An artificial intelligence capable of predicting the taste of beer based on its chemical composition could assist in creating non-alcoholic beers that taste just like their alcoholic counterparts. Recent research findings on this topic have been published in Nature Communications.
With the help of artificial intelligence, various beers can be brewed using only a small amount of raw materials. Image Source: Kutredrig/Getty Images
Predicting taste from compounds is challenging due to the complex interactions between ingredients and sensory perception. To tackle this, Kevin Verstrepen and colleagues at KU Leuven in Belgium developed an AI model that can forecast the taste of beer based on its chemical composition and offer suggestions on how to improve it.
A panel of 16 professional beer tasters rated 50 attributes of each beer. The model was trained using these evaluations from the panel, as well as 180,000 public reviews from an online beer rating website. It compared these subjective descriptions with measurements of 226 compounds in 250 Belgian beers.
"Hundreds of compounds are received by our noses and mouths, then processed in the brain to create what we perceive as flavors. Now, you can accurately predict this through machine learning. It's really amazing," says Verstrepen.
Using this AI model, Verstrepen's team predicted how blending mixtures of specific compounds such as lactic acid and ethyl acetate could enhance beer taste. The final beers received higher ratings from the tasting panel.
In another unpublished study cited by Verstrepen, after implementing changes suggested by the model, non-alcoholic beer became indistinguishable from regular beer.
Brewers should implement the AI model's recommendations by adjusting formulations rather than simply adding flavors. Verstrepen argues that solely adding pure aroma compounds in beer brewing is unacceptable.
"I see the model as a tool, particularly for brewing better non-alcoholic beers, without depriving the art of crafting good beer by hand," says Verstrepen.
Related Paper: https://doi.org/10.1038/s41467-024-46346-0
