Evaluation and application of chemometric and spectroscopic techniques for application-based analysis
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Chemometric analysis has been demonstrated as a powerful, and robust tool for the elucidation of additional information from chemical measurements. The specific preprocessing methods and model used to draw new conclusions have been evaluated across multiple measurement techniques and data sets. In addition, a novel classification visualization method is proposed and demonstrated for the creation of network guided decision tress for classification of large interconnected data set. ☐ Specifically shown in this dissertation, first edible oil peroxide values was predicted from Near-infrared spectroscopy measurements with a minimum prediction error of only 1.80. While no best combination of preprocessing method and prediction model was found, it was shown that in data sets with a few, broadband features, models such as LASSO regression, which reduce regression coefficients fully to zero produce on average better results, it was also shown that in these same limited broadband data sets, box-car averaging can benefit prediction be having a greater noise reduction effect than signal loss. ☐ Next, a novel method for visualizing classification success through network analysis was presented. This method was shown to easily visualize large, and interconnected classification results, and to provide easy identification for stepwise classification models. This process was demonstrated using a trace elemental analysis of exotic hardwoods, resulting in classification Cohen’s kappa accuracy of prediction of 0.93 in the best performing models. More interestingly, the network guided stepwise classification was able to produce classification models with as few as half of the misclassifications as the single pass flat classifier. Lastly, variable importance was used to demonstrate that as few as five factors, or elements, are needed to classify these exotic hardwood samples with Cohen’s kappa accuracies for prediction greater than 0.9. ☐ A combination of Gaussian mixture modeling with expectation maximization was shown successful at identifying nanoparticle sub-populations via AFM. The AFM imaging showed the degradation of virus-like particle (VLP) nanoparticles, and the degradation products. This process shows the possibility of discrete degradation steps rather than degradation continuum for these nanoparticles as the degradation products for distinct populations. Furthermore, the compressibility of the VLP nanoparticles was shown to be an indicator of interior particle degradation before the morphological aspects of this process are exhibited. ☐ Lastly, the performance of a liquid core wave guide (LCWG) enhancement system of Raman spectroscopy was evaluated. The LCWG system showed signal enhancement of up to 100x, but was unable to produce Raman Optical Activity, likely due to inadequate circular polarization of the incident beam. However, due to the significant signal enhancement exhibited by the LCWG system, the instrument was altered to detect low concentration amino acids and VLPs. This analysis showed 13.3x signal enhancement of these biomolecules demonstrating potential as an analysis technique for the study of low concentration VLPs.
Description
Keywords
Chemometric analysis, Chemometrics, Classification, Raman spectroscopy, Regression, Spectroscopy, Virus-like particle