From Molecules to Menus: How AI Is Decoding the Chemistry of Flavor and Scent

Artificial intelligence is revolutionizing food and flavor science by decoding the chemistry of taste and scent. AI systems like FlavorGraph and the "AI Connoisseur" are enabling novel ingredient pairings, guiding crop breeding for better flavor, and predicting scent from molecular structures.

Dr. Lena Hanson

May 29, 2026 · Updated November 6, 2026 · 4 min read

Artificial intelligence is now being used to predict what foods will taste like together, which crops breeders should grow for better flavor, and even how a molecule will smell before anyone takes a sniff. Three research efforts illustrate the range: Sony AI and Korea University built FlavorGraph to suggest novel ingredient pairings, University of Florida researchers created an "Artificial Intelligence Connoisseur" to guide fruit breeding by consumer taste, and separate work on graph neural networks aims to predict scent directly from molecular structure. All tackle the same stubborn problem: the leap from a chemical structure to a human perception is anything but straightforward.

AI for Food Pairing: The FlavorGraph System

FlavorGraph is a graph embedding model—built on the metapath2vec framework—developed by researchers at Sony AI and Korea University. Its job is to predict how two ingredients will pair together, using two very different kinds of data at once: recipe knowledge accumulated across cultures and the chemical structures of flavor molecules.

As reported by Isha Salian on the NVIDIA Technical Blog, FlavorGraph was trained on a million recipes plus chemical structure data from more than 1,500 flavor molecules, with the team using PyTorch, CUDA, and an NVIDIA TITAN GPU to train and test the resulting large-scale food graph. The model outperformed other baseline methods for food clustering and, more interestingly, revealed connections between foods' flavor profiles and their underlying chemical compounds.

The ambition here is explicitly creative. The researchers behind FlavorGraph aim to enable the discovery of new recipes and novel flavor combinations, and to identify potential substitutes for unhealthy or unsustainable ingredients. In other words, the pairing tool is a discovery engine: not just confirming that tomato and basil work, but surfacing mash-ups no chef has tried, grounded in shared chemistry rather than tradition alone.

AI for Crop Breeding: The AI Connoisseur

Where FlavorGraph starts from recipes, the University of Florida's "Artificial Intelligence Connoisseur" starts from the fruit itself. Marcio Resende, a breeder and geneticist at the university's Institute of Food and Agricultural Sciences, is behind a system that determines which chemical compounds produce the best fruit flavors, based on consumer preference data.

As Agriculture Dive reported in November 2023, the system works by linking the chemistry of fruit—its flavor compounds—to what consumers actually say they enjoy. That answers a question breeding programs have historically sidestepped: crop breeding traditionally prioritizes traits such as yield and disease resistance over taste and nutritional value.

The stakes are nutritional as much as culinary. "By using AI to help breed and plant crops that are not only more flavorful, but also better for our body and mind, we can contribute to a healthier society," one of the project's researchers said. The AI Connoisseur effectively gives breeders a shortcut: instead of waiting years for field trials and taste panels, they can identify which chemical signatures correlate with consumer-approved flavor and breed toward them directly.

Comparing AI's Role in Flavor and Scent Innovation

Set side by side, these systems show that "AI for flavor" is really several different problems wearing one label—one of prediction, one of optimization, and one of structural interpretation. The table below compares the objectives, methods, inputs, and outputs of each approach.

Three distinct applications of AI in flavor and scent science
Application	Goal	AI Model / System	Input Data	Example Outcome
Food pairing	Predict which ingredients will pair well and suggest new combinations	FlavorGraph, a graph embedding model built on metapath2vec	A million recipes plus chemical structures of over 1,500 flavor molecules	Pairing suggestions that outperform baseline clustering methods, revealing links between flavor profiles and compounds
Crop breeding	Identify which chemical compounds produce the best fruit flavors	The "Artificial Intelligence Connoisseur" from the University of Florida	Fruit flavor chemistry and consumer preference data	Breeding guidance toward crops that are both flavorful and nutritious, rather than yield-focused alone
Scent prediction	Predict how a molecule will smell from its structure	Graph neural networks operating directly on molecular graphs	Molecular structure	Predicted scent descriptors tied to chemical structure rather than to recipe or preference data

The comparison exposes a real distinction in method. FlavorGraph fuses cultural data (recipes) with chemistry; the AI Connoisseur fuses chemistry with human preference; scent prediction models skip human context entirely, mapping molecular structure straight to perception. Each answers a question the others cannot.