Abstract

Beer production consists of a series of complex chemical, physical, and biological transformations. Although modern industrial production protocols are highly standardized, external and process disturbances often lead to degradation in beer quality. While models are available to predict beer quality, their widespread use is currently limited. Typically, these models rely on input variables from products that are almost complete; therefore, corrective measures cannot be implemented on time. Advanced modeling tools, such as digital twins, are effective alternatives to tackle this limitation, as they can integrate real-time process data into a digital model of the physical system to provide online predictions. To advance the development of such a tool, we have developed a hybrid model for beer quality prediction by combining different modeling frameworks and industrial process data. First, a dynamic model of industrial beer fermentation was calibrated that satisfactorily captured the kinetics of primary (extract) and quality-associated volatile compounds (pentanedione and butanedione). Second, a Naïve Bayes classifier for predicting beer quality was trained using physicochemical variables that identified the most critical attributes in a high-quality beer (ethyl acetate, total esters, foam stability, bitterness, and isoamyl acetate). Lastly, a hybrid regression model was constructed using fermentation model descriptors and external process data to predict the latter attributes with high prediction fidelity (less than 10% relative root mean squared error). The model identified yeast handling—including storage and propagation—and wort preparation as critical determinants of final product quality. Overall, this work represents a step toward developing a digital twin that can provide real-time process descriptors and integrate industrial data to optimize production and enhance beer quality.