Machine learning can now predict how some molecules may smell and help electronic noses recognize chemical patterns, opening possibilities in medicine, food safety and environmental monitoring.
THE UNIVERSAL RECORD
Sourced reporting. No opinions.
Brad Socha | September 11, 2026 | 9:15 AM EST
Artificial intelligence already processes language, images, video and sound. A less familiar frontier is AI built around another human sense: smell.
Computers do not smell the way people do. They have no nose, olfactory receptors or conscious experience of an aroma. Instead, researchers are combining machine learning with molecular data and chemical sensors to predict how molecules may smell or recognize patterns in volatile compounds in the air.
The distinction matters. Some systems begin with the molecular structure of a chemical and attempt to predict whether people will describe its odour as floral, fruity, woody or something else. Electronic noses take a different approach: arrays of physical sensors react to airborne chemicals, while algorithms interpret the resulting pattern.
Together, these technologies are beginning to give machines capabilities that resemble limited parts of biological olfaction, and could eventually make smell another form of information computers can analyze.
How AI Is Mapping Smell
Predicting an odour from a molecule is unusually difficult.
Colour can be represented using a relatively compact set of dimensions, and sound can be measured through properties such as frequency and amplitude. Human olfaction involves hundreds of receptor types, while molecules with apparently similar structures can sometimes smell very different.
A major advance came from researchers at Google Research, the Monell Chemical Senses Center and collaborating institutions. In research published in Science in 2023, the team described a machine-learning system called the Principal Odor Map.
The researchers represented molecules as graphs, with atoms and chemical bonds forming their underlying structures. A graph neural network then learned relationships between those structures and human descriptions of smell.
The system was tested prospectively on 400 previously uncharacterized odorants. A panel of 15 trained participants smelled the compounds and described their qualities. According to the study, the model’s predicted odour profiles were closer to the panel’s average ratings than the descriptions given by the median individual panelist.
That does not mean the AI itself experienced those smells.
It learned a mathematical relationship between molecular structure and human perception. Given a molecule it had not previously encountered, the system could predict how people were likely to describe its odour.
The Principal Odor Map also arranged molecules so that substances perceived as smelling similar were generally located closer together in its mathematical representation. Researchers found that this representation could be applied to several additional odour-prediction tasks.
The work points toward what researchers sometimes call the digitization of smell: developing computational representations that allow computers to predict, classify and eventually help reproduce odours.
But predicting smell from a chemical structure is only one side of the problem.
Electronic Noses Turn Chemicals Into Data
An electronic nose typically starts with physical sensors rather than a molecular formula.
When volatile organic compounds, or VOCs, encounter an array of sensors, they produce measurable responses. Different chemicals can generate different patterns across the array. Machine-learning algorithms can then be trained to recognize those patterns and classify what the sensors are detecting.
The objective is not necessarily to identify every individual molecule. Like biological smell, an electronic nose can use a pattern produced by a complex chemical mixture as a recognizable signature.
This approach has potential applications in food safety. A 2025 review in Food Research International examined machine-learning-assisted electronic noses and related technologies for detecting contamination and adulteration, including microbial contamination, pesticide residues and toxins. The authors found considerable potential but cautioned that many applications remain primarily laboratory-based and that a lack of standardized protocols limits wider deployment.
Another 2025 study combined gas sensors with a transformer-based machine-learning model to classify gas leaks and food spoilage. The experimental system achieved strong results on its test data and operated with low latency, although results from a controlled research system should not be interpreted as proof of equivalent performance across real-world environments.
Environmental monitoring presents another possibility. Electronic noses can detect patterns involving VOCs and hazardous gases, potentially providing portable systems for monitoring industrial facilities, indoor environments or air pollution. Reviews of the technology nevertheless identify persistent problems including sensor drift, sensitivity, selectivity and differences in how data are collected and processed.
Could Machines Smell Disease?
Perhaps the most consequential application involves human health.
Metabolic processes produce volatile chemicals that can leave the body in exhaled breath. Disease can alter these chemical patterns, raising the possibility that an electronic nose could recognize a distinctive “breathprint” without requiring researchers to identify every compound individually.
Cancer detection is one area receiving substantial attention.
A 2025 systematic review and meta-analysis examined 35 studies involving 4,483 patients in which electronic noses analyzed exhaled air for lung cancer. Across those studies, pooled sensitivity was 90% and pooled specificity was 89%. The researchers concluded that electronic noses show promise as non-invasive diagnostic tools but emphasized that further research and validation are necessary.
Research published in 2026 continues to identify significant barriers to clinical use, including sensor drift, cross-sensitivity, variations among patients, differences in breath collection and the need for sufficiently large training datasets.
Those limitations illustrate why saying that AI can “smell cancer” would go beyond the evidence. Researchers are investigating whether sensor systems can recognize chemical patterns associated with disease. That is not yet equivalent to a universally validated medical diagnostic test.
The same caution applies to the broader idea that AI has acquired a sense of smell. Today’s systems divide the problem into measurable tasks: sensing airborne chemicals, analyzing molecular structures, recognizing patterns and predicting human odour descriptions.
What is changing is the sophistication with which those pieces can be connected.
Sight and sound became extraordinarily useful forms of digital information because computers learned to represent, analyze and generate them. Researchers are now attempting something similar with a sense that has historically been much harder to quantify.
AI cannot smell as a human does. But it is getting better at extracting information from the chemistry behind smell, and that could eventually allow machines to detect things that human noses cannot reliably perceive.
Sources:
Science — A Principal Odor Map Unifies Diverse Tasks in Olfactory Perception
https://www.science.org/doi/10.1126/science.ade4401
PubMed — A Principal Odor Map Unifies Diverse Tasks in Olfactory Perception
https://pubmed.ncbi.nlm.nih.gov/37651511/
Nature Microsystems & Nanoengineering — Research Progress of Electronic Nose Technology in Exhaled Breath Disease Analysis
https://www.nature.com/articles/s41378-023-00594-0
PubMed — Progress in Machine Learning-Supported Electronic Nose and Hyperspectral Imaging Technologies for Food Safety Assessment
https://pubmed.ncbi.nlm.nih.gov/40253192/
Journal of Food Science — Advancing Electronic Noses With Transformers: Real-Time Classification of Hazardous Odors and Food Freshness
https://ift.onlinelibrary.wiley.com/doi/full/10.1111/1750-3841.70609
Trends in Analytical Chemistry — Electronic Nose for the Detection and Discrimination of Volatile Organic Compounds: Application, Challenges, and Perspectives
https://www.sciencedirect.com/science/article/pii/S0165993624004412
ERJ Open Research — The Electronic Nose in Lung Cancer Diagnostics: A Systematic Review and Meta-analysis
https://pubmed.ncbi.nlm.nih.gov/40391063/
ACS Sensors — Emerging Electronic Nose Design for Breath-Based Cancer Diagnostics: Advances in Machine Learning Approaches and Sensor Architecture Design
https://pubmed.ncbi.nlm.nih.gov/42003788/
About the Author
Brad Socha is the founder of The Universal Record, focused on sourced, factual global reporting. Coverage includes international news, geopolitics, technology, and major developments.