Fast Nitrogen Dioxide Sensing with Ultralow‐Power Nanotube Gas Sensors

The study reports fast, ultralow‐power operation of carbon nanotube‐based nitrogen dioxide (NO2) sensors enabled by nanotube self‐heating and transient sensing. The self‐heating effect in the nanotube channel significantly accelerates the desorption of gas molecules, reducing the sensor recovery time to a minute. As gas molecules re‐adsorb on the nanotube after cooling, the initial rate of the sensor transient is used to determine NO2 concentration within a few minutes. To accelerate and optimize the operation of the sensor, the study considered temperature profiles along the self‐heated carbon nanotube, their effect on different sensing regions, and a physical model‐based fit. As a result, the nanotube‐based NO2 sensor demonstrates recovery and readout times below 5 min and an extrapolated limit of detection below 10 ppb. The peak power consumption of this operation mode is below 6 µW. The combination of fast readout, fast recovery, low limit of detection, and ultralow power consumption demonstrated in this work shows strong promise of carbon nanotube‐based NO2 sensors in mobile or Internet‐of‐Things (IoT) applications.


Introduction
Nitrogen dioxide (NO 2 ) is a toxic gas that poses a threat to human health at low concentrations.Monitoring NO 2 levels is thus important for ensuring compliance with environmental regulations and protecting public health.The critical concentration levels defined for NO 2 by governmental organizations are in the range of a few tens of parts per billion (ppb). [1]Therefore, the monitoring of NO 2 requires sensing systems with a very low limit of detection (LOD).In addition, due to the prevalence of combustion processes-a major source of NO 2 -in various industries, [2] there is an increasing need to monitor this gas with mobile systems.Current mainstream technologies such as electrochemical DOI: 10.1002/adsr.202300081 and semiconductor sensors, however, have limitations in mobile applications.5] In the past decades, several NO 2 sensors based on nanostructures including metal oxide (MOX) nanowires/nanoparticles, [6][7][8] lowdimensional materials, [9][10][11] and metalorganic frameworks (MOFs) [12] have demonstrated great potential as an alternative, especially in terms of LOD and power consumption.However, most of the studies in the literature do not report sufficiently short readout and recovery times, which should be on the order of a few minutes for practical NO 2 monitoring in the field. [13]Readout schemes that rely on steady-state signals usually require tens of minutes or even a few hours. [8,9][12] However, the conductance is still changing rapidly at this point in time, so it is questionable if a response defined this way can translate into the readout of the analyte concentration in real-life situations, in which NO 2 concentration varies unpredictably.
Carbon nanotube (CNT)-based NO 2 sensors have also suffered from slow readout and recovery.Although they function as highly sensitive, low-power NO 2 sensors at room temperature, [14] the conventional steady-state sensing scheme requires a long readout time (>2 h) [15] and unassisted recovery at room temperature takes even longer (>12 h). [16]ne effective way to shorten the recovery time while maintaining CNT's low power consumption is to accelerate the desorption of gas molecules by using the self-heating effect in a suspended CNT (Figure 1a). [16]When a nanotube is attached to a planar substrate, the heat dissipation into the substrate hinders the nanotube from heating up.In contrast, when an individual nanotube is suspended between source and drain contacts, it behaves like an ideal heating filament that can be heated up significantly at only a few μW, [17] as there is almost no heat dissipation from the nanotube into the substrate.
Resetting of the sensor surface via self-heating can also be used as an opportunity to shorten the readout time.As gas molecules ) obtained with a pulsed V gs sweep. [34]c) Output characteristic (I d -V sd ) of the device showing a peak I d at V sd of ≈0.8 V followed by a negative differential conductance (NDC) regime.The bias conditions of V sd = 1.5 V and V gs = −10 V (marked with red circle), which belong to the NDC regime, are used to induce the nanotube self-heating effect for accelerated sensor recovery in our gas measurements.
