(Photo: Jane Barlow/PA Wire URN:59277880 Press Association via AP Images)
Earlier this month, German scientists created a machine-learning model that can purportedly detect whisky aromas and origins on par with — and occasionally, better than — a human expert.
The algorithm, developed at the Fraunhofer Institute for Process Engineering and Packaging IVV, was tasked with sniffing out the difference between 16 whisky samples, nine of which were Scotch and seven American. Brands included household names like Jack Daniel's, Talisker, Laphroaig and Maker's Mark. Detecting keyword descriptors like "caramel-like," "apple-like" and "solvent-like," the model was reportedly able to distinguish the difference between the two countries of origin with 93.75% accuracy.
Researchers adjusted their approach accordingly. Narrowing in on the finer details, the team fed the AI a reference dataset of 390 molecules commonly found in whiskies. The neural network, dubbed CNN, was subsequently able to differentiate American whiskey from Scotch in 100% of cases. While the presence of menthol and citronellol indicated the former, methyl decanoate and heptanoic acid were telltale signs of the latter.
The model then competed against 11 human panelists to detect the top five odors in each sample based on their chemical composition. Each group was judged on a sliding scale from 1, for perfect accuracy, to 0, for consistent inaccuracy, The neural network achieved 0.78. The human experts, only 0.57.
“The beautiful thing about the AI is that it is very consistent,” lead researcher Dr. Andreas Graskamp told The Guardian. “You have this subjectivity still in trained experts. We are not replacing the human nose with this, but we are really supporting it through efficiency and consistency.”
The glaring caveat of all this comes down to subjectivity.
Though the neural network may have beaten humans at detecting the "correct" aromas, tasting notes don't boil down to a list of chemical compounds — at least not in the eyes of most enthusiasts. Here at Bottle Raiders, our spirit reviews frequently point out such far-fetched scents as potpourri, sawdust and gummy bear sweetness. On more than one occasion, we've differentiated between buttercream, brown butter, buttery pastry and butterscotch. There's a certain poetry to the details, one that we imagine a mechanical model wouldn't (and couldn't) make.
But perhaps that's the point. As billion-dollar giants like Jack Daniel's and Johnnie Walker scale up their production capacity, a certain degree of objectivity is necessary to maintain consistency. Dr. William Peveler, a senior lecturer in chemistry at the University of Glasgow, says that AI models might be the solution.
“The flavour notes of a whisky brand could be quickly checked from batch to batch or blend to blend based on the chemical signature alone, to try to ensure a consistent house style," Peveler added.
