Whisky, wherever it’s made in the world, is well-known for its complex aroma. On a chemical level, whisky can have over 40 different chemical compounds that are contributing to what we smell when we raise our nose to the glass.
One of the joys of whisky is the subjectivity. Not only in what we like and dislike, but what we experience when we nose or taste a whisky. If you’ve been to a whisky tasting you’ll have likely heard the person leading the group say “there’s no right or wrong answers” when asking their audience for suggestions about tasting notes.
But what if there was a scientific way to determine the ‘right’ answer? Could an electronic nose accurately assess a whisky’s aromas, removing human subjectivity from the equation?
This is the question asked by a team of researchers at the Fraunhofer Institute for Process Engineering in Germany. The team there have developed a machine learning algorithm that studies the molecular composition of whisky samples to predict their sensory profiles.
AI models were trained by pairing chemical analyses of these samples with human sensory data. The main objective was to identify key molecules within the samples in order to distinguish between different types of whisky, such as where that whisky could be made. Additionally, the models were used to predict the main aroma descriptors - or what you and I would refer to as a whisky’s nose. The data from the AI models was then evaluated against notes from a panel of 11 industry experts.
For the study, 16 different whiskies were used - nine Scotch and seven American - and included some well known names such as Jack Daniels and Talisker. Characteristics such as ‘caramel-like’ were attributed to American whiskies, while ‘phenolic’ or ‘apple-like’ were used for Scotch. Because the molecules that contribute to these aromas are so different, it helped the AI detect where the samples were from.
When the molecular profiles were fed into the AI, one of the models was able to predict the whisky’s origin with 100% accuracy. This is certainly an incredible result, despite some of the limitations of the methodology (which we’ll get to in a bit).
Both AI models were also used to predict the top five aroma descriptors based on the molecular data. When using commonly used terms such as ‘woody’, ‘fruity’, and ‘smoky’, one of the models was able to achieve a score of over 70% accuracy, which was a higher score than the consensus reached by the panelists.
While these results are impressive on the surface, there are definitely some limitations with this research. Firstly, there was an incredibly limited dataset and the researchers are currently unsure of how it would perform with a larger sample size.
The nature of the samples also feels as if it set up an easy score for the AI. Even the most untrained nose would be able to pick out some difference between Makers Mark and Laphroaig or JD and Talisker. Also, identifying which is which wouldn’t be difficult if that person was told before hand that the crazy smoky and phenolic one is the Scotch.
If more work is done in this area through there’s definitely some positive, practical implications for this research. For example, molecular-driven sensory analysis could help prove authenticity and detect counterfeit products.
AI could also be used for quality control. In larger distilleries much of the production process is already computer-driven and automated, so implementing AI for quality control could be the natural next step. Swedish distillery Mackmyra have also explored using AI to generate new whisky recipes. Right now these whiskies are generally regarded as a fun, experimental novelty rather than ‘serious’ whisky, but that doesn’t mean there isn’t scope for the to change in the future.
Returning to the question we asked earlier - is there a way to determine the right answer when it comes to whisky tasting notes. The answer certainly appears to be yes. The fact that technology has developed to be able to accurately identify aromas is impressive, but it does raise another question: do we need it?
As we’ve already said, the inherent subjectivity of whisky tasting is one of the things that makes it so enjoyable.
The differences are an interesting point of conversation. Half-drunkenly quibbling with a friend over whether you’re picking up notes of tablet or fudge on a given dram is part of the fun. Do we really need a computer to tell us that what we’re really smelling is the chemical compound most commonly associated with caramel? It’s a bit like that one person you know who annoyingly butts in with “well, actually, you’ll find that…” whenever anyone gets something slightly wrong in conversation.
Generally speaking, AI-driven technology is created with the goal of using human-like intelligence and experience to perform a given task. A comprehensively crafted AI should, in theory, be able to complete that task in the way a human would.
Right now, the relationship between whisky and artificial intelligence is at an interesting stage. AI models are being taught how to detect flavours and make accurate assumptions based on chemical analysis, but I think it’s all still too early to even consider the idea that technology might soon replace the role of whisky experts in the real world - be they master blenders, brand ambassadors, or even your mate who knows more about whisky than anyone you’ve ever met.
So don’t worry, there robots won’t be hosting your local whisky tasting events any time soon. Maybe.
