Title: Qualitative and Quantitative Analysis of Milk Protein Using Chemometrics Coupled With Infrared Spectroscopy

Program: Computing PhD

Committee: Dr. Timothy Andersen (Chair), Dr. Owen McDougal (Co-Chair), Dr. Edoardo Serra

Abstract: Traditional wet chemistry methods for milk protein analysis are accurate but time-consuming and unsuitable for real-time monitoring. Infrared (IR) spectroscopy offers a rapid, non-destructive alternative; however, the complex, overlapping signals in spectral data require chemometric approaches to deconvolute and extract meaningful information. This dissertation develops a chemometric framework that integrates IR spectroscopy with machine learning for rapid, accurate quantification of milk proteins.

The research is guided by Process Analytical Technology (PAT) principles; emphasizing real-time monitoring, chemometric analysis, and process control, but applied retrospectively to laboratory and industrial datasets. The framework achieves high predictive accuracy through systematic investigation of intelligent calibration sample selection, spectral preprocessing, wavenumber selection, and model optimization.

Chapter 2 demonstrates an optimized SVR chemometric model for MIR-based whey protein analysis (ß-lactoglobulin R2P = 0.928; a-lactalbumin R2P = 0.927). Chapter 3 introduces automated preprocessing algorithms using Bayesian optimization, achieving R2P values above 0.90 for all evaluated milk components, including true protein, fat, lactose, and total solids, using the MIR dataset from Agropur. Chapter 4 validates the framework using an industrial NIR dataset from Daisy Brand (casein R2P = 0.994; true protein R2P = 0.997). Although the primary focus of this dissertation is on proteins, the framework also demonstrates strong performance for other components, including lactose, total solids, and fat.

The primary contributions are: (1) a chemometric framework integrating IR spectroscopy with machine learning; (2) automated spectral preprocessing to eliminate subjective parameter selection; (3) dataset-specific automated wavenumber selection; and (4) validation using industrial datasets from Agropur and Daisy Brand.