With the support of artificial intelligence, researchers want to create new beers that delight consumers. To date, it has only been possible to a limited extent to find out through consumer tests which alcoholic and non-alcoholic flavors are well received by the market, explains the team from Belgium in the specialist journal “Nature Communications”. The new method could help food manufacturers meet specific consumer needs more efficiently and cost-effectively.
Researchers led by Kevin Verstrepen from the Catholic University of Leuven explain that it is generally complex to predict which new food flavors consumers will like. This is mainly due to the fact that there are an immense number of taste-active chemicals in food. There are also interactions and complex enhancing or diminishing effects on taste perception. Sweetness and bitterness, for example, masked each other.
High costs due to tasters
The use of trained tasters is common, but this causes high costs. Online rating databases, on the other hand, are prone to errors because factors such as the price or the current cult status of a product are also reflected there.
The scientists now recorded over 200 chemical properties of 250 Belgian beers, which belong to 22 different beer styles such as Blond, Tripel and Lager. These were linked to descriptive sensory profile data from a trained tasting panel of 16 people on hops, malt and yeast aromas, off-odors and spices, as well as data from more than 180,000 consumer reviews from an online beer review database.
The resulting data set was used to train and test ten machine learning models, which then predicted consumer taste and appreciation. The effectiveness of the most powerful AI approach was tested by implementing predictions to modify an alcoholic and a non-alcoholic commercial beer. In tastings, the AI beers received a better overall rating from the testers, it was said.
Taste is complex
The study confirms that the concentration of flavor compounds does not always correlate with perception – suggesting complex interactions that are often overlooked by conventional approaches.
Overall, the models used are still immature, the research team says. More comprehensive data sets are crucial for further improvements. It must also be clear that the models can only recognize correlations and not causal relationships.
The team is nevertheless convinced that AI could be a basis for the development of novel, tailor-made foods with tastes that are perceived as excellent. This could possibly be used to supplement or replace the usual evaluation of new recipes by trained tasters, which is expensive and time-consuming and can produce varying results.
Verstrepen’s team also combines their study with a warning: such AI approaches should not lead to increasing the addictive potential of alcoholic drinks.
