The Scent of Data: Artificial Intelligence Conquering the Sense of Smell

Artificial Intelligence is digitizing the last of our senses: smell. Researchers have trained algorithms capable of translating the chemical structure of an unk

In the laboratory of a prestigious perfume maison in Paris, a professional "nose" no longer stares at a wall of glass vials, but at a computer screen. In front of him, an algorithm suggests that a particular combination of molecules, never tested before, will produce a perfume with notes of sandalwood, black pepper, and an unexpected saline accent. Until a decade ago, discovering that combination would have required months of trial and error. Today, Artificial Intelligence (AI) promises to digitize the most elusive, subjective, and primordial sense we possess: smell.

The entry of algorithms into the art of perfumery is redefining the very concept of chemical creativity. But between mathematical prediction and the emotion evoked by a fragrance on human skin, there is an abyss of complexity that the machine has not yet bridged.

In this in-depth piece for the Scenarios and Reflections column, we will map out this invisible revolution. We will discover how AI manages to translate the structure of a molecule into words, how commercial systems already assist perfumers in formulation and, above all, we will explore the insurmountable limits of a technology that, however advanced, can never "smell" the world as we smell it.

1. The Principal Odor Map: Translating Molecules into Words

The fundamental problem of artificial olfaction is its intrinsic non-linearity. While in vision we can map colors onto a spectrum of wavelengths, for odors there is no simple chemical rule. Molecules with nearly identical structures can have radically different odors (e.g., a floral scent versus the smell of urine), while structurally opposite molecules can produce the same olfactory perception.

To solve this puzzle, researchers at the Monell Chemical Senses Center and the startup Osmo have taken a historic step, described in a study published in the journal Science ("A Principal Odor Map Unifies Diverse Tasks in Human Olfactory Perception") [1838]. The team trained a neural network on a database of approximately 5,000 odorant molecules, associated with linguistic descriptions provided by experts (e.g., "fruity," "woody," "herbaceous").

The result is a "Principal Odor Map." During tests with hundreds of molecules completely new to the model, the AI managed to predict their olfactory descriptors with surprising accuracy. In over 50% of cases, the machine's predictions came closer to the average of human evaluations than even a single individual's judgment did.

This study, and others similarly analyzed in reviews on PMC ("A comparative study of machine learning models on molecular odor prediction"), demonstrate a revolutionary principle: AI has learned to associate the complex geometry of a molecule with the perception described in words by human beings [1817].

2. The Nose's Assistant: Philyra and Carto in Commercial Practice

While basic science maps individual molecules, the perfume industry is already using AI to orchestrate complete formulas, drastically reducing development times (time-to-market). The machine does not replace the perfumer, but amplifies their ability to explore uncharted territories.

The best-known example is Philyra, the Artificial Intelligence-based assistant developed by the fragrance house Symrise in collaboration with IBM Research [1827]. Philyra was trained on millions of historical formulas, demographic data, sales trends, and physical information about raw materials. When given a demographic target (e.g., "Brazilian millennials interested in fresh fragrances"), the AI proposes combinations and accords, often suggesting dosages or ingredients that a human perfumer would consider counterintuitive, leading to the creation of fragrances successfully launched on the global market since 2019.

Another example of human-machine integration is Carto, the tool launched by the Swiss giant Givaudan [1826]. Carto presents itself as a large touch-screen on which the perfumer visually organizes raw materials to build their idea. Once the formula is defined, an integrated robot comes into play, instantly producing test samples, eliminating the downtime of experimentation. In all these industrial cases, the company specifies one non-negotiable point: the human "nose" remains at the center of the creative process. AI proposes and accelerates; the human chooses and refines.

3. The Limit of the Algorithm: Perception Cannot Be Calculated

However, predicting that a compound will be associated with "green and citrusy notes with a mineral nuance" is a statistical problem; ensuring that a human being will feel joy or repulsion upon smelling it is a biological mystery.

As highlighted by a comprehensive review published in Artificial Intelligence Review ("Advances in artificial intelligence for olfaction and gustation"), the limits of this technology are structural and deeply tied to our physiology [1810].

Artificial Intelligence analyzes the chemical potential of an isolated molecule, but in real perfumery, perception depends on variables that escape the algorithm:

  • The Mixture Effect: Concentration drastically alters the odor (molecules that smell of jasmine at low doses can smell rotten when concentrated), and the interaction with hundreds of other notes in the same bottle creates effects that cannot be calculated in advance.
  • Genetic Subjectivity: Each of us possesses about 400 olfactory receptors, but there are millions of genetic variations (polymorphisms) among individuals. A molecule that reminds me of musk might be odorless to another person (specific anosmia) or unpleasant.
  • The Chemistry of the Skin: The perfumer's masterpiece comes to life only when it warms on human skin, blending with the pH, diet, and sweat of the wearer, transforming the perfect formula into a unique and unrepeatable experience.

Conclusions: The Resilient Magic of Human Synthesis

Artificial Intelligence is not killing perfumery, but it is freeing it from the burden of random trial. If a perfumer is looking for a note of "aged leather and spiced vanilla," they no longer have to manually test a thousand vials; the system will suggest the ten molecules with the highest statistical probability of success, also proposing the formula to blend them and extracting a physical sample via a robotic arm.

Yet, at that point, the calculation stops. Someone will have to close their eyes, spray that liquid onto a strip of paper (the mouillette), and smell it. They will have to understand whether that combination of atoms, statistically perfect, manages to evoke a childhood memory, to seduce, to comfort, or to surprise. AI can predict the "scent of data," but until machines have memories, skin, and heart, the alchemy of transforming an odor into an emotion will remain an exclusively human miracle.

Bibliographic References and Sources

Article by the Editorial Team of La Bussola dell'IA