AI is giving computers a sense of smell
Computers have learnt to recognise faces, generate voices and interpret medical scans. Smell has proved considerably harder.
That may be starting to change. In July 2026, Japanese start-up ScentifAI announced a proof-of-concept project with an unnamed global technology company to develop what it calls “next-generation smell AI”. The company, previously known as Aroma Bit, is building semiconductor sensors that translate interactions with airborne chemicals into digital patterns.
It is one of several efforts to give machines something resembling a sense of smell. Researchers are using artificial intelligence to predict odours from molecular structures, while companies are developing electronic noses for factories, laboratories and consumer products. Others are exploring whether a scent can be analysed in one place and recreated in another.
The ambition is not merely to build a better air-quality monitor. If smell can be represented as data, computers could help formulate perfumes, detect spoiled food, identify industrial emissions and discover new insect repellents. In time, it may even become possible to send a recognisable scent as readily as an image or audio file.
Why smell is difficult to digitise
Sight and sound are relatively easy to capture because both can be measured as waves. Cameras record light across a grid of pixels, while microphones convert changes in air pressure into electrical signals.
Smell works differently. An odour is usually produced by volatile molecules entering the nose and binding to hundreds of types of olfactory receptor. The brain interprets the resulting combination of signals as coffee, smoke, cut grass or thousands of other impressions.
A rose does not release a single “rose molecule”. Its aroma is produced by a mixture of compounds whose relative concentrations, interactions and evaporation rates affect what a person perceives. Temperature, humidity and an individual’s genetics can alter the experience further.
There is also no established equivalent of the red, green and blue values used to encode colour. Scientists do not yet agree on a small set of primary smells from which all others can be constructed.
Those complications have left digital olfaction researchers with two related problems: how to predict what a molecule will smell like and how to measure the much richer mixtures encountered outside a laboratory.
AI builds a map of smell
A major step came from research led by Alex Wiltschko, now the founder and chief executive of US start-up Osmo.
While at Google Research, Wiltschko and his colleagues trained a graph neural network to examine the structure of an odorous molecule and predict how people would describe it. Instead of treating the molecule as a string of symbols, the system represented its atoms and chemical bonds as a graph.
The model produced what the researchers called a Principal Odor Map: a numerical space in which molecules expected to smell alike sit close together, even when their chemical structures appear quite different.
For a study published in Science in 2023, a trained panel assessed 400 previously unseen molecules against 55 descriptive labels, including “floral”, “mint”, “sulphurous” and “fruity”. When the model’s predictions were compared with the panel’s average ratings, it performed better than the median individual panellist for 53 per cent of the molecules.
That does not mean the AI smells as a person does. It predicts the language people are likely to use after encountering a particular molecular structure. Even so, the research showed that machine learning could uncover useful relationships between chemistry and perception that were difficult to capture with conventional rules.
From individual molecules to real-world scents
Most natural smells contain tens or hundreds of components. Adding those ingredients together does not always produce a straightforward blend of their individual odours. One compound can mask another, while some combinations generate an impression that is difficult to infer from the ingredients alone.
Researchers are therefore extending molecular odour maps to mixtures. A project called POMMix combines graph neural networks, which represent individual molecules, with an attention-based system that estimates the similarity between blends.
The model has been tested on several existing datasets and was designed to work despite the limited amount of public information about how people perceive mixtures. The researchers describe it as a step towards representing complex scents, although the work remains research rather than a finished system capable of reliably decoding any smell.
The shortage of training data is a persistent obstacle. Image models can learn from billions of labelled pictures found online. There is no comparable archive of smells, complete with molecular compositions and consistent human descriptions. Producing one requires laboratory analysis and panels of people willing to sniff and rate thousands of samples.
Descriptions also vary. One person’s “woody” may be another’s “smoky”, while culture and memory can influence how a scent is named.
AI enters the fragrance laboratory
The perfume industry offers an immediate commercial application because it already has large collections of ingredients, formulas and sensory assessments.
