Skoltech researchers have designed and tested an easy-to-manufacture semiconductor-based gas sensor that is inexpensive to make and operate. Within 40 seconds, it can detect harmful pollutants and other substances of interest in the air. Published in Sensors and Actuators B: Chemical, the study demonstrates 90% accuracy in detecting three hazardous gases: the highly flammable acetone and the toxic hydrogen sulfide and nitrogen dioxide. That said, the device will also respond to other compounds, and subsequent measurements improve the accuracy. The distinctive feature of the new detector is the use of carbon nanotube “fabric” suspended between two electrodes as the sensitive element. This is the first nanotube-based sensor of such design.

Sensors of this kind have applications that go beyond air quality monitoring at industrial facilities or in areas with heavy traffic. By analyzing gas concentrations in a greenhouse or an industrial refrigerator, one can tell when vegetables and fruits are ripe and whether meat has started to go bad. And acetone content in the air exhaled by patients with Type 1 diabetes is a good indicator of when the next insulin injection is due. 

Gas sensors based on different technologies tend to serve distinct applications. Semiconductor sensors such as the one created at Skoltech do well where constant monitoring is required. They are good for frequent measurements because of the low power consumption and cost per sensor and per measurement. That said, they are also very sensitive, but this comes with a problem: These devices respond to the presence of all sorts of compounds in the air, so scientists have to find ways to make them specific to whichever gases are of interest.

The primary approach to boosting semiconductor sensor selectivity involves mimicking the way the mammalian sense of smell works.

“The sensitive material, which in our case is a piece of carbon nanotube ‘fabric,’ has certain electrical resistance that varies in response to the presence of a whole lot of gas species. But the effect on resistance is different depending on the particular compound and concentration, and this is the key to selectivity,” says the study’s lead author, Skoltech research intern Konstantin Zamansky, who is also a doctoral student of the Institute’s Life Sciences program. “Similarly, human olfaction relies on epithelial receptors responding to all sorts of compounds. Yet distinct groups of receptors respond to each odor in their own characteristic ways. Based on the activation of the associated neurons, the brain can tell which odor the nose is dealing with.”

In a similar fashion, if many semiconductor sensors are used simultaneously, each of them can play the role of a “receptor,” and in the presence of a gas, the entire system of sensors (the “nose”) will generate a multidimensional response. That signal can then be investigated for signs of a certain pattern (“smell”) by means of machine learning methods (the “brain”). “It’s just that we work in a slightly different manner and generate the multidimensional signal not with many sensors but with one and the same sensor at different temperatures,” Zamansky added.

The sensor design proposed by the research team incorporates a piece of single-walled carbon nanotube fabric, which serves as the sensitive material, suspended in the air between two gold electrodes sputtered onto a polycrystalline aluminum oxide substrate. Such a solution enables nearly instantaneous cooling and heating of the sensor to requisite temperature. By virtue of that, the device can measure the nanotubes’ electrical resistance at 400 distinct temperatures between 25 and 125 degrees Celsius within merely 40 seconds. The resulting 400 values constitute the multidimensional sensor response pattern.

A machine learning model attributes that sensor response pattern to one of the target gases or air with an average accuracy of 90%. Each consecutive 40-second cycle reduces the probability of error roughly by a factor of 10, bringing the confidence in the result to about 99.9% (this is a somewhat simplified calculation) in just two minutes. Importantly, the cost of one such measurement for nitrogen dioxide detection, for example, is lower than with other sensors.

By filtering the data (to establish a baseline), the creators of the sensor were able not just to interpret the signal but also to account for the so-called device aging effect. This refers to the sensor altering its response to the same gases under unchanged conditions after prolonged operation. Introducing a correction for sensor aging enabled the team to cut the error rate by half, bringing the accuracy from 80 to 90%. This problem is relevant to all semiconductor sensors, yet it is scarcely accounted for. The Skoltech team successfully resolved it for an interval of at least 10 hours of continuous operation.

The study reported in this story was supported by Russian Science Foundation Grant No. 20-73-10256.