Spectrofluorometric fingerprinting to track the evolution of red wine and model the relationship with compositional aspects

Abstract

Red wine contains an abundance of phenolic constituents that crucially contribute to sensory appeal. These compounds undergo dynamic reactions that evolve the wine’s phenolic profile, finally shaping its quality and sensory attributes. Given the importance, this work aimed to apply a rapid spectral method and machine learning to investigate red wine compositional modifications over time. Various red wines were studied under different evolution conditions using spectrofluorometric fingerprinting. Specifically, Mataro and Grenache wines treated after fermentation with seven common additives at three levels were analysed immediately after addition and after a period of storage. In addition, four Grenache, two Shiraz, and two Cabernet Sauvignon wines produced commercially were analysed at three successive time points during maturation. Pinot Noir and Petite Sirah wines with five cap management treatments were tracked during natural and accelerated bottle ageing. Phenolic and chromatic parameters were measured periodically for the wines under maturation or bottle ageing, along with determining tannin composition and molecular size for bottle-aged wines. Machine learning was used for spectral data analysis to classify different additives in Mataro and Grenache and to explore the traceability of additive treatments based on their spectrofluorometric fingerprint. Moreover, the relationship between spectral data and phenolic and chromatic compositions was established via extreme gradient boosting regression (XGBR), enabling the simultaneous prediction of these parameters. Spectral fingerprints of the bottle-aged wines and their isolated tannins were also compared to provide further insights into the compositional changes of red wines during ageing. XGBR models were validated to predict the molecular size and composition (according to subunit composition) of tannins based on fluorescence data of wines. The Train models reached near-perfect accuracy (R2 = 0.9999), with Cross-validation and Test models maintaining robust performance based on R2 values of 0.9271–0.9891. This study illustrates the ability of fluorescence spectroscopy to rapidly detect differences and changes in red wine phenolic profile induced by additives, short-term storage, maturation, and bottle ageing. It offers a powerful and simple spectroscopic approach for efficient monitoring and classifying of red wines, and enables the prediction of multiple phenolic-related parameters in a single method.