Math & Engineering

AI Perfumery

Learning odor space from molecular graphs

Can machine learning predict how a molecule will smell? Companies like Osmo (spun out of Google Research) are training graph neural networks to map molecular structure to olfactory perception. The "principal odor map" creates a navigable space of smells—dimensionality reduction applied to human sensation. Predicting smell from structure has been impossible; now it's becoming computable.

The Representation Challenge

Molecules can be represented as graphs: atoms are nodes, bonds are edges. Graph neural networks process these structures, learning features that predict properties. For odor prediction, the network must learn which molecular features—functional groups, ring structures, bond patterns—correlate with which smells.

The training data comes from perfumery databases: molecules with expert labels describing their odors. But smell labels are messy—subjective, culturally variable, and inconsistent across experts. The AI must find signal in noisy, human-generated descriptions.

Why It Matters for Luxury

If smell becomes predictable from structure, perfumery gains a powerful design tool. Rather than synthesizing and smelling thousands of candidates, chemists could search computationally for molecules with desired olfactory profiles. The craft of perfumery wouldn't disappear—but the ratio of creative vision to random exploration would shift dramatically. AI as perfumer's assistant, if not perfumer.

News & Coverage