Selective Detection of Toxic Gases by Arrays of Single-Layer Graphene Sensors Functionalized with Nanolayers of Different Oxides †

: Graphene provides an ideal platform for chemiresistive gas sensors as the material is fully exposed to the surrounding environment. For practical use in an ambient atmosphere, its sensitivity and selectivity should be evoked by functionalization by defects and dopants or by decoration with nanophases of metals or metal oxides. Here, we demonstrate a few successful cases of selectivity enhancement by functionalizing the graphene with different oxide layers and applying machine learning to the resulting sensor array.


Introduction
Chemiresistive graphene gas sensors are appealing for e-nose applications where their easy production and miniaturization potential can be exploited.Relevant use cases include monitoring indoor and outdoor air quality, performing medical self-diagnosis by breath analysis, and controlling industrial processes.However, for effective use in practical sensors, the sensing properties of graphene have to be improved and controllably modified.We have shown nearly a 100 times enhancement of graphene's gas sensitivity as well as improved selectivity through functionalization by pulsed laser deposition [1,2].In this work, we demonstrate the modification of graphene with different metal oxides to controllably induce partial selectivity toward the harmful gases NH3, H2S, NO2, and O3.Selected combinations of these sensors were integrated into arrays, whereby machine learning applied to the output signal pattern of each array allowed the successful differentiation of various gases and their mixtures.

Materials and Methods
Chemically vapor-deposited single-layer graphene was transferred onto Si/SiO2 electrode substrates (inset in Figure 1) or special CMOS sensor substrates with built-in microheaters.A KrF excimer laser was used to deposit thin oxide layers on top of the graphene from respective ceramic targets [1].The substrate temperature, background gas type and pressure, and the thickness of the functionalizing layer were optimized for the best performance of the sensors.

Discussion
Figure 1a shows the responses of the individual sensors in an array for the simultaneous detection of NO2 and NH3 gases.The functionalizing materials were chosen to achieve partial selectivity of the individual sensor, either toward NO2 (TiN, HfO2) or NH3 (V2O5, SnO2).For machine learning studies, an extended (>50 h) gas sensing experiment was carried out, where NO2 and NH3 were simultaneously present and their concentrations in synthetic air were randomly (but gradually) changing within 25-390 ppb and 4.5-80 ppm, respectively.Simple artificial neural networks (containing a few neurons in a single hidden layer) applied to the sensor array containing three of the most stable sensors could reliably distinguish and quantify NO2 and NH3 in a mixture over a period of tens of hours, maintaining the mean relative error ≤14%.
Figure 1b shows the responses of the individual sensors in another array tailored for differentiation between O3 and NO2 gases.Although all the tested materials showed selectivity toward O3, functionalizing materials were selected to exhibit different response ratios to O3 and NO2 gases.During 72 h, the sensors were exposed to randomly generated concentration cycles of 30 ppb NO2, 30 ppb O3, 60 ppb NO2, 60 ppb O3, and 30 ppb NO2 + 30 ppb O3 in synthetic air.Various properties of the dynamic responses (amplitude, response rate, and recovery rate) were considered features for machine learning, enabling clearly distinguishing these five gas compositions with an accuracy of ~94%.

Discussion
Figure 1a shows the responses of the individual sensors in an array for the simultaneous detection of NO 2 and NH 3 gases.The functionalizing materials were chosen to achieve partial selectivity of the individual sensor, either toward NO 2 (TiN, HfO 2 ) or NH 3 (V 2 O 5 , SnO 2 ).For machine learning studies, an extended (>50 h) gas sensing experiment was carried out, where NO2 and NH3 were simultaneously present and their concentrations in synthetic air were randomly (but gradually) changing within 25-390 ppb and 4.5-80 ppm, respectively.Simple artificial neural networks (containing a few neurons in a single hidden layer) applied to the sensor array containing three of the most stable sensors could reliably distinguish and quantify NO2 and NH3 in a mixture over a period of tens of hours, maintaining the mean relative error ≤14%.
Figure 1b shows the responses of the individual sensors in another array tailored for differentiation between O 3 and NO 2 gases.Although all the tested materials showed selectivity toward O 3 , functionalizing materials were selected to exhibit different response ratios to O 3 and NO 2 gases.During 72 h, the sensors were exposed to randomly generated concentration cycles of 30 ppb NO 2 , 30 ppb O 3 , 60 ppb NO 2 , 60 ppb O 3 , and 30 ppb NO 2 + 30 ppb O 3 in synthetic air.Various properties of the dynamic responses (amplitude, response rate, and recovery rate) were considered features for machine learning, enabling clearly distinguishing these five gas compositions with an accuracy of ~94%.

Figure 1 .
Figure 1.(a) Responses of an array (inset) of graphene sensors functionalized with different PLD films for differentiating NH3 and NO2 gases.Measured at RT under UV light excitation.(b) Responses of an array of graphene sensors functionalized with different PLD films for differentiating O3 and NO2 gases.The inset shows clustering of the data points in the 2D feature space of the response amplitudes.

Figure 1 .
Figure 1.(a) Responses of an array (inset) of graphene sensors functionalized with different PLD films for differentiating NH3 and NO2 gases.Measured at RT under UV light excitation.(b) Responses of an array of graphene sensors functionalized with different PLD films for differentiating O3 and NO2 gases.The inset shows clustering of the data points in the 2D feature space of the response amplitudes.