Osmo describes its technology as “olfactory intelligence”: software intended to help analyse and create fragrances. In 2025, the company launched an AI-assisted fragrance house that allows clients to develop scents from written concepts, with human perfumers refining the results.
Established fragrance companies are developing their own tools. Givaudan’s Carto system gives perfumers access to a digital formulation interface connected to a robot that produces physical samples. The system can draw on the company’s accumulated knowledge of ingredients and suggest workable formulas, but the perfumer still directs the creative process.
Such systems could reduce the number of experiments needed to reach a desired result. They may also help formulators replace an ingredient that has become expensive, scarce or restricted without losing the character of a product.
The same approach can be used outside fine fragrance. Algorithms could search chemical space for compounds with a desired scent or biological effect. Osmo has reported developing potential insect-repellent molecules through its earlier research, although performance claims still require independent validation and regulatory testing before they translate into products.
Can AI send a smell as data?
Osmo has also demonstrated a process it calls “scent teleportation”.
The company used gas chromatography–mass spectrometry to analyse the volatile compounds emitted by a plum. Its software interpreted the results and generated a formula, which was then blended from available fragrance ingredients to create an approximation of the original aroma.
Osmo says its 2024 demonstration recreated the plum’s scent without a perfumer manually correcting the formula. It was a notable technical exercise, but the term “teleportation” deserves qualification. No smell travelled through a network. Chemical measurements and a reconstruction recipe were transmitted, after which a new physical mixture was produced at the destination.
The process exposes a fundamental limitation. A screen can generate millions of colours using three types of light emitter. A scent machine needs cartridges of physical chemicals, and no compact, universally accepted palette can reproduce every possible odour.
For now, sending a smell is closer to transmitting instructions to a sophisticated mixing machine than streaming music to a speaker.
Electronic noses learn to recognise odours
Other companies are concentrating on recognition rather than reproduction.
French firm Aryballe combines peptide-based sensors with silicon photonics. When volatile compounds interact with the sensor array, the device records a characteristic response. Machine-learning software can then compare that pattern with a library of known samples.
Its NeOse systems are aimed at uses such as product development, quality control and monitoring variations between batches. They do not identify every molecule in the air. Instead, they can be trained to recognise whether a sample resembles a previously recorded odour profile.
ScentifAI is pursuing a semiconductor-based approach. Its smell-imaging chips use multiple sensing membranes that react differently to airborne compounds, producing a pattern that software can classify. The company’s 2026 name change reflects its attempt to position this combination of sensors and AI as a broader platform rather than a single analytical instrument.
Meanwhile, French environmental technology company Ellona sells connected sensor systems for monitoring gases, particles and odour-related conditions around waste sites, ports, factories and other infrastructure. These networks can provide continuous measurements and help operators investigate when and where an emission occurred.
Similar technology could eventually detect food spoilage, monitor crop health or flag a chemical leak before it becomes obvious to people nearby. Medical applications are also under investigation because disease can alter the volatile compounds in breath, urine or sweat. Electronic noses proposed for diagnosis, however, still face demanding clinical-validation and regulatory requirements.
The future of AI and digital smell
AI has not given computers a human sense of smell. Today’s systems operate within constrained settings: predicting broad descriptions for individual molecules, comparing samples with a trained reference library or reconstructing selected aromas from chemical analysis.
Yet the components of a digital smell ecosystem are coming together. Molecular models can connect chemical structure with human descriptions. Sensor arrays can capture repeatable patterns from the air. Generative tools can search for new scent molecules, while automated dispensers can turn a digital formula back into a physical blend.
The progress is unlikely to produce a general-purpose “smell camera” soon. Scent remains chemical, contextual and personal in ways that images and sound are not. But machines do not need to experience an odour exactly as people do to make the information useful.
If a system can detect contamination, maintain the aroma of a product or identify a promising repellent faster than existing methods, it has already begun to turn smell into something computers can work with.