprocessing and the confidence interval of the bias showed that optical sampling is unbiased and there is an absence of systematic error by the near infrared method. A calibration model developed with 50 film layers presented high statistical errors and bias different from zero, indicating a sampling error due the depth of penetration of near infrared radiation. The impact of number of samples in the calibration and validation set was also evaluated. The results showed that bias was significant when the number of samples was less than 11. This finding highlights the lack of systematic error in the near infrared method, as long as the number of samples in the calibration set is representative of the variation to be modelled by a partial least squares regression and the sampling error is reduced.