Mapping the Global Senses (Part 3 of 3): Flavors Meet the Machines

Part 1 of this series appeared in the April 2026 issue of Perfumer & Flavorist+; Part 2 of this series appeared in the September edition of the magazine. -Editor

Saikindō, a cocktail bar inside the Four Seasons Hotel Abu Dhabi, draws from Japan's listening bar culture—spaces where the music is not background but the point—and filters it through omotenashi, the Japanese hospitality philosophy built around anticipating what a guest needs before they say it. EdNurg at Adobe Stock

Editor's introduction: From an AI-driven cocktail bar in Abu Dhabi to machine-learning systems reshaping formulation, the boundaries between sound, emotion, memory and flavor are beginning to blur. But as algorithms get better at predicting what consumers might like, author Mindy Yang says one stubborn question remains: can AI learn what flavor actually feels like?

Abu Dhabi: When Sound Becomes Something You Can Taste

Everything discussed so far in this series—fermentation, emotional coding, the tension between novelty and familiarity—converges in an unlikely place: a cocktail bar inside the Four Seasons Hotel Abu Dhabi.

Saikindō opened in late 2025, designed by AvroKO, a firm whose portfolio includes James Beard Award-winning restaurants and projects across 22 countries. The concept draws from Japan's listening bar culture—spaces where the music is not background but the point—and filters it through omotenashi, the Japanese hospitality philosophy built around anticipating what a guest needs before they say it. 

Japan's listening bar culture features spaces where the music is not background but the point.Szymon Shields at Pexels

The interior pulls from two unexpected sources: Japan's post-war Metabolism movement, which favored modular, adaptable architecture, and the Bosozoku biker subculture, with its customized motorcycles and embroidered jackets. L-Acoustics speakers are built into wooden panels at mid-height rather than aimed at the room, dispersing sound evenly so that every seat hears the same thing.

The bar is led by beverage director Marco Corallo, who previously worked at London's Artesian—the bar that held the World's Best Bar title for four consecutive years, from 2012 to 2015—alongside head mixologist Ann Pinsuda, formerly of The Bamboo Bar at Mandarin Oriental Bangkok, which debuted at number 35 on the World's 50 Best Bars list. The credentials are serious. But what's technically interesting for anyone working in flavor development is what they're attempting with AI.

The team used machine learning to translate musical elements—tempo, tone, key—directly into cocktail recipes, creating drinks that correspond to whatever is playing on any given night. This is not as speculative as it might sound. 

The research foundation already exists. Cédric Colas, working out of a computational neuroscience background, trained a self-supervised transformer model on 30,000 MIDI files and mapped musical pieces into a shared representational space with cocktail recipes. His system represents each cocktail in a 13-dimensional "taste space" that includes alcohol content, sourness, sweetness, bitterness, herbiness, fruitiness and complexity. It learns to connect music and taste through shared semantic labels—"Cuban" applying to both Latin jazz and rum-and-mint cocktails, "Romantic" applying to both Chopin nocturnes and complex, bittersweet drinks.

Before Colas, Tramontina partnered with Spotify and neuroscientist Dr. Marcelo Costa to build a synesthetic algorithm that translated millions of songs into flavor profiles. Beethoven's Fifth became foie gras terrine with salmon caviar. A samba rhythm mapped to dense, layered food. The results were imperfect—synesthesia is not a solved problem—but they demonstrated that the connection between sound and taste is not purely metaphorical. There are measurable neurological overlaps between auditory and gustatory processing, and machine learning is starting to find them.

What Saikindō is doing is essentially a commercial deployment of this research: a bar where the cocktail menu shifts with the music, where every drink is available in both alcoholic and non-alcoholic versions, built around seasonal Japanese ingredients, sake, and shōchū. Whether it works—whether guests actually experience cocktails that feel compositionally linked to what they're hearing—remains to be seen. But it represents something genuinely new: the first serious attempt to use AI-driven synesthetic flavor translation in a real hospitality context, at a venue with the talent and the backing to do it properly.

Cédric Colas trained a self-supervised transformer model on 30,000 MIDI files and mapped musical pieces into a shared representational space with cocktail recipes. His system represents each cocktail in a 13-dimensional "taste space" that includes alcohol content, sourness, sweetness, bitterness, herbiness, fruitiness and complexity. It learns to connect music and taste through shared semantic labels such as "Romantic," which applies to both Chopin nocturnes and complex, bittersweet drinks.Emilia at Adobe Stock

The food follows an izakaya rhythm—compact, shareable, designed for comfort rather than spectacle. Maguro crunch with spicy tuna and grilled Japanese eggplant, a melt-in-your-mouth wagyu skewer that is likely to become one of the venue's defining dishes.

