Neural Brewmeister: Modelling Beer Fermentation Dynamics Using LSTM Networks

Fermentation is a complex biochemical process that transforms brewer’s wort into beer. Beer fermentation is driven by yeast and is influenced by process parameters such as the content of fermentable sugars in wort, temperature, and pH. Traditional methods of modelling this process rely heavily on empirically tuned kinetic models. However, these models tend to be recipe-specific and often require retuning when processes change. This paper proposes a data-driven approach using a Long Short-Term Memory (LSTM) network, a type of recurrent neural network, to model beer fermentation dynamics. By training the LSTM model on real-world fermentation data (1305 fermentations across ales, IPAs, lagers, and mixed-culture beers), including variables such as apparent extract (derived from specific gravity), temperature, and pH, we demonstrate that this technique can accurately predict key fermentation trajectories and support process monitoring and optimisation. When evaluated on representative medoid fermentations as one-step-ahead roll-outs over 0–300 h, the model produces accurate predictions with low errors and minimal residuals. These results show that the LSTM-based model provides accurate and robust predictions across beer styles and operating conditions, offering a practical alternative to traditional mechanistic kinetic models. This work highlights the potential of LSTM networks to enhance our understanding, monitoring, and control of fermentation processes, providing a scalable and efficient tool for both research and industrial applications. The findings suggest that LSTM models can be effectively adapted to model other fermentation processes in beverage production, opening new possibilities for advancing food science and engineering.