Can You Ask AI to Make a Scent More Floral?

I want to add a floral quality while keeping the character of this scent. Could a computer turn that kind of perfumery request into a formula? One study tested a method that goes a step beyond reproducing an existing scent to adding a specific desired feature.

In a study published in 2024, Aleixandre and colleagues used an AI trained on essential-oil mass spectra and odor descriptors to search for new formulas. They did not stop at computation: they mixed actual liquids and conducted tests in which people smelled them. Success varied depending on the feature they wanted to add.

Linking essential-oil analysis to the language of scent

The starting point was analytical data collected for 94 essential oils and 39 odor descriptors. These included floral, sweet, fresh, and woody. Whether a descriptor applied to an oil was recorded as 1 or 0. These were not scores rating how strong each quality was.

A mass spectrum is an analytical pattern showing the amount of signal at different mass-to-charge ratios after molecules have been ionized. The researchers trained a neural network to predict odor descriptors from this pattern. Essential oils contain many molecules, but the method does not identify every constituent individually before making a prediction.

When the trained predictor was tested by leaving out one essential oil at a time, its overall balanced accuracy was about 0.74. This measure evaluates cases where a descriptor applies and cases where it does not without favoring either group. It does not mean that the system can make the desired perfume 74% of the time.

Using 20 blended materials as the basis for formulas

In a separate preparatory step, the mass spectra of 180 essential oils were organized so that they could be approximated by combinations of 20 basic blended materials. These 180 oils included the 94 with odor descriptors. The number of oils used to train the predictor therefore differed from the number used to establish the basis for the formulas.

The 20 materials were not simply 20 fragrance molecules or a selection of 20 essential oils. Each was a mixture of several essential oils. Rather than directly adjusting a large number of oils, the researchers varied the proportions of these 20 materials to approach the target analytical pattern.

The number 20 was chosen partly with a view to eventual use in a scent-delivery device with 20 channels. In this study, the sensory tests used liquids prepared according to the calculated formulas. The researchers did not demonstrate a device automatically releasing scents to accompany images.

Working backward from an AI prediction to a formula

The formula search begins with equal proportions of the 20 materials. The mass spectrum of that mixture is calculated, and the AI predicts its odor descriptors. The mixing ratios are then gradually changed in a direction that reduces the difference from the target combination of descriptors.

What changes here is the formula. The AI itself is not trained again from scratch for every search. A predictor that links an analytical pattern to a likely scent is used in reverse to search for a formula likely to produce the desired scent.

The method assumes that a mixture’s mass spectrum can be obtained by adding the patterns of its ingredients in proportion to their amounts. It does not prove that perceived scent itself is always a simple sum. Even if the computation approaches the target, whether people actually perceive it that way needs to be tested separately.

Can people identify a sample with added floral character?

The sensory tests used four essential oils as starting points: chamomile, citrus, cedarwood, and oregano. The same search method was used for a replicated formula targeting the original oil’s characteristics and a modified formula with one additional descriptor.

The four features to be added were floral, sweet, fresh, and woody. However, the researchers did not test every possible combination of oil and feature. They selected eight combinations, taking into account the predictor’s performance and how closely the computation converged on the target.

They recruited 26 untrained participants aged 21–64. Samples were presented in randomized order, and neither the evaluators nor the people administering the tests knew which was the modified version. In the first test, participants had to choose which of two samples they thought had the specified added feature.

For floral, the modified version was selected in 21 of 30 pooled responses, a difference above chance in the two-choice test. For sweet, 25 of 38 responses selected the modified version, and for fresh, 22 of 38 did so; neither met the statistical criterion. For woody, 13 of 30 responses selected it, so the choices did not favor the expected direction.

These are pooled counts of sample choices, not numbers of people. Although the paper reports recruiting 26 participants, its condition-level tables contain 15 or 19 responses, and the breakdown of whether everyone completed every condition is unclear.

Across all conditions, 81 of 136 responses were in the expected direction, a result above chance. That aggregate does not mean that all four features could be added equally successfully. It is important to look at the results for each feature.

Which formula smells closer to the original?

In the second test, participants first smelled the actual essential oil, then chose whether the replicated or modified formula was closer to it. The replicated formula was selected as closer in 92 of 120 pooled responses.

This supports the idea that people can distinguish between a formula targeting the original characteristics and one changed by adding a feature. It does not prove that the replicated formula smelled exactly like the actual oil. The test asked which of two alternatives was closer to the reference.

Formula quality depends on predictor quality

Floral also performed relatively well at the stage of predicting descriptors from analytical patterns, whereas woody performed less well. The authors interpret this as meaning that, because the formula search relies on the predictor, features that are difficult to predict are also difficult to add deliberately.

Suggested improvements include increasing the amount of essential-oil and evaluation data used for training and improving the neural network’s design. This study used a limited dataset of 94 oils, and the network settings were not fully optimized.

The authors also propose collecting users’ responses, such as satisfaction with a scent, to update the predictor. Their longer-term vision is to connect this formula search to a multichannel scent-delivery device and create scents suited to visual media or entertainment settings. The work is better understood as a proof of principle toward that goal than as a finished general-purpose perfumery service.

Summary

An AI that predicts odor descriptors from analytical data can be used to work backward toward a formula. Tests in which people smelled actual essential-oil mixtures supported the addition of floral character, but did not establish the same success for sweet, fresh, or woody qualities.

The advance here is a step beyond imitating existing scents toward designing them with a particular added feature. The study points to a way of combining better prediction data with human evaluation to bring computational targets closer to what people actually smell.

References

Aleixandre, M., Prasetyawan, D., & Nakamoto, T. (2024). Automatic scent creation by cheminformatics method. Scientific Reports, 14, 31284. DOI: 10.1038/s41598-024-82654-7 / Full text.Original article: © The Author(s) 2024, CC BY 4.0.