Creating an artificial wine taster: Inferring the influence of must and yeast from the aroma profile of wines using artificial intelligence

Abstract

The human brain is able to compute information from very complex olfactorical impressions. The special pattern of the concentrations of hundreds of aroma constituents allows an experienced wine taster to determine special features of the wine, for instance grape variety or vintage.
Artificial Neural Networks are often used to recognize shapes and patterns like faces or finger prints. Here we use Artificial Neural Networks to mimic the abilities of a wine taster to deal with very complex olfactorical patterns. We produced 120 unique wines combining twelve different grape musts and ten yeast strains and determined the aroma profile (83 aroma constituents) of all wines. We analyzed the ability of a well trained neural network to recognize the used must variety and the fermenting yeast strain from unknown aroma profiles. Furthermore we investigated the capability to predict the aroma profile of a wine with a must variety/yeast strain combination that is new to the neural network.
In 96 % of all trials the neural network identified the must that was used for wine production correctly (expected random propability: 8 %). An accurate identification of the yeast strain, used for fermentation, occurred in 67 % of all trials (propability by chance: 10 %).
The aroma profiles of the must/yeast combinations new to the neural network were forecasted with a divergence of only 2.1 % compared to the actual wine of this production characterization. Thus we conclude that a comprehensive description of wines using neural networks is possible.