re-adsorb on the reset surface, the initial slope of the sensor transient can be read out to determine the analyte concentration.Such readout is intrinsically faster than steady-state sensing, which would have to wait until the sensor signal stabilizes. [15,18]revious work from the authors' research group presents early demonstrations of these two methods to report sensor signal recovery in 10-45 min and transient readout in 10-40 min. [15,16,19]lthough these are significant improvements, further reductions of recovery and readout times down to a few minutes would be required for practical applications.In case of initial slope sensing, simple linear fitting has been used to estimate an LOD of 23 ppb. [19]his work aims to advance the understanding of the physical phenomena during the self-heating-assisted desorption and subsequent re-adsorption-discussions of which are largely missing in the earlier studies-and exploit it to optimize the operation scheme of CNT-based NO 2 sensors.Our analysis of the changing temperature profile along the self-heated nanotube and its effect on different sensing regions identifies an optimal gate bias (V gs ) for a much faster recovery of device current (I d ) signal.Consequently, our gas measurements demonstrate reproducible sensor signal recovery in 1 min.When gas molecules return to the surface of the recovered nanotube after a self-heating pulse, we fit a physics-based nonlinear response function to the sensor transient to determine its initial response rate.This approach allows us to determine NO 2 concentration within 1-5 min.We periodically alternate between the "recovery mode" to reset the sensor and the "readout mode" to determine NO 2 concentration, completing a measurement cycle in a few minutes.The highest power consumption during this operation scheme is ≈5.6 μW, and the extrapolated LOD is 9-61 ppb.To the best of our knowledge, the short readout and recovery times below 5 min and an LOD below 10 ppb at an ultralow power (<10 μW) have not been reported elsewhere for a nanostructure-based NO 2 sensor.

Optimizing Gate Bias for Fast Sensor Signal Recovery
The gas sensor used in this work is a carbon nanotube fieldeffect transistor (CNTFET), in which an individual nanotube is suspended between source and drain contacts (Figure 1a).The device shows a hysteresis-free transfer characteristic (I d -V gs ) and low ON-resistance of 97 kΩ at V sd = 0.1 V, V gs = -10 V (Figure 1b), which indicates strong electrical contact between the nanotube and electrodes.The output characteristic (I d -V sd ) shows current peaking followed by a negative differential conductance (NDC) regime (Figure 1c), a sign of substantial Joule heating of the suspended CNT channel. [17]In our gas sensing measurements, we regularly apply biases in the NDC regime (V sd = 1.5 V, V gs = −10 V) to induce the self-heating effect in the suspended nanotube channel and thereby accelerate sensor recovery.
In previous work, the self-heating effect in a suspended nanotube has been demonstrated to recover sensor signal in 10-45 min. [16,19]][21] In this section, we show that it is possible to reduce the recovery time by another order of magnitude, down to 1 min, by applying a gate bias V gs near the threshold voltage (V th ) in the readout mode.To justify this choice of V gs , we need to understand changes in different sensing regions during the nanotube self-heating phase.
Consider the temperature profile along a self-heated, suspended nanotube.Hsu et al. measured Raman G mode downshifts to show that the temperature change, ΔT, is largest in the middle of the tube and decreases drastically near contacts (Figure 2). [22]The desorption rate of gas molecules is expected to depend on the local temperature T in an exponential fashion, i.e., desorption is orders of magnitude faster in the middle of the tube than on the contacts.Therefore, the time required for bulk desorption-i.e., desorption from the suspended nanotube channel-is mainly determined by ΔT in the transition regions marked with red ovals in Figure 2, which would heat up to intermediate ΔT values.These different sensing regions correspond to different responses in the device's transfer characteristics (I d -V gs ).Adsorption on the CNT channel (bulk) shifts the threshold voltage (V th ), whereas adsorption on contacts influences the tilt of the transfer characteristic. [14,23]Therefore, monitoring I d near V th , instead of in the ON-state, should allow us to observe the bulk desorption by nanotube self-heating faster.
To verify the reasoning above, we performed a gas measurement in humid air and increasing concentrations of NO 2 .In this measurement, we repeatedly applied 1 min self-heating pulses (recovery mode), which are significantly shorter than those used in previous work, [16,19] and monitored I d near V th (V gs = 0.5 V) Figure 2. Qualitative illustration of the temperature profile along a selfheated, suspended carbon nanotube.ΔT in the transition nanotube regions between the contacts and the middle of the nanotube (marked with red ovals) has been shown to be smaller than that in the middle of the tube. [22] the readout mode.As shown in Figure 3a, at all NO 2 concentrations ranging from 0 ppb (no NO 2 ) to 1 ppm, I d undershoots to approximately the same level after 1 min of nanotube self-heating, which is well below the pre-heating level in ambient conditions.This suggests successful recovery of the nanotube channel.In contrast, if we track I d at V gs = −3.0V (ON-state) from the same measurement (Figure 3b), we see that 1 min selfheating pulses are not sufficient to recover the baseline at this large V gs .This can be explained by the nonuniform temperature profile mentioned earlier; 1 min self-heating pulses appear to be sufficient for desorption in the transition regions and the middle of the nanotube, but they appear to be too short to desorb molecules from contacts.
