Assessment of the Sensory Properties of Wines Using Artificial Intelligence Technologies

Machine learning (ML) algorithms can effectively analyze complex chemical-sensory relationships in wine. Five different ML methods, characterized by different functionalities, were used to predict the sensory evaluation of 150 wine samples based on the content of amino acids, volatile compounds, total content of higher alcohols, acetic acid, and furfural. The concentration of analytes in the wine was determined by gas chromatography and capillary electrophoresis. Initial values of the sensory properties were evaluated by qualified wine tasters. Based on the results of model training, automated neural networks proved to be the most effective method. Boosted trees were also used for predicting sensory evaluations in test data. The concentrations of proline, threonine, higher alcohols, and methanol made the greatest contribution to the prognostic model. An increase in the concentration of proline and furfural increased the contribution to sensory evaluation, while the concentration of threonine and arginine, as well as higher alcohols such as acetaldehyde, acetic acid, and ethyl acetate, decreased it. It has been shown that despite being oriented toward solving large-dimensional problems, ML methods can be successfully applied to smaller data sets, in particular to predicting sensory evaluation of wine based on its component composition. This ability to process data with any dimension and model complex nonlinear relationships makes them promising for use in artificial intelligence systems for assessing food quality.