Advanced Methods for the Diagnosis of Volatile Organic Compounds

Abstract In recent years, volatile organic compounds (VOCs) have received increasing attention due to their impact on agri-food quality, human health, and industrial sustainability. Among these compounds, chlorine-based VOCs such as 2,4,6-trichloroanisole (TCA) and its precursor 2,4,6-trichlorophenol (TCP) represent a serious problem in the wine and cork industries, where extremely low concentrations (ng/L) are sufficient for sensory defects. The detection of such contaminantsincomplexnaturalmaterialsischallengingbecauseitrequiresanalyticalapproaches that combine high sensitivity with suitability for rapid, large-scale screening. Analytical techniques, such as Gas-Chromatography-Mass-Spectroscopy (GC-MS), are sensitive but are often costly, time-consuming, and poorly adapted for large-scale or in-line industrial applications. Moved by these motivations, this PhD thesis aims to introduce a methodological advancement in the field of spectroscopic VOC diagnosis, focusing on cork materials as an industrially relevant case study. The core objective of this work is the development of a strategy based on Fourier Transform Infrared (FTIR) spectroscopy and Artificial Intelligence (AI) techniques, designed to improve sensitivity to VOCs and enable reliable discrimination between contaminated and non-contaminated samples. Because of the weak spectral signatures of VOCs within the heterogeneous cork matrix, a full- spectrum approach was adopted. Attenuated Total Reflectance FTIR (ATR-FTIR) spectroscopy was combined with unsupervised and supervised Machine Learning (ML) models to capture subtle patterns across the infrared spectrum that cannot be identified using conventional band- by-band analysis. ML models represent a powerful and valuable tool for the recognition of spectral signatures and are widely used across multiple scientific and industrial fields. However, their performance critically depends on the availability of sufficiently large and representative experimental datasets for training, which can be a limitation, particularly when data collection is limited by experimental costs and time constraints. To address this challenge, this thesis develops a software framework that generates realistic artificial spectroscopic data from experimental measurements. These artificial datasets expand the available data and enable the systematic application and comparison of ML models for distinguishing contaminated from non-contaminated samples. Moreover, the framework allows controlledanalysisofhowfactorssuchasdatasetsize, contaminationlevel, noiselevel, background variation, and spectral region influence classification performance. The proposed methodology is validated through the detection of TCA contamination in cork samples, demonstrating that ML-assisted FTIR analysis in some models, such as Support Vector Machine (SVM), can effectively discriminate contaminated from non-contaminated materials even though in low concentrations (0.1%). The results presented in this thesis highlight the feasibility of combining FTIR spectroscopy and artificial intelligence as a rapid, non-destructive, and cost-effective approach for VOC detection, with strong potential for industrial quality control and broader agri-food applications.