To quantify the temperature increase that is required for bulk desorption within 1 min, we refer to transition state theory, which relates the desorption time  des to the binding energy E b as follows: [24] Here,  0 is the attempt frequency (≈10 12 s −1 for NO 2 [24] ), k B is the Boltzmann constant (8.62 × 10 −5 eV K −1 ), and T is temperature in K.The 12 h recovery time of carbon nanotube-based NO 2 sensors at room temperature corresponds to E b = −1 eV. [24,25]A reduction of recovery time from 12 h to 1 min corresponds to a temperature increase of ≈50 K in the transition regions marked with red ovals in Figure 2. Hsu et al. show ΔT of ≈40 K in the similar transition regions at an electrical heating power I d × V sd of 2.67 μW.They note that the bias conditions they use to heat the nanotube do not belong to the negative differential conductance (NDC) regime, which, as explained earlier, is an indication of substantial Joule heating and optical phonon scattering in suspended nanotubes. [17,26]On the other hand, the heating power of our device is significantly higher at ≈5.6 μW, and the bias conditions used in our recovery mode clearly belong to the NDC regime (Figure 1c).These support the feasibility of ΔT close to 50 K in our suspended nanotube, which enables effective bulk recovery in 1 min.
As shown in this section, using a gate bias near V th in the readout mode allows unprecedentedly fast signal recovery within 1 min when the self-heating effect in a suspended nanotube is used as a recovery method.Therefore, from now on, we will focus on I d at V gs = 0.5 V for our sensor readout.

Nonlinear Transient Readout
In the re-adsorption phase after a self-heating pulse, we determine the concentration of the analyte by using the initial slope or rate of the device current (I d ) transient.Previous work on initial slope sensing uses simple linear fits to sensor transients and does not consider the roles of other gas species in the system. [15,18,19]ere, we propose an alternative transient readout procedure based on multi-species adsorption dynamics, which does not only describe physical phenomena more accurately but also results in practical benefits such as lower LOD.
We hypothesize that nanotube self-heating desorbs environmental species like N 2 , O 2 , and H 2 O, in addition to analyte gas molecules-NO 2 in our case.Therefore, the re-adsorption of the environmental species creates the "baseline" transient signal ΔI d, env .To understand ΔI d, env , we performed a measurement, in which 1 min self-heating pulses were applied repeatedly in ambient conditions without NO 2 .As shown in Figure 4a, after each self-heating pulse, I d undershoots and then returns to the pre-heating level.This is consistent with our hypothesis that environmental species are desorbed by nanotube selfheating and gradually re-adsorb after the nanotube cools down.A similar undershoot after self-heating has been observed in an earlier study, both in the presence and absence of NO 2 . [16]igure 4b, which displays the five transient waveforms from Figure 4a on the same axes, shows that these baseline transients are highly reproducible.A rising exponential, a(1 − e −bt ), has been fitted to the data points in the first minute to yield Δ I d,env = (11.41± 1.57 nA) ⋅ (1 − e −(0.23±0.03min −1 )t ), where t is time after self-heating in minutes and ΔI d, env is in nA.Similar exponential fits can be obtained with the data points in the first 3 and 5 min (Figure 4b insets).The fact that an exponential fit with a single time constant accurately describes the transients suggests that, for up to 5 min after nanotube recovery, the transient is determined mostly by one dominant species.
In the following measurements, we introduced NO 2 into the system and gradually increased the concentration.Figure 5a shows I d records from two repeated measurements, the first of which has been presented in Figure 3a already.When multiple gas species exist in the system, multiple Langmuir isotherms are expected.Our measurement environment can be approximated as a two-species system consisting of NO 2 and the dominant environmental species identified in the previous paragraph., and cross terms such as ΔI d,env ⋅ ΔI d,NO 2 .We start with the simplest model, which is the sum of individual responses:    In Figure 5a, the initial slope of I d after self-heating is noticeably larger at higher NO 2 concentrations.Following Equation (2), we subtract the baseline exponential fit shown in Figure 4b from all waveforms and plot the resulting ΔI d,NO 2 transients on the same axes in Figure 5b.As shown in Figure 5b, the ΔI d,NO 2 transients are clearly distinguishable even at low concentrations between 0 and 100 ppb.
