What if artificial intelligence surpassed humans in the art of choosing a single malt? Machine learning algorithms were able to predict the dominant aromas of different whiskeys better than an expert, according to a study published Thursday.
In our environment, the majority of odors are made up of a complex mixture of molecules that interact in our olfactory system to create a specific impression.
This is the case for whiskey, whose aromatic profile can be determined from more than 40 compounds and which can contain even more non-odorous volatile compounds.
This makes it particularly difficult to evaluate or predict the aromatic characteristics of a whiskey when based solely on its molecular composition.
Yet this is what chemists managed to do thanks to two machine learning algorithms, according to the results of a study published Thursday in Communications Chemistry.
The first algorithm, OWSum, is a statistical tool for predicting molecular odors developed by the study authors.
The second, CNN, is a convolutional neural network, which helps discover relationships in very complex data sets. Like those between “the most influential aroma molecules and attributes” in a mixture of whiskey, explains to AFP Andreas Grasskamp, researcher at the Fraunhofer Institute for Process Engineering and Packaging IVV, in Freising (Germany), and main author of the study.
The researchers have “trained” algorithms by providing them with a list of molecules detected by gas chromatography and mass spectrometry (two techniques for separating molecules in mixtures and identifying them) in 16 whiskey samples: Talisker Isle of Skye Malt (10 years old 'age), Glenmorangie Original, Four Roses Single Barrel, Johnnie Walker Red Label or even Jack Daniel's…
They also gave them the aroma descriptors determined for each sample by a panel of 11 experts.
The algorithms were then used to identify each whiskey's country of origin and its five dominant notes.
Detect counterfeits
OWSum was able to determine whether a whiskey was American or Scottish with over 90% accuracy.
The detection of menthol and citronellol molecules was strongly associated with an American classification, while the detection of methyl decanoate and heptanoic acid was mainly linked to a classification as Scotch whisky.
The algorithm also identified caramelized notes as the most characteristic of American whiskeys, while notes “apple”, “solvent” et “phenolic” (often described as a smoky or medicinal smell) were the most characteristic of Scottish whiskeys.
The researchers then asked OWSum and CNN to predict the olfactory qualities of whiskeys based either on the molecules detected or on their structural characteristics.
Both algorithms were able to identify the five dominant notes of a given whiskey more accurately and consistently on average than any human expert on the panel.
“We found that our algorithms aligned better with the panel results than each panelist individually, providing a better estimate of overall odor perception.”underlines Mr. Grasskamp.
These machine learning methods could be used to detect counterfeits. Or to evaluate whether a mixture of whiskey “will have the expected aroma, helping to reduce costs by limiting the need for evaluation panels”he believes.
Could similar results be obtained with wine? “In theory yes, all these tools need is a list of compounds detected in the sample and their corresponding descriptors”according to Mr. Grasskamp.
“The challenge remains in the finer details, like whether wine aromas are distinct enough for an AI algorithm”he adds.
“All rights of reproduction and representation reserved. ©2024 Agence France-Presse. All information (text, photo, video, fixed or animated infographics, sound or multimedia content) reproduced on this page is protected by current legislation on intellectual property rights. Consequently, any reproduction, representation, modification, translation, commercial exploitation or reuse in any manner whatsoever is prohibited without the prior written consent of AFP, with the exception of personal non-commercial use. The AFP cannot be held responsible for delays, errors, omissions which cannot be excluded in the field of press information, nor for the consequences of actions or transactions carried out on the basis of this information. AFP and its logo are registered trademarks »
