Can We Find an Odor Blend in a Single Vessel?

Liquids are mixed little by little to approach a target odor. If the blend does not work, a different ratio is tried. When automating this search, must each attempt require a new container and fresh materials?

In 2024, Yota Fukui and colleagues reported a system that approaches a target sensor response by adding liquids to a single vessel and removing some of the mixture as needed. Alongside automating the search for a blend, they aimed to reduce sample consumption, control liquid volume, and reduce the effort of replacing vessels.

Adding alone makes it hard to reverse a blend

Continuing to mix in one vessel avoids replacing it. But if liquid is only added, the volume keeps increasing. When too much of one ingredient has been added, substantially changing its proportion may require adding large amounts of the others.

Changing liquid volume also matters for measurement. The device measures vapor in the space above the liquid. Large changes in liquid volume can alter gas concentrations in that space and the sensor response.

The researchers therefore allowed the system to choose not only what to add but also how much mixture to drain. Removing part of a well-mixed liquid does not itself change its proportions. Adding another ingredient afterward, however, changes the blend without continually increasing total volume.

Connecting measurement, prediction, and mixing

The apparatus included pumps for adding three liquids and another for removing the mixture. While stirring the vessel and controlling temperature and gas flow, it measured the vapor with a 12-channel sensor array.

The sensors convert tiny changes in force, caused when surface materials take up molecules, into electrical signals. Materials with different responses provide a combination of signals.

First, the system measures vapor from the target liquid. It then measures the current mixture and quantifies how far its response differs from the target. A computer uses that result to decide which liquid to add next and how much mixture to remove.

Bayesian optimization chooses the next blend to try

The researchers used Bayesian optimization to select blends. Based on previous trials, it predicts the performance of untested blends and takes uncertainty into account when choosing the next candidate to measure.

Instead of trying every combination in turn, the method seeks candidates likely to approach the target or yield useful new information. The researchers adapted it to the constraints of adding and draining liquid in a single vessel.

What it learns here is not the probability that a person will say two odors smell alike. Its objective is to reduce differences between the sensor signals of the target and the mixture.

Reselecting useful sensors during the search

All 12 channels need not be equally informative for every ingredient. Including responses with little useful information could interfere with the search.

The researchers therefore incorporated a method that, after each measurement, selects four channels whose accumulated data best support prediction of the mixing ratio. Four features are extracted from each channel, describing the signal’s rise, settling, recovery, and magnitude.

Differences from the target are evaluated across the 16 features from the selected four channels. The selected channels can change as the search accumulates data. They are chosen for their ability to reveal the proportions of the ingredients being mixed, not for similarity to human olfaction.

Comparing methods on 100 computational targets

Before testing the apparatus, the researchers compared search methods on a virtual blending problem whose performance was determined by a known mathematical function. They randomly generated 100 target ratios of three ingredients and tested approaches including emptying the vessel each time, adding only, and allowing drainage.

In this computational comparison, the method allowing drainage reduced total liquid use while maintaining search performance broadly comparable to preparing a fresh blend each time. It also controlled the amount in the vessel and performed more favorably than adding alone.

The number 100 does not mean that 100 real odors were reproduced. It refers to a comparison of search progress and material consumption for 100 mathematical targets.

Approaching the ratios of three liquids in the apparatus

The researchers then used 1-octanol, isopropyl alcohol, and methanol, with a mixture at a predetermined ratio as the target. Starting from a different initial blend, the apparatus repeated measurement, addition, and drainage.

For one target, the final mixture approached the correct proportions, and the selected channels’ signals closely reproduced the target. The researchers also confirmed that the system could approach the correct blend for another target ratio.

This was a demonstration with three known ingredients. It did not show that the apparatus can automatically identify every ingredient or completely recover the formula of an unknown perfume. These experiments were also conducted for research, not as instructions for blending fragrances at home.

Mirin and ponzu reveal a difference between signals and perception

The researchers also tried to approach the sensor signals of mirin and ponzu using pure water, cooking sake, and fish sauce as ingredients. For both targets, they produced signals resembling the responses of the selected sensors.

Cooking sake became the main ingredient in the blend targeting mirin. The authors explain that this is consistent with mirin containing ethanol and sugars and sharing odor characteristics with cooking sake.

For the blend targeting ponzu, however, they reported that despite similar sensor signals, the mixture lacked the sour odor associated with vinegar or citrus. The paper does not specify the number of evaluators or a formal sensory-test procedure for this assessment. Nevertheless, the authors identify the failure to fully reproduce the human-perceived odor of ponzu as an important limitation.

The sensors’ surface materials are not the same as human olfactory receptors. This example illustrates that a mixture judged similar by a machine need not smell the same to a person.

Bringing sample-efficient blending closer to perception

The researchers hope to apply the approach to perfume and food-aroma blending. Achieving this will require not only automated mixing but also improvements on the measurement side, such as sensor materials with properties closer to human olfactory receptors, to improve correspondence with perception.

The system of repeated addition and drainage in a single vessel is not limited to these sensors. The authors consider it applicable to alternatives such as metal-oxide or quartz-crystal sensors. Their vision is a basis for finding blends with less sample material, paired with measurements suited to the objective.

The research received support including Japanese Grants-in-Aid for Scientific Research. The authors declared no conflicts of interest.

Conclusion

Incorporating removal as well as addition into the search enabled automated blending that approached a target sensor response while controlling liquid volume. Computational comparisons also showed reduced sample consumption relative to preparing a fresh blend each time.

The study’s contribution is to connect a search method with a physical apparatus and address practical problems of vessel replacement and material consumption. To approach human-perceived odors, however, the measurement of “similarity” must improve too. Connecting efficient mixing with measurements that reflect perception is the next challenge.

References

Fukui, Y., Minami, K., Shiba, K., Yoshikawa, G., Tsuda, K., & Tamura, R. (2024). Automated odor-blending with one-pot Bayesian optimization. Digital Discovery, 3, 969–976. DOI: 10.1039/d3dd00215b. Full text. © 2024 The Author(s). Published by the Royal Society of Chemistry. The original article is available under CC BY 3.0.