[29] Therefore, instead of a simple linear fit, we use an exponential fit function derived from the Langmuir dynamics to determine the initial rate of ΔI d,NO 2 transients: [15] Δ Here,  is the proportion of adsorption sites occupied by the analyte (surface coverage).p is the partial pressure of the analyte, which is equivalent to gas concentration at a given atmospheric pressure.K ads and K des are rate constants, which depend on the gas-surface system and temperature.𝜕I  is a conversion function between  and current.In case of our device, we assume a constant I  of 32 nA and K des is found to be 0.026 ± 0.003 min −1 , as determined by a decaying exponential fit (ae −K des t ) to an unas-sisted desorption phase of this device, i.e., from Min 91 to Min 101 in the measurements shown in Figure 5a.
Equation ( 3) is derived from traditional Langmuir dynamics involving one species.As our model assumes re-adsorption of multiple species at the same time, it would be more consistent if we used a multi-species adsorption model like the extended Langmuir (EL) model to describe ΔI d,NO 2 . [30]In Supporting Information, we compare Equation ( 3) with an equation derived from the multi-species, extended Langmuir (EL) model and show that Equation ( 3) is a good approximation in the time window we focus on, which is the first few minutes after nanotube self-heating.
Our main signal R is the initial rate of ΔI d,NO 2 transient directly after recovery ( dΔI d,NO 2 dt at t = 0).Derivations from Equation (3) suggest that R is I  ⋅ K ads p, and K ads p is the only unknown in Equation (3). [15]At all concentrations shown in Figure 5b, we fit Equation (3) to the data points in the first 1, 3, or 5 minute(s) to find the initial rates.As exemplified in Figure 6a with 1 ppm data, the ΔI d,NO 2 transient can be fitted very well by Equation (3).The extracted values of K ads are almost constant over the range of NO 2 concentrations, as expected from measurements performed at the same temperature: 0.20 ± 0.01 min −1 ppm −1 (50 ppb), 0.15 ± 0.005 min −1 ppm −1 (100 ppb), 0.16 ± 0.002 min −1 ppm −1 (200 ppb), 0.19 ± 0.002 min −1 ppm −1 (500 ppb), and 0.25 ± 0.002  2)) can explain the net ΔI d transients very well for up to 5 min after nanotube self-heating.Although we do not rule out higherorder terms and cross terms, we leave it for future work to assess the magnitude of these terms with specifically designed experiments.
The consideration of environmental species in the system (Equation ( 2)) and the use of a physical model-based, nonlinear fit function (Equation ( 3)) allow us to determine initial rate R consistently over a range of readout times (Figure 6a,b).In contrast, previous approaches using linear fits underestimate the initial slope S of the concave sensor transients as the readout window becomes longer (Figure 6c,d).
In addition to furthering our understanding of multi-species adsorption dynamics, our nonlinear transient sensing method offers a practical benefit in lowering the sensor's LOD, which we define as 3 noise /(sensitivity).Here,  noise is the standard deviation of initial rates or slopes of five baseline transients shown in Figure 4a, after the subtraction of ΔI d, env .
Figure 7a shows that  noise decreases with increasing readout time for both linear transient sensing and nonlinear transient sensing.This is as expected, since longer readout time entails more averaging in a statistical sense.On the other hand, the sensitivity changes differently in the two methods.In case of linear transient sensing, the diminishing initial slope with increasing readout time, which is exemplified in Figure 6c, results in decreasing sensitivity (slopes of red dashed lines in Figure 6d) with increasing readout time.This consequently limits the improvement in LOD with increasing readout time; the extrapolated LOD is calculated to be 66 ppb with a readout time of 1 min and 35 ppb with a readout time of 5 min (Figure 7b).In contrast, with our nonlinear fit function, the sensitivity stays almost constant over a range of readout times (slopes of red dashed lines in Figure 6b), and as a result, the extrapolated LOD fully benefits from the  noise reduction with increasing readout time.Thus, the extrapolated LOD, which is 61 ppb with a 1 min readout window, goes down to 9 ppb as the readout window is extended to 5 min (Figure 7b).
