Wine fermentation relies on the conversion of sugar and nitrogen into ethanol, carbon dioxide, and aroma compounds, the latter contributing to the sensory profile of the wine. Although this process is conducted industrially in batch fermenters, it remains poorly optimized. Key control parameters, such as temperature profiles and nitrogen supplementation, could be used to adapt the aromatic profile to consumer preferences. In this context, mechanistic models of aroma synthesis are highly relevant: they enable deeper understanding and prediction of fermentation dynamics, identification of optimal conditions, and design of real-time control strategies. However, existing models are scarce and overly simplistic, often relying on empirical or constant yield coefficients to describe sugar-to-aroma conversion, and thus fail to capture the impact of nitrogen addition or temperature variations on aroma synthesis. This study presents a parsimonious, fully mechanistic model of wine fermentation integrating the dynamics of sugar, ethanol, nitrogen, biomass and isoamyl acetate—a key aroma compound highly sensitive to nitrogen addition. The model represents two key nitrogen-dependent processes — sugar transport and enzyme-mediated aroma synthesis — whose dynamics successfully reproduce the experimentally observed shift in isoamyl acetate yield. The model is calibrated using a robust sequential calibration strategy designed to exploit its hierarchical structure. We use a comprehensive dataset of 81 fermentations conducted at varying initial sugar and nitrogen concentrations, with different temperature profiles and nitrogen supplementation regimes, and for which high-frequency measurements are available. Finally, model performance is assessed using error metrics and is compared with that of an existing model from the literature.
Ensuring both high productivity and viable cell survival remains a major challenge in industrial-scale probiotic manufacturing due to the multistage and highly variable nature of fermentation and downstream processing. This study presents an exploratory, machine learning-based framework for analyzing batch-to-batch quality variability in the production of an L. rhamnosus-based probiotic, using data from 178 independent production batches together with process variables measured across the entire manufacturing pipeline, from seed culture through centrifugation, freeze-drying, and sieving. Three quality attributes — total production, viable cell count (VCC) reduction relative to theoretical values, and VCC change after one week of storage — were each dichotomized into below- and above-mean groups, framing batch-to-batch quality variation as a binary classification problem. XGBoost models were implemented to predict each quality attribute; the predictions of total production and VCC reduction achieved higher classification performance than the prediction of VCC change after one week. Feature importance analysis identified cryoprotectant volume, centrifugation-related variables (e.g., number of discharges), and pressure-related variables as prominent predictors, although their stability varied across quality attributes and sensitivity analyses. To verify the robustness of the results, a range of sensitivity analyses was additionally conducted, including comparisons across missing-data handling thresholds and classification cutoff definitions (mean vs. median). These analyses showed that the prediction of total production remained comparatively stable, whereas the predictions of VCC reduction and VCC change were somewhat sensitive to these methodological choices. These findings offer exploratory, process-level insights into predictive quality assessment based on routinely collected process data, while also indicating that their interpretation warrants considerable caution.
