Use of mathematical modelling and multivariate statistical process control during alcoholic fermentation of red wine

Abstract

Cyberphysical systems can be seen in the wine industry in the form of precision oenology. Currently, limitations exist with established infrared chemometric models and first principle mathematical models in that they require a high degree of sample preparation, making it inappropriate for use in-line, or that few oenological parameters are considered. To our knowledge, a system which incorporates a more comprehensive mathematical model as well as in-line spectroscopic monitoring for the purpose of precision oenology has not yet been presented.

The use of first principle mathematical modelling was employed to predict the trends of alcoholic fermentation and oenological parameters in a four-phase model based on initial conditions. The components of interest were sugars, alcohol, biomass, nitrogen, carbon dioxide, phenolic parameters, and pH. The phases considered included the lees, the cap, the must, and an intermediate liquid phase present in the cap. For each phase, a system of ordinary differential equations was developed to describe the change of each of the components listed. Parameters such as mass transfer coefficients and partition coefficients need to be determined via regression during the model development stage. To obtain the necessary data, fermentations using three different cultivars (Shiraz, Merlot, and Cabernet Sauvignon) were conducted using three different temperatures (20oC, 25oC, and 28oC). Samples were taken once per day and chemical analysis took place for each of the components. A functional mathematical model capable of generating accurate forecasts for different oenological components using the chemical composition of grapes was attempted. Additionally, the model should describe the change in parameters due to cap mixing and increasing ethanol concentration. The model includes the boundary conditions which can be used to determine if a fermentation is deviating from desired progression.

To complete this process control system, it is still necessary to utilize partial least squares (PLS) calibration models for real time monitoring. Additionally, outlier identification, caused by abnormal spectra, was performed using statistical analysis allowing samples to be re-analysed. The use of machine learning techniques and the development of local and incremental models was explored to assess a live updating of the PLS models. The expected outcome of this study is a combined system using dynamic modelling to predict the fermentation and extraction trends and the monitoring with real time predictions generated by PLS models