To investigate the reproducibility of our sensor behavior, the second device has been calibrated and tested in the same way as the first device.As summarized in Table 1, the second device shows noticeable similarity to the first device in terms of K des , K ads , R ON , onset of NDC, and self-heating power.As a result, key performance metrics also lie in similar ranges.This second device shows successful signal recovery after 1 min of self-heating at ≈4.2 μW and shows Langmuir-like transients after signal recovery.The extrapolated LOD of this second device is 23 ppb with a readout time of 5 min.
One important consideration is how our readout scheme would work in a real environment, in which the NO 2 concentrations change gradually rather than in stepwise increments as in our measurements.Our operation scheme recovers the sensor signal to the baseline and reads out the analyte concentration over a short time period of 1 to 5 min.Therefore, our readout algorithm can be relevant unless the analyte concentration changes significantly within this 1 to 5 min window.Measurements from air quality monitoring stations in urban and traffic environments suggest that NO 2 concentrations change rather slowly (a few ppb in an hour). [13]However, in case of an abrupt concentration change within the readout window, the adsorption dynamics would indeed be more complicated, and our algorithm would yield an intermediate value between the lowest and highest NO 2 concentrations in the readout window.

Comparison with Other Nanostructure-Based NO 2 Sensors
Table 2 compares our sensor with selected nanostructure-based NO 2 sensors demonstrating good performance in LOD, power Table 1.Comparison of two gas sensors in terms of key parameters required for sensor calibration.K ads is the mean of the values calculated at 50 ppb, 100 ppb, 200 ppb, 500 ppb, and 1 ppm.Measurement data for Device 2 are available in Supporting Information.
V sd for I peak (onset of NDC) Self-heating power Device 1 0.026 ± 0.003 min −1 0.  [9] Self-heated graphene Our work provides a promising alternative that only requires a few μW for significantly shorter readout and recovery times.

Investigating the Effect of Repeated Self-Heating on CNT with Raman Spectroscopy
In previous work, self-heating-based recovery has been demonstrated in dry air, after NO 2 flow is turned off. [16]Our experiment in this work shows that nanotube self-heating is also effective in conditions of a typical field test containing humid air and NO 2 .
In addition, it has not been studied whether repeated self-heating in the presence of NO 2 , a strong oxidizer, damages the nanotube.Therefore, we performed Raman spectroscopy to estimate the structural quality of the nanotube after our measurement campaign.The gas measurements and initialization procedures before and after the measurements included ≈3 h of self-heating pulses in total.The Raman image presented in Figure 8a, obtained with a confocal Raman microscope inVia from Renishaw, shows an individual carbon nanotube suspended between source and drain contacts.Multiple Raman spectra, a representative of which is shown in Figure 8b, were obtained with area imaging scans with a power density of 30.2 W mm −2 and a measurement time of 60 s per spot.None of them shows a visible D mode (1300-1400 cm −1 ) indicative of structural defects. [31]This suggests that the nanotube has not been damaged by hours of repeated self-heating in NO 2 (Table 2).

Conclusion
In summary, a fast, ultralow-power NO 2 sensor based on a selfheated, suspended CNT has been demonstrated.We reset the suspended nanotube channel by accelerating gas desorption via nanotube self-heating and read out the initial rate of sensor transients, as gas molecules re-adsorb on the nanotube surface, to determine the analyte concentration.By considering the temperature changes during nanotube self-heating and their effect on different sensing regions, we find that a gate bias (V gs ) near the threshold voltage (V th ) enables sensor signal recovery in 1 min at a very low power of ≈5.6 μW.In addition, readout based on a physics-based, nonlinear determination of the initial rate of the sensor transient requires only 1-5 min.If the readout mode and recovery mode were alternated, our measurement cycle would take only 2 min and still yield an extrapolated LOD of 61 ppb.The readout time can be extended to a few minutes if a lower LOD is desired.An extrapolated LOD of 9 ppb has been demonstrated with a readout time of 5 min, which is a four-fold improvement in LOD over previously reported linear fit-based sensing.
The meaningful improvements reported in this work bring us closer to practical NO 2 sensing with carbon nanotubes in low-power mobile or Internet-of-Things (IoT) sensing systems.We expect future work to build upon our studies and further investigate other practically important topics such