Deep Learning Approaches to Chemical Property Prediction from Brewing Recipes

Despite the explosion of craft beer brewing over the last decade, there is virtually no work in the public domain exploring machine learning approaches to understand and optimize the brewing process. Learning to map between representations of an object across different domains is one of the fundamental challenges in machine learning. There are at least three distinct representations of beer that one may wish to learn to map between: 1) the brewing recipe, 2) the chemical composition of the resulting beer and 3) the written reviews of the beer. The mapping between any pair of these three domains is highly non-linear. Brewing beer involves complicated biological and chemical processes, while the human qualitative perception of a beer may be even more complex. In the work described in this paper, we focus on the former: mapping between the recipe and chemical attribute domains. We use two deep learning architectures to model the non-linear relationship between beer in these two domains, classifying coarse- and fine-grained beer type and predicting ranges for original gravity, final gravity, alcohol by volume, international bitterness units and color. Such models could be used to optimize recipes to produce desired chemical properties of the beer, allowing brewers to design better tasting beer, faster and with less waste. Using a set of approximately 223K brewing recipes from homebrewing site http://brewtoad.com, we find that deep and recurrent neural network models significantly outperform several baselines in predicting these attributes, offering relative reductions in classification error by 20%+ and reducing the root mean squared error for the attribute ranges by 44% relative to the best baseline.