Robert Coleman* and Roger Boulton
*Treasury Wine Estates, Beringer Winery, Saint Helena, CA 94574 (bob.coleman@tweglobal.com)

Dynamic models can provide insights into yeast cell performance, nutritional limitation, and the influence of temperature on the rate and completion of wine fermentations. Such models have been used successfully to fit hundreds of complete fermentation density-time data. These models have also been applied from the onset of fermentation to provide a diagnostic and predictive tool for winemakers to intervene in the case of developing sluggish or incomplete fermentations. The models can also have an application in the estimation of cooling loads and the management of refrigeration loads and electrical energy usage during harvest. The Boulton Fermentation Model was used to evaluate the predictive performance of alternative parameter estimation routines using density-time information at early, middle, and late stages of fermentation (20, 12, and 8 Brix) in five white and five red commercial fermentations. This comprehensive set included fermentations of Chardonnay, Sauvignon blanc (2), Riesling, Muscat, Pinot noir, Syrah, and Cabernet Sauvignon (3) and covered fermentation temperatures between 12.8 and 30.6°C (55 and 87°F). The parameter estima- tion routines (Bard’s, Differential Evolution, Genetic Algorithm, and Particle Swarm Optimization) were compared for their ability to reduce the sum of squares deviation and to converge quickly. Prediction improvement was achieved using correlations with initial fermentation temperature to constrain some parameter boundaries during the fitting process. The most efficient parameter estimation routine for reducing the sum of squares error on this wine fermentation test set was Differential Evolution, with acceptable performance by all parameter estimation routines. Bard’s routine was most efficient in terms of time; however, the resulting parameter fit was more sensitive to the initial parameter guesses.

Funding Support: Treasury Wine Estates