Key Compounds Mining Related to Aroma Richness of Cabernet Sauvignon Dry Red Wine Integrated with Machine Learning

To establish an objective evaluation method for the aroma richness of Cabernet Sauvignon dry red wine integrated with machine learning and identify the key aroma compounds influencing wine aroma richness, volatile compounds in 14 Cabernet Sauvignon dry red wine samples from different production regions and vintages were subjected to qualitative/quantitative analysis and odor activity value (OAV) analysis by gas chromatography-mass spectrometry, basic physicochemical indicators determination, and sensory fuzzy comprehensive evaluation method based on fuzzy mathematics. Contents of titratable acid, residual sugar, volatile acid, alcohol content, free SO₂, total phenols, tannins and anthocyanins, and the pH value were also determined. Orthogonal partial least squares discriminant analysis (OPLS-DA) was used to mine key differential compounds affecting aroma richness, and machine learning algorithms were finally applied to evaluate wine aroma richness. The results showed that 14 wine samples could be classified into four aroma richness grades (high, medium, low, and extremely low) using the sensory fuzzy comprehensive evaluation method. Further comparative analysis based on these four grades revealed that wine aroma richness was significantly positively correlated with titratable acid and volatile acid contents, as well as volatile fatty acid compound contents. OAV analysis indicated that the OAVs of ethyl acetate, isoamyl acetate, phenethyl acetate, and α-damascenone were significantly positively correlated with aroma richness. OPLS-DA results demonstrated that isobutanol, ethyl decanoate, decanal, phenethyl acetate, and α-damascenone (VIP > 1, P < 0.01) were the key differential aroma compounds affecting aroma richness, with vintage and production region significantly influencing damascenone content. Among three machine learning algorithms, backpropagation (BP) neural network, random forest (RF), and support vector machine (SVM), the RF algorithm exhibited high stability, and the important feature factors screened for aroma richness were mostly esters and volatile fatty acids. This study was expected to provide theoretical guidance for objective aroma evaluation and quality control in the actual production of Cabernet Sauvignon dry red wine.