What Can an Electronic Nose Read from Response Waveforms?

Even when a sensor responds to two gases with a similar magnitude, the way that response changes need not be the same. Does it rise abruptly or gradually? How does it recover when returned to air? Could these waveforms retain information that helps distinguish gases?

In 2024, Seung-Hyun Sung and colleagues reported an artificial olfactory system that emphasizes curves changing over time, rather than response magnitude alone. Their aim was not simply to improve AI analysis, but to design the entire process together, from sensors that produce reproducible signals to methods for interpreting the data.

A large response is not the same as discrimination

The sensors used in this study change their electrical resistance when gases react at their surfaces. A large gas response can make small changes easier to detect. But if two gases produce responses of similar magnitude, that number alone makes them difficult to distinguish.

In mixed gases, components interact at the sensor surface, so responses cannot necessarily be understood by adding together the responses to individual gases. The authors focused on the possibility that reducing a curve to a few numbers, such as its maximum, discards information.

Nine different responses on one chip

The researchers built nine sensing elements based on thin, rod-like tin oxide structures, varying the combinations of metals and other surface modifications. Changing the gas-sensitive material and surface morphology alters not only the magnitude of resistance changes, but also the waveforms.

They also compared the thicknesses of the added surface layers. Excessively thick layers could raise resistance and destabilize signals, while very thin layers did not produce sufficiently distinct responses. The study selected a thickness of 1 nanometer for each metal layer.

They also adopted a method capable of producing multiple chips through the same fabrication process and standardized conditions such as temperature and humidity in the measurement system. If signals varied greatly between measurements before reaching the AI, learning the curves’ features would be difficult.

Three gases that overlapped in response magnitude

The researchers first examined six gases: acetone, toluene, xylene, ethanol, ammonia, and hydrogen sulfide. Summarizing the response magnitudes of the nine sensing elements separated ethanol, ammonia, and hydrogen sulfide relatively well, but acetone, toluene, and xylene overlapped.

They therefore analyzed waveforms from these three difficult-to-distinguish gases and mixtures containing them in different proportions. Measurements at different concentrations also examined whether characteristic curves associated with each material–gas combination appeared repeatedly.

The authors called these characteristic response curves “eigengraphs.” Here, it is helpful to think of an eigengraph as a distinctive response curve produced by a particular sensing material and gas combination. It is not a universal signature belonging to the gas alone, but a signal arising in relation to sensor materials and measurement conditions.

Organizing electrical signals with a method from speech analysis

Instead of treating each curve as a long list of raw numbers, the researchers adapted a method used in audio processing called MFCC. This technique summarizes features of a time-varying signal in a relatively small number of values.

First, changes in a curve are represented as a combination of components fluctuating at different rates. These features are then compressed and supplied to the AI. The idea is that even curves reaching the same height may contain useful differences in their rise and other characteristics.

Although the method originated in audio analysis, the sensors were not listening to sounds. Their input was the change in electrical resistance over time during exposure to gas. Nor was this an experiment reproducing human hearing or smell itself.

117 classes do not mean 117 different odors

At this stage of the analysis, the researchers used 117 labels: 13 gas conditions, including individual gases and mixture ratios, multiplied by nine sensing elements. In other words, they did not prepare 117 different odors. They classified combinations of which gas condition was measured by which sensing element.

They compared the use of all 342 points in a curve with the first 171 points and with just the first 50 points, covering the initial reaction to the gas. Even these shorter sections retained features useful for classification in this dataset.

Accuracy on the training data reached the chosen targets of 99.5% or 99.9%, and restricting input to 50 points also shortened training time. These figures describe performance during training, however. They do not by themselves show that newly measured gases or unknown mixtures can be identified with the same accuracy.

Separate validation using exhaust-like standard gases

As an application to more complex combinations, the researchers also examined two standard gas mixtures simulating emissions from gasoline and diesel vehicles. They used certified standard gases with specified compositions, rather than collecting exhaust from actual vehicles.

Adding the individual gases carbon monoxide, carbon dioxide, nitric oxide, and nitrogen dioxide produced a six-class experiment. Measurements were made under controlled conditions, with gases diluted in dry air and sensors operated at 300°C.

This experiment used four chips made through the same process, recording four responses from each. Unlike the initial 117-label analysis, data from the nine sensing elements were combined for each gas condition, giving six classes.

The researchers also used four-fold cross-validation, dividing the data into four groups and rotating which groups were used for training and evaluation. Classification accuracies for three AI models were 96.1%, 99.8%, and 100%. A separate validation comparing four gas mixtures also produced high classification performance.

These results show that differences between the selected standard gases could be distinguished using the waveforms and analysis methods. They do not establish equivalent classification of real roadside exhaust with varying humidity, temperature, and composition, or the ability to quantify every component in an unknown mixture.

Good AI is not enough

A major implication highlighted by the authors is that the apparatus producing the data should not be separated from the method interpreting them. Accumulating meaningful features first requires reliably similar waveforms from the same material–gas combination.

The paper broadly proposes the existence of waveforms characteristic of material–gas interactions. Results for the materials and gases tested here, however, do not establish a universal law explaining all chemical reactions or unknown odors.

The authors therefore identify standardization of sensors and measurement conditions, along with data centers collecting waveforms from different gases and environments, as next steps. Their vision is that expanding the range of materials and conditions and enabling reference to known responses will provide a future foundation for handling unknown signals.

The study is relevant to interest in complex mixtures such as perfume, but it did not measure people’s scent impressions, pleasantness, or perceived intensity. The relationship between chemical responses distinguished by an “electronic nose” and human perception must be investigated separately.

The work received support from the National Research Foundation of Korea and public research funding in industrial and environmental fields, among other sources. The authors declared no competing interests.

Summary

Combining nine sensors with different responses and waveform analysis derived from audio processing extracted gas characteristics that were difficult to distinguish from response magnitude alone. Separate cross-validation using standard gas mixtures also achieved high classification performance within the selected conditions.

The study points toward an approach beyond pursuing more sensitive sensors or more complex AI in isolation. Material–gas interactions need to be recorded as stable waveforms and developed into comparable data. The authors propose that jointly designing fabrication, measurement, analysis, and data sharing will advance artificial olfaction.

References

Sung, S.-H., Suh, J. M., Hwang, Y. J., Jang, H. W., Park, J. G., & Jun, S. C.(2024). Data-centric artificial olfactory system based on the eigengraph. Nature Communications, 15, 1211. DOI: 10.1038/s41467-024-45430-9. Full original article. © The Author(s) 2024. The original article is published under CC BY 4.0.