AI takes on whisky and it's doing just as well as the experts — if not better
Could AI outsniff human experts? Researchers have developed machine learning models that decode whisky aromas with remarkable accuracy,.
Whisky, whether a smoky Scotch or a delicate Bourbon, has a complex aroma. Dozens of different compounds (often, over 40) intertwine to create a complex and unique olfactory profile. But what if artificial intelligence (AI) could break down these complexities to predict the aromas better than human experts?
Some odors are single molecules, but most food aromas consist of a whole range of molecules. From caramel and fruity notes to phenolic and smoky hints, the wide-ranging aromas of whisky are shaped by its chemical composition. Based on odor and taste, experts can tell where a whisky comes from or what type it is… most of the time.
Traditionally, whisky aromas are evaluated by sensory panels — groups of trained humans who describe smells using agreed-upon descriptors. However, this approach has limitations. Odor perception is subjective and there’s always a bit of bias involved. Additionally, the training and convening of expert panels is time-consuming and expensive.
That’s where an electronic nose would come in.
A team of researchers at the Fraunhofer Institute for Process Engineering in Germany developed a new machine learning algorithm that analyzed the molecular composition of whisky samples obtained through chemical analysis. The algorithm then predicted their sensory aroma profiles, outperforming human panels in accuracy and consistency.
What’s in a whisky
The team used a dataset of 16 whiskies (9 Scotch, 7 American) including Jack Daniel’s, Maker’s Mark, Laphroaig, and Talisker. They also had a panel of 11 experts evaluating the whiskies. The researchers paired chemical analyses with human sensory data to train AI models. Their goal was twofold: identify key molecules distinguishing whisky types and predict their most prominent aroma descriptors.
The researchers employed gas chromatography-mass spectrometry (GC-MS) to analyze whisky samples. These molecular profiles were then fed into two AI models:
- OWSum (Olfactory Weighted Sum): A linear classifier designed to distinguish between American and Scotch whiskies based on molecular or sensory data.
- Convolutional Neural Network (CNN): A model that leveraged structural information of molecules to predict specific odor attributes.
Using OWSum, the team successfully classified whisky origin with remarkable accuracy. When relying on molecular data, the model achieved 100% accuracy. Descriptors such as “caramel-like” for American whiskies and “phenolic” or “apple-like” for Scotch whiskies further refined predictions, enabling the AI to tell where the whisky is from. By analyzing molecular features, researchers pinpointed compounds like menthol and citronellol, unique to American whiskies, and methyl decanoate and heptanoic acid, exclusive to Scotch.
As good as human experts — or better
Both models offered insights into the molecular basis of aroma descriptors. For example, the compounds driving “caramel-like” notes differed significantly from those linked to “apple-like” aromas. Visual tools like molecular influence diagrams showcased these relationships, bridging the gap between chemistry and sensory perception.
The research went a step further, using AI to predict the top five aroma descriptors for each whisky. These descriptors included terms like “fruity,” “smoky,” “vanilla-like,” and “woody.” The CNN outperformed OWSum in this task, achieving an accuracy of over 70%. For context, these scores surpassed the inter-panelist agreement.
This research has implications far beyond the whisky industry. The methodology could be applied to other complex odor mixtures, from perfumes to food products. The most straightforward application would be against counterfeiting, ensuring product authenticity. Whisky fraud is nothing new, and it’s a widespread problem. AI tools could also be used for quality control in distilleries, or even to develop new flavor profiles.
However, the dataset was limited to 16 whiskies, and it’s not clear how the algorithms would perform on a larger sample size. Future research could also explore dynamic aroma profiling, capturing how whisky odors evolve over time as they interact with air.
The study was published in Nature Communications Chemistry.
