AI beer and the future of flavor

St. Austell Brewery began in 1851 in Cornwall, England when Walter Hicks mortgaged his farm to start a small wine business. Twelve years later, he mastered brewing, and by 1893, he established the brewhouse where St. Austell still produces its renowned ales, lagers, and IPAs. This family-owned brewery in southwest England prides itself on tradition. Yet, sometimes, bucking tradition can lead to remarkable results. 

Every year, St. Austell brewers participate in a Cask Club, developing unique recipes for around 40 casks of limited-edition beers sold to select pubs. Barnaby Skerrett, the assistant brewing manager, works on brewhouse automation using artificial intelligence. That’s when lightning struck.

“I told it to write me a recipe based on broad parameters of colors and flavors, and it sent me some ideas. So, I turned them into a recipe,” Skerrett said this March. The resulting beer included Willamette, Cascade, and Sultana hops, lending a tropical, juicy flavor to what Skerrett called Hand Brewed by Robots. It’s a 4.2% “AI-PA,” a play on the traditional IPA label. This spring, pint glasses filled with the AI-designed beer had no problem finding the hands of human drinkers.

Cracking the flavor code

AI tools, often linked with disruption in cost-sensitive fields like content creation, are now transforming the food industry. Despite its apparent simplicity, food is complex. Your favorite type of coffee, for instance, contains hundreds of compounds indiscernible to humans that provide taste, texture, and smell. 

But artificial intelligence is rewriting the rules. The new tools of machine learning can parse the minutiae of the chemical makeup of foods and drinks. In turn, AI is allowing brewers, chefs, and food scientists to more efficiently analyze data about the convoluted world of flavor—sparking rapid innovations in how we create food and drinks.

“We’ve described the culinary world as ‘culinary arts’ for so long. AI is allowing us to take that art and make it into science,” says Jay Gilbert, Director of Scientific Programs and Career Pathways at the Institute of Food Technologists. “It’s giving us a true foundational understanding of the molecules that are in food.” 

Engineering taste from the ground up

In some cases, that understanding starts at the very beginning. Christine Diepenbrock is a faculty member and molecular breeder at the Artificial Intelligence Institute for Next Generation Food Systems at UC Davis. Right now, she’s trying to build a better lima bean. The dry lima bean is one of California’s primary food exports; the state itself produces virtually all of the dry lima beans in the U.S. Diepenbrock and her colleagues are now conducting field trials throughout the state to improve the bean’s yield and pest tolerance. 

But the foundational work occurs in the lab. Each lima bean plant is composed of tens of thousands of protein-coding genes that determine how the plants grow and respond to pests. Identifying genes that confer resistance to threats like beetles and spider mites—or those that enhance crop yields—presents a daunting challenge for humans. Machine learning, however, is well-suited for this complex task.

“Humans can’t intuit this information from thousands of different data points,” Diepenbrock says. “But AI is really good at it.” 

Using custom models, Diepenbrock and her colleagues scan the vast DNA sequence of the lima bean plant to rapidly find the genes that code for specific, desired traits — traits that make lima beans that taste good to humans, but not to beetles.

Processing palates 

Despite these advancements, the subjective nature of taste means that human judgment remains critical. Why rely on a computer if the outcome could be food more suitable for the trash than the dinner plate?

“All human beings perceive things differently because we don’t have the same receptors,” says Julien Delarue, a sensory and consumer scientist at UC Davis. “But the link between chemistry and perception is not that obvious.” 

Two people might find a dish or drink bitter; one might enjoy it, the other might not. Delarue’s role involves analyzing data from consumer studies to establish how people’s perceptions relate to food composition, aiming to identify the compounds that contribute to flavor profiles.

According to Delarue, many foods contain hundreds or even thousands of active chemicals. Chocolate, for example, is not just processed cocoa beans. It’s an amalgamation of 1,500 flavor compounds. There are flavonoids, fatty acids, and even a compound called phenylethylamine, which stimulates the brain’s pleasure centers. 

Generalizing the link between those compounds and someone’s taste is where Delarue’s team employs machine learning models and partitioning algorithms. Using R, a programming language commonly used for data analysis, as well as Python, they run multivariate analyses that allow them to connect the language people use when they describe food with the compounds they find in foods themselves using chromatography and spectroscopy, two techniques used to identify chemical compounds.

“The techniques provide a wealth of data that human beings just cannot handle, especially if you want to relate the data,” Delarue says. “That’s how we help food scientists improve or create new products.”

The challenge of aligning these data with human preferences just reveals the complexity of taste: What delights one person may repel another. Yet, AI's growing capability to parse these vast data sets might increasingly enable the creation of universally appealing flavors.

An algorithmic ale

In the same month Skerrett was perfecting his robot-assisted beer, researchers at KU Leuven University in Belgium were unveiling the results of three years of research. They had successfully used artificial intelligence to enhance the taste of Belgian beers. Initially, they analyzed the chemical makeup of 250 different beers—ranging from lagers and fruit beers to ales, blonds, and even non-alcoholic varieties—focusing on aspects such as sugar and alcohol levels, and around 200 different compounds. A panel of 16 tasters then rated each beer on attributes like sweetness and acidity. 

Simultaneously, the team reviewed 180,000 beer ratings from RateBeer, initially reading thousands manually before using a Python-built model to analyze the rest. With multiple datasets in hand, they processed the chemical and sensory data using R, developing machine learning models to not only predict beer flavors but also suggest improvements.

“A lot of people were convinced that flavor was way too complex to model, but for the first time we now had a data set to actually try it out,” says Michiel Schreurs, a doctoral student at KU Leuven and one of the researchers. “By predicting appreciation from flavor attributes, we identified which attributes are important to make a good beer.”

The results were clear: Enhancing an existing Belgian beer with higher concentrations of lactic acid and glycerol led the tasters to rate it higher for both sweetness and body.

Flavor frontiers

While AI has the potential to transform food creation and design, its effectiveness is currently limited by the data available for training. "For AI, you need a fairly large set of data to have a highly accurate model," Diepenbrock says.

Looking ahead, with ample data, AI could revolutionize entire kitchens: Molecular gastronomy, which examines the physical and chemical transformations of food during cooking, is particularly amenable to AI integration. Big name chefs like Heston Blumenthal and Ferran Adrià are already experimenting with AI to devise new dishes.

But the transformation of food by AI doesn't signal the end of traditional roles. "An AI can never properly tell you whether that beer tastes fantastic or if it’s not up to scratch," Skerrett says. Robots might brew the beer, but ultimately, its taste will be judged by human palates.