Bob Suri, general manager of Four Seasons Hotel Abu Dhabi, described it as "an inspiring journey of creativity, collaboration and innovation," developed with support from Abu Dhabi's Department of Culture and Tourism. The programming extends beyond the bar itself: vinyl takeovers, guest DJ sets, record collectors' nights, masterclasses on Japanese spirits. Every detail is orchestrated, which is either the purest expression of omotenashi or the most elaborate marketing exercise in the history of cocktail bars. Possibly both.

Lausanne, Switzerland: When the Algorithm Learns Restraint

Everything discussed so far in this series is being fed into algorithms now—the fermentation data, the emotional coding, the nostalgia research, the sound-to-flavor experiments. The question is whether the algorithms are sophisticated enough to do anything useful with it.

Givaudan's ATOM tool uses AI to identify positive and negative flavor drivers and explore ingredient synergies. In an early project focused on reducing salt in cheese snacks, ATOM identified the ideal recipe quickly enough to justify the investment, delivering a 33% salt reduction that scored as highly as the original in blind testing. The company's Carto system takes it a step further: an intuitive touchscreen interface where perfumers select raw materials and a robot produces instant samples, collapsing the development cycle from weeks to hours. 

Most flavor preference research captures what people say they like—survey responses, purchase history, click patterns. What it doesn't capture, or captures very poorly, is what people actually experience when they taste something.Gaspar Zaldo at Pexels

Elsewhere, Symrise collaborated with IBM on Philyra, an AI assistant designed specifically for perfumery innovation. In addition, Firmenich (prior to the formation of dsm-firmenich) announced its first AI-created flavor in partnership with Microsoft in 2020 and followed it with Scentmate in 2021.

The tools are impressive. The limitation is always the data feeding them. Most flavor preference research captures what people say they like—survey responses, purchase history, click patterns. What it doesn't capture, or captures very poorly, is what people actually experience when they taste something. Those are not the same thing, and the gap between them is where AI flavor development currently lives.

The Human Problem Nobody's Solving Fast Enough

Paul Breslin at Rutgers has spent decades studying the intersection of taste, smell and what we actually call flavor—which turns out to be a far more complicated neurological event than most people realize. His work focuses on taste discrimination, enhancement, suppression, and the interactions among gustation, chemesthesis and olfaction that together constitute the experience of flavor.

The core finding is this: how we perceive flavor when eating is fundamentally different from how we perceive it when smelling. The route aroma takes during eating—up through the back of the throat to the olfactory epithelium, what scientists call retronasal olfaction—creates a different neurological signature than smelling through the nose. Research has shown that retronasal odorants activate encoding patterns in the insula similar to their paired tastants, with overlapping representations particularly in the ventral anterior insula. Retronasal odor perception requires taste cortex. Orthonasal does not. These are not subtle differences.

What we actually call flavor turns out to be a far more complicated neurological event than most people realize. Cottonbro at Pexels

And yet people describe "vanilla" the same way whether they're smelling it or tasting it. Most AI flavor models are trained on exactly this kind of descriptive data—language that collapses two distinct neurological experiences into a single word. The models don't know the difference. Neither, in most cases, do the datasets.

Then there's the temporal dimension, which is even harder to capture. Flavor perception is dynamic. The intensity of what you're tasting changes moment to moment as you chew, breathe, salivate, move your tongue, and swallow. Flavor release doesn't happen all at once—it occurs in steps, producing gustatory, olfactory, and trigeminal sensations that layer over each other in ways that time-intensity research is only beginning to map. Studies on potato chip consumption have shown that chewing rate significantly affects maximum flavor intensity. A flavor that works beautifully at 40 chews per minute might fall apart at 120. The algorithm doesn't know that yet.

Clinical trials that track real sensory perception—neurological response, temporal flavor release, emotional associations measured over time—are expensive, slow, and difficult to scale. But they're necessary, because without them we are training machines on proxies. On language instead of experience. On what people say instead of what they feel. 

Where This Lands

The brands that seem to be winning right now are doing something that sounds simple but is remarkably difficult in practice: holding cultural specificity and scalable systems in the same hand. 

Data still can't tell you whether a flavor feels right at 8 a.m. versus 8 p.m. Whether the third bite tastes different from the first. Those are still human questions. And for now, at least, they still need human answers.Helena Lopes

The flavorists who thrive in the next decade won't be the ones who can replicate traditional flavors most accurately, or the ones who can deploy AI most aggressively. They'll be the ones who understand that technology is only useful when it amplifies what humans actually experience—not what humans say they experience, not what humans click on, but what happens in the mouth, in the nose, in the parts of the brain that process taste and memory and emotion all at once.

The data still can't tell you whether a flavor feels right at 8 a.m. versus 8 p.m. Whether the third bite tastes different from the first. Whether it triggers comfort or curiosity. Whether it works at 40 chews per minute or needs 80. Whether it will make someone feel something they didn't expect to feel when they sat down to eat.

Those are still human questions. And for now, at least, they still need human answers.

 FOOTNOTES