Ask most people what flavour is and they’ll say taste and smell. Ask a sensory scientist and you’ll get a longer, stranger answer: flavour is something your brain constructs from many channels at once — aroma and taste, yes, but also texture, temperature, sound, and even colour. The crunch changes how fresh a crisp seems. A warmer soup tastes saltier. A redder drink tastes sweeter before you’ve swallowed. The mouth is only part of the instrument.
This is why the most interesting development in AI flavour models is a quiet shift in architecture: from systems that read mainly aroma and taste chemistry, to a multi-sensory integration framework that combines tactile, thermal and visual inputs alongside the chemical profile. Kerry’s 2026 taste forecast even names “maximalist”, layered multi-sensory flavour as a trend. The models are, belatedly, being built the way perception actually works.
1Why single-channel models hit a ceiling
A model that predicts flavour from aroma compounds and taste chemistry alone can be impressively accurate — and still be blindsided in the real world, because it is missing information the eater is using. Two products with an identical aroma-and-taste fingerprint can land completely differently if one is smooth and one is grainy, one is warm and one is cold, one is vivid and one is dull. The chemistry says they match; the human says they don’t. That gap is not noise. It is the other senses doing their job, and a model that ignores them will keep being confidently wrong in the same direction.
Flavour isn’t in the food. It’s assembled in the head — from every sense at once.
2Tune the model’s senses
Set how much attention a flavour model pays to each sensory channel and watch its predicted match to real human experience. The lesson is built into the maths — and into perception: pouring everything into aroma and taste gets you part-way, but the score only climbs when the model covers the senses the way a person does.
3The honest limits
Two cautions keep this useful rather than breathless. First, cross-modal data is scarce and hard: we have mountains of aroma-compound data and comparatively little clean data linking texture, temperature and appearance to perceived flavour, so multi-sensory models are often data-starved exactly where they’re most novel. Second — and this is the one I’d underline — these models augment trained human panels; they do not replace them. The human panel remains the ground truth a model is trying to approximate; without it there is nothing to train against and nothing to check against. Anyone promising to retire the sensory panel has misunderstood what the panel is for.
The market noise doesn’t always help here. Forecasts for AI in food and drink are eye-watering — one widely-cited estimate puts the sector on a compound growth rate above 40% into the 2030s — but those figures span everything from supply-chain software to this kind of sensory modelling, and estimates vary enormously depending on what’s counted. Growth is real; the specific numbers deserve a pinch of salt.
4Where I think this goes
The version of this I find most compelling isn’t a model that guesses flavour from a photo. It’s a system that fuses three things a good product team already generates: hard analytical data (what a GC-MS says is actually in the aroma), structured sensory-panel data (what trained humans reliably perceive), and these cross-modal models (how texture, temperature and appearance are likely to bend that perception). Kept honest by real panels and real instruments, that fusion is a genuinely useful design tool — a way to reason about the whole eating experience earlier, instead of discovering at launch that a technically-correct flavour felt wrong in the mouth. Flavour was always multi-sensory. It’s good that our tools are finally catching up to the senses.