Method for Using Historical Winery Data to Improve Wine Quality
We are developing an optimization method for wine processing based on historical winery data and artificial neural networks (ANNs). This method will allow winemakers to fully use the information collected in their wineries over the years in order to produce a final product which has the qualities that they desire, even as grape characteristics vary from year to year. To date, we have established that ANNs can be used successfully to correlate wine processing inputs with chemical and sensory properties of the finished wine. In particular, we have been able to predict primary and malolactic fermentation kinetics based solely on grape characteristics and intended processing. This result is the basis of using the methodology developed in the Long-Term AVF Grant on Sluggish and Stuck Fermentations. We have also used this data to evaluate how well the neural network can interpolate and extrapolate with data not used in training. In addition, we have been able to predict how the color of Cabernet Sauvignon wines that we have made is a function of the fermentation temperature, skin contact time, and macerating enzyme addition, as well as extending our knowledge of the effects of these treatments on the final phenolic profile of the wine. Using neural networks trained with the data from the wines produced here, we have evaluated several types of optimization methods for choosing optimal processing inputs to achieve the desired outputs. These include gradient methods, simulated annealing, and genetic algorithms. Of these, genetic algorithms look the most promising for wine processing cases. The optimization methodology developed to date has been used for several optimization case studies, including Sauvignon blanc processing, Cabernet Sauvignon processing, blending of Chardonnay wines, and the effect of field temperature profiles on Cabernet Sauvignon quality.