Neural-Network-Assisted Optimization of Wine Blending Based on Sensory Analysis

Abstract

Because common sensory characteristics of wine are frequently the result of many different compounds with varying perception thresholds, a nonlinear relationship often exists between the desired target attributes of a final blend and the individual attributes of the base wines, thus complicating the blending process. To address this complication, a blending optimization method has been developed that uses artificial neural networks to model the potentially nonlinear response of the blending based on sensory data from the base wines and a limited number of blends. This method has been developed and verified by constructing a series of 24 wines from three base wines. Each wine was profiled by descriptive analysis with a trained panel, and the sensory data was modeled with an artificial neural network. After choosing specific target attributes for the final blend, an optimization algorithm was employed to predict the optimal blend for this set of goals. Optimal blends chosen with this methodology had sensory characteristics close to the goal characteristics and to experimental blends with similar composition. Reduction of the training data to a single experienced judge and elimination of 30% of the trial blends did not change the optimal blend identified significantly (less than 2% difference in any fraction). A reduction of 50% of the trial blends led to changes of up to 11%, demonstrating that caution must be exercised in reducing the data collected for blending.