Predicting oxygen consumption rate by tannins through sweep linear voltammetry and machine learning models

Abstract

Nowadays, it is well known that oxygen significantly impacts wine quality. The amount of oxygen wine consumes during the winemaking process depends on several factors, such as storage conditions, the number of rackings, the materials used for aging, and the type of closure chosen for bottling. The primary substrate for oxidation is polyphenols, whose oxidation leads to the formation of quinones. These quinones, in turn, can react with numerous compounds in wine, causing changes of varying depth depending on its chemical composition (Ugliano, 2013). For this reason, tools capable of assessing the oxidative potential of wine are essential for improved winemaking management.

To address this, linear voltammetry was employed to determine the potential oxygen consumption of wine in which tannins were added. The experiment provided 43 enological tannins from three different botanical sources: oak, grapeseed, and gallnut. 500 mg/L of each tannin was added to a Corvina red wine without sulfur dioxide and with 25 mg/L of free sulfur dioxide, respectively, for a total of 86 samples. After tannins addition, the sample was saturated by handshaking up to 7,5 mg/L of oxygen. Samples were kept in an oven at 25 °C. Oxygen consumption was monitored every 24 hours, and the experiment was concluded when the samples contained less than 1 mg/L of oxygen.

The oxygen consumption data were subjected to clustering analysis, which divided the samples into two groups based on their hourly oxygen consumption rates, which have been called “Fast” and “Slow”. A Kruskal-Wallis test was performed to verify that the oxygen consumption rate between the two groups was statistically different. The voltammograms of the samples were used to train machine learning models to predict whether a sample would consume oxygen at a fast or slow rate based on these initial measurements. Voltammetry is a quick, simple, and cost-effective analytical technique. However, its primary drawback lies in the interpretability of voltammograms. In studies like Ugliano et al., 2020, a convolutional analytical approach was applied to accentuate differences in voltammograms that might hold crucial interpretive information.

A single dataset was created, combining the linear voltammetry data with their first and second derivatives of themselves. A Random Forest was employed to avoid feature redundancy during training, followed by Recursive Feature Elimination (RFE) for precise feature selection. The resulting dataset was used to train a Support Vector Machine (SVM).

The SVM model achieved an overall accuracy of 94% on the test set, with mean precision and recall of 96% and 92%, respectively. The AUC-ROC score was 1.00, indicating perfect discrimination between the “Fast” and “Low” oxygen consumption classes. These results demonstrate the effectiveness of the feature selection and hyperparameter optimization approach in this context. The study highlights the potential of voltammetry, combined with advanced machine learning techniques, to provide valuable insights into the oxidative potential of wine.

References

Ugliano, M. (2013). Oxygen contribution to wine aroma evolution during bottle aging. Journal of agricultural and food chemistry, 61(26), 6125-6136.

Picariello, L., Slaghenaufi, D., & Ugliano, M. (2020). Fermentative and post‐fermentative oxygenation of Corvina red wine: influence on phenolic and volatile composition, colour and wine oxidative response. Journal of the Science of Food and Agriculture, 100(6), 2522-2533.