Tuesday, November 17, 2026

Can AI learn to design food flavours?

Taste may feel subjective, but AI is helping scientists uncover the molecular patterns that make certain flavours work.
2 mins read
Robotic arm decorating colorful cupcakes and pastries in a bakery kitchen


Nearly 4,000 years ago, cooks in Mesopotamia were already experimenting seriously with flavour. Recipes preserved on Babylonian clay tablets describe elaborate stews combining meat, beer, leeks, garlic, herbs and spices. Yet there was little of the precision we associate with a modern recipe. Much was left to judgement, smell and taste.

Centuries later, Greeks and Romans perfected another remarkable exercise in flavour engineering: garum. Fish and salt were left to ferment until enzymes and microorganisms transformed the mixture into an intensely savoury condiment that became ubiquitous in Roman cooking. Its makers understood nothing of volatile molecules, amino acids or microbial metabolism. They simply knew that manipulating ingredients, time and biological processes could transform taste.

Humans, in other words, have been engineering flavour for millennia.

What is changing is the instrument.

Artificial intelligence is now entering a field known as flavoromics, where scientists attempt to map the molecular architecture behind what we taste and smell.

Modern analytical instruments can take food apart chemically. Gas chromatography-mass spectrometry can identify volatile compounds associated with aroma. Electronic noses detect patterns among airborne molecules. Electronic tongues analyse chemical signatures associated with taste.

But measuring molecules creates another problem: an enormous amount of data.

A food can contain hundreds of volatile compounds, and abundance alone does not determine importance. Some molecules influence aroma at extremely low concentrations, while others contribute little despite being plentiful. Compounds also interact, and perception changes with temperature, texture and the food matrix surrounding them.

This is where machine learning becomes useful.

Algorithms can search chemical and sensory datasets for relationships that would be extraordinarily difficult to identify manually. Scientists can ask which combinations of molecules are associated with a particular aroma, which fermentation conditions produce desirable sensory characteristics or which chemical signatures distinguish one product from another.

Researchers studying baijiu, the complex Chinese distilled spirit, have already combined sensory analysis, flavoromics and machine learning. In one study, scientists identified 18 important odour-active compounds and developed models capable of distinguishing quality grades. Explainable-AI techniques were then used to determine which compounds contributed most strongly to those predictions.

The important point is not that AI has learnt to enjoy baijiu.

It hasn’t.

AI does not possess a palate. What it can increasingly do is connect molecular measurements with human sensory judgements.

That distinction could transform food development.

Traditionally, creating a new flavour or reformulating an existing food involves repeated experimentation: change an ingredient, alter a process, produce another prototype and return it to a sensory panel.

Computational models could increasingly narrow that search before the first bite is taken.

A food scientist trying to reproduce a roasted aroma, for example, might use modelling to identify the compounds most strongly associated with it and then investigate which ingredients or processing conditions generate them.

Eventually, the question could even be reversed.

Instead of asking, What will this formulation taste like?, developers could begin with a desired sensory profile and ask, What chemistry and processing conditions are most likely to produce it?

That possibility is particularly important for foods that are difficult to reformulate. Removing salt, sugar or fat changes much more than nutrition. It can alter mouthfeel, aroma release, texture and the overall sensory experience. Alternative proteins pose an even more complicated problem because reproducing meat requires the interaction of aroma, fat behaviour, texture and chemical reactions during cooking.

AI could help scientists explore these combinations far more systematically.

But there is a boundary that technology has not crossed. Flavour is not chemistry alone. Memory, culture, expectation, temperature, texture and individual biology all influence whether something tastes wonderful or terrible.

The Babylonian cook adjusted a stew through experience. The Roman maker of garum worked with knowledge accumulated over generations.

Tomorrow’s food scientist will still need the human palate.

But for the first time, there may also be an algorithm capable of showing us some of the hidden molecular logic behind why that palate says yes.

The author is a journalist and AI enthusiast who believes technological progress is most meaningful when it expands human possibility without shrinking nature’s space

The views expressed in this article are entirely those of the author and do not necessarily reflect those of the organisation.

Don't Miss

When AI becomes your colleague

The working day of 2035 could begin without opening an inbox. Instead,…

There is a lot of creativity in mathematical traditions: Prof Ramasubramanian

The scholar described how patterns of short and long syllables could be…