Near Infrared Spectroscopy and Chemometrics for in situ Process Monitoring

This thesis investigates the application of near-infrared (NIR) spectroscopy as a process analytical technology (PAT) tool for in situ monitoring of physico-chemical processes, at ambient or industrially relevant conditions such as high temperature, high pressure, or in corrosive environments. The work originates from a collaboration with an industrial partner aiming at the development of a NIR-based method for monitoring urea synthesis in simulated process conditions. Starting from this, the thesis was extended to fundamental studies on simpler molecular systems, allowing systematic explorations of the application of NIR spectroscopy across different matrices. Classic chemometric methods such as principal component analysis (PCA) and partial least squares (PLS) regression were applied as primary tools for qualitative and quantitative analysis respectively. For quantitative tasks, machine learning algorithms were also evaluated, and their performances were systematically compared with PLS regression in different case studies. NIR thermometry was investigated using two probe molecules with different intermolecular interactions, i.e. CO2 (weak quadrupole-quadrupole interactions) and water (extensive H-bonding framework). Continuous temperature ramp protocols were studied as alternative to discrete measurement approaches widely used in literature. This approach yielded more representative calibration datasets with PLS models achieving prediction accuracies below 1°C for both molecules. The inverse problem, i.e. the prediction of NIR spectra from process parameters was then investigated by ML techniques, using different validation frameworks. This revealed that models excelling according to cross-validation can fail on truly independent data. The choice of preprocessing was shown to be as relevant as the choice of regression algorithm. A comprehensive framework for their systematic evaluation was developed. The quantification of infrared-transparent species (e.g., inorganic salts dissolved in water) exploits the aquaphotomics principle, where perturbations of the H-bonding network by dissolved ions are reflected by quantifiable effects on the absorption bands of water. The extension from aqueous solutions of single salts at ambient temperature to multi-component systems at elevated temperatures demonstrated the potential of this approach also for industrially relevant mixtures whose analytes are not directly detectable. Finally, NIR spectroscopy was applied to real-time monitoring of the urea synthesis at high temperatures and high pressures. PLS models developed for both reactants and products achieved accuracies suitable for industrial process control. Across all case studies, it was evidenced that no single algorithm universally outperforms the others. Rather, the optimal combination of preprocessing and regression approach is case-dependent, making a comprehensive and systematic evaluation of all the available combinations the most reliable a