Resolving NO₂, O₃ and $\mathrm {H}_{2} \mathrm{O}$ Concentrations in Gas Mixtures with Dual CNT-FET Sensors and Machine Learning

In this work, we present a dual suspended carbon nanotube field-effect transistor (CNT-FET) sensor system, with a power consumption ~500 nW, for selective gas quantification in ternary gas mixtures. The CNT-FETs were evaluated in synthetic air under varying concentrations of NO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> and O<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> (0–50 ppb), and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm {H}_{2} \mathrm{O}$</tex> (0-60% RH). To extract relevant information from the device transfer characteristics, we derived four typical transistor features per device: threshold voltage <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathrm{V}_\mathrm{T})$</tex>, ON-state current <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathrm{I}_{\text{ON}})$</tex>, ON-state transconductance <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathrm{g}_{\text{ON}})$</tex> and subthreshold swing (SS). These features were used as input to machine learning non-linear regression algorithms to predict gas concentrations. Our results show that combining all four features from both CNT-FETs yields the highest prediction accuracy. Specifically, we achieved a mean absolute error (MAE) of ~6.8 ppb and a resolution of ~12.9 ppb at 0 ppb for NO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>, and an MAE of ~7.0 ppb and a resolution of ~15.0 ppb at 0 ppb for O<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf>.