AI Smell Research: Molecules, Mixtures and Sensors

A headline saying that AI can smell invites a wonderfully simple picture: a machine encounters a scent and recognises it. The research is more varied, and more interesting. Some systems begin with a molecule’s structure. Others work with descriptions of blends. Others receive recordings from physical sensors.

Those starting points matter. A fragrance maker considering new ingredients, a business checking a delivered sample and a researcher studying human perception need different evidence. Understanding the input and the intended output makes the progress easier to appreciate—and the practical claims easier to assess.

Predicting how a molecule might smell

In a Science paper published online on 31 August 2023, Brian Lee, Emily Mayhew and colleagues introduced a principal odour map. Their graph neural network linked molecular structure to descriptions of odour quality.

The team characterised 400 new odorants prospectively; after exclusions for issues including low intensity and contamination, 323 remained in the performance analysis. The model’s profile more closely matched the panel mean than the median panellist’s profile did for 53% of those molecules. This measures agreement with a trained group’s description, rather than superiority on every smell-related task.

Here, the useful starting information is a molecular structure. The result concerns how an odorant is described. Whether a customer would enjoy a finished product is a further question. Treating those outcomes separately preserves the meaning of the achievement.

Blends introduce another prediction problem

Laura Sisson and colleagues’ ACS Omega study, published on 3 March 2025, addressed pairs of aroma chemicals. They assembled molecular structures and odour labels from the GoodScents repository, then developed graph neural network models to predict blend descriptors.

The dataset had a consequential limitation: relative concentrations in the blends were unavailable. Essential oils and flavour extracts were also filtered out. Those details define the material behind the reported results; they are not a verdict on everything the wider field might achieve.

My practical reading is that a successful prediction for a labelled pair should lead to a carefully scoped next experiment. It does not, by itself, establish how an arbitrary finished fragrance will be perceived at a chosen recipe. Ingredient identity, formulation and the intended judgement need to stay visible in the claim.

Reading physical samples through sensors

SmellNet, published in the ICLR 2026 proceedings, starts from physical measurements. Dewei Feng and colleagues collected gas-sensor time series covering 50 substances, including herbs, spices, fruits, vegetables and nuts, plus 43 controlled mixtures. Their ScentFormer model learns from those changing sensor responses.

The authors’ accompanying repository distinguishes recognition of a single substance from predicting mixture proportions. It also separates known mixtures recorded in new sessions from previously unseen combinations. The reported mixture results were substantially weaker for unseen combinations.

This makes generalisation a concrete issue. A demonstration can show that a recording has been classified within a defined task. To use that result elsewhere, ask which substances and combinations were tested, how the samples were recorded and what counted as a correct answer.

Let the business question choose the demonstration

Imagine an independent fragrance studio assessing a proposed AI service. This is an illustrative scenario, not a report about an existing company. The studio has three requests:

• Explore an ingredient: suggest likely descriptive qualities for a candidate whose chemical identity is already known.
• Check a delivery: help determine whether a physical sample matches the material the studio expected to receive.
• Choose a collection: establish which of three finished scents the intended customers prefer.

A supplier’s demonstration might be relevant to one request and leave the others unanswered. The studio should ask for the evidence separately. A persuasive description of an ingredient is useful for discussion, but it cannot stand in for a demonstrated delivery-checking process or customer-preference study.

Suppose the business’s immediate problem is investigating inconsistent batches. It could write a brief asking a proposed system to distinguish approved reference samples from specified alternatives, under agreed sampling conditions. That is a clearer request than buying a general promise to understand fragrance.

Before any pilot, decide who will establish the reference identity, which cases will be withheld from development and what happens when the output is uncertain. These are proposed evaluation choices for the studio, not procedures claimed to have been tested in the three papers.

Ask for an error you can interpret

An overall score becomes useful when you understand the errors behind it. For the hypothetical batch-checking task, ask the provider to show individual cases: a correct match, a wrong match, an expected material absent from the available choices and a sample outside the agreed conditions.

Define how those cases should be handled before viewing the results. If the proposed product can flag an unknown sample, test that behaviour explicitly. If it always selects a listed label, record that limitation rather than reading every label as confirmed identification.

Keep a separate note of the decision the result would support. Would it send a sample for further assessment, narrow a set of candidates or approve a delivery? The acceptable evidence depends on that role. A useful aid to investigation and an approval process can require different demonstrations.

For readers examining a paper themselves, MaryChuks.com’s guide to critical research appraisal offers a route through the question, methods, finding and boundary. In this field, add the input and output to that note.

Appreciate the progress at the right level

These studies show why a single phrase can conceal several kinds of technical progress. Molecular prediction, blend-description modelling and sensor-based recognition each address a defined part of the problem. Their results become more useful when the boundaries travel with them.

For creators and business owners, the first productive step is a clear sentence: what information will the system receive, what should it return, and which decision will that help us make? Once that sentence is settled, a research result can become the start of a credible application.

Sources

• Lee, Mayhew and colleagues: A principal odor map unifies diverse tasks in olfactory perception, Science, online 31 August 2023.
• Sisson and colleagues: Deep Learning for Odor Prediction on Aroma-Chemical Blends, ACS Omega, 3 March 2025.
• Feng and colleagues: SmellNet, ICLR 2026 proceedings; accompanying research code and benchmark descriptions, checked 24 October 2026.