Presentation Type

Poster

Abstract

The number of craft breweries has exploded in last decade: there are around a dozen breweries in Bellingham alone. Each brewery must assemble a lineup of beers, but this process of designing new beers usually relies on some combination of brewer instinct and trial and error. Because brewing beer involves complicated biological and chemical processes, the mapping from recipe to the beer it will produce is non-trivial to predict. In this project, we consider mapping between representations of beer in three distinct domains. In one view, a beer can be described by a recipe, which specifies the particular hop, malt and yeast varieties required, their quantities and the times at which they are added during the brewing process. In another view, a beer can be described by its objective chemical attributes, including color, alcohol by volume, international bitterness units, quantity of flavor and aroma compounds, and so forth. Finally, a beer can be viewed through the way it is perceived subjectively; e.g., in the words used to describe it in written reviews. In this project, we focus on mapping between the recipe and chemical attribute domains. We use sophisticated deep learning techniques to model the highly non-linear relationship between beer in these two domains. Our goal is to be able to optimize the recipe given some constraints on the objective properties of the beer, which would allow brewers to design better tasting beer more efficiently. Towards this goal, we developed a model that takes a recipe as input and predicts its attributes. Our preliminary results on roughly 200,000 brewing recipes scraped from brewtoad.com are encouraging.

Start Date

6-5-2017 12:15 PM

End Date

6-5-2017 2:00 PM

Location

Miller Hall

Share

COinS
 
May 6th, 12:15 PM May 6th, 2:00 PM

RED BeeTL: Recipe Encoder Decoder Beer Translator LSTM

Miller Hall

The number of craft breweries has exploded in last decade: there are around a dozen breweries in Bellingham alone. Each brewery must assemble a lineup of beers, but this process of designing new beers usually relies on some combination of brewer instinct and trial and error. Because brewing beer involves complicated biological and chemical processes, the mapping from recipe to the beer it will produce is non-trivial to predict. In this project, we consider mapping between representations of beer in three distinct domains. In one view, a beer can be described by a recipe, which specifies the particular hop, malt and yeast varieties required, their quantities and the times at which they are added during the brewing process. In another view, a beer can be described by its objective chemical attributes, including color, alcohol by volume, international bitterness units, quantity of flavor and aroma compounds, and so forth. Finally, a beer can be viewed through the way it is perceived subjectively; e.g., in the words used to describe it in written reviews. In this project, we focus on mapping between the recipe and chemical attribute domains. We use sophisticated deep learning techniques to model the highly non-linear relationship between beer in these two domains. Our goal is to be able to optimize the recipe given some constraints on the objective properties of the beer, which would allow brewers to design better tasting beer more efficiently. Towards this goal, we developed a model that takes a recipe as input and predicts its attributes. Our preliminary results on roughly 200,000 brewing recipes scraped from brewtoad.com are encouraging.