University of Adelaide students teach AI to brew beer
People have been brewing beer for more than 5,000 years, so from the very beginning the process has not only changed but also become more complicated. As part of a machine learning internship, two computer science students from the University of Adelaide, Christopher Fusco and Josh Vira, created a neural network that was able to learn to brew beer by studying a huge number of brewing records.
The result is a unique IPA (India pale ale) created with artificial intelligence, which will go on sale in early 2022. Rodney AI²PA is named after Rodney Brooks, an Australian robotics pioneer and co-founder of iRobot (the company that made the Roomba robot vacuum).
Working under the guidance of AIML (Artificial Intelligence Markup Language) machine learning researchers and experts from the Barossa Valley Brewing brewery, the two students began building a large dataset of more than 260,000 existing craft beer recipes available online, before creating a neural network that learns from the data to make beer.
Christopher said: "We created 200,000 new recipes, and then trained the neural network to select the best of them and evaluate them." Beer typically contains only four main ingredients: malt, hops, water and yeast. Minor changes in these ingredients, as well as precise changes in timing and temperature at certain stages of the brewing process, lead to the huge variety of beer available today.
But if creating an AI capable of producing its own beer is not a difficult enough task, then creating an AI that can actually make good beer is far more challenging.
“In fact, it was very difficult. We had to come up with our own mathematical formula, using statistics from those original recipes”
Josh noted: "By obtaining statistics on these variables, we were able to assess the significance of each variable. This also helped us deal with any possible errors that might have arisen in the data."
Data such as the number of online views of certain beer recipes and the number of people who said they had brewed them make it possible to judge a beer's popularity. The students then created a neural network that learned to evaluate its own AI-generated recipes and give each one a popularity rating.
The AIML neural network created recipes with about 60 data points, or "features," which included not only the necessary ingredients and their quantities, but also specific process information, such as how to handle hops, the fermentation temperature of the yeast and the boiling time, as well as predictive indicators of bitterness, color and alcohol content.
“I think this is a very exciting time for machine learning, because we are approaching a stage where we can collaborate with other industries, using what we know and what we are good at. There is significant growth in this field”
The project had its technical challenges: the neural network sometimes produced strange outliers in its features, such as absurdly large or small quantities or very high temperatures.
“We used some statistical and graphical methods to analyze our data, and then decided to limit some of these outliers. This method helped us preserve the inherent variability of the data, as well as assess whether the decision we made accurately reflected the final product we wanted to create”
The result was 30 potential beer candidates selected by the AI. The system left the final brewing decision to the experts at Barossa Valley Brewing.
Brewery founder Denham D'Silva was happy for the opportunity for artificial intelligence to expand his company's creative process, but initially believed the technology was too new to be useful.
“Willingness to experiment and create interesting, premium beer styles has been the foundation of the brewery for 16 years. In short, it was terrifying to hand this process over to AI. Brewing is traditionally a very practical process, especially for a small craft brewery like Barossa Valley. When you run a small craft brewery, you cannot compete on scale, so you have to be different and smart”
Demand for craft beer in Australia is high. While per capita beer consumption has generally declined in recent decades, craft beer sales are growing at 10% per year.
“Most people think of artificial intelligence and machine learning as something only huge tech companies can do. This project showed us that AI can use our craft skills and expand them so we can compete. It is very important for small and medium-sized businesses to take advantage of the opportunities AI offers. Craft producers need to get involved to ensure that AI includes the artistry that makes us special”
Christopher and Josh plan to pursue careers in machine learning. Both students also intend to continue their studies in graduate school after completing their bachelor's programs at the end of 2022.
To complement the high-tech beer experience, the engineers also created a robotic "bartender" that can automatically detect when an empty glass is placed on the bar and quickly fill it with cold beer straight from the keg.
As for the important question of what AI beer actually tastes like, Denham D'Silva is enthusiastic: "It tastes like the future! Seriously, it's a fruity IPA that I'm very proud of and look forward to releasing."
Rodney AI²PA from Barossa Valley Brewing will go on limited retail sale from mid-January 2022.
The Australian students are not the first to use AI to brew beer. In 2016, IntelligentX began producing a line of AI-crafted beer. The company constantly collects feedback from its customers and updates its beer styles based on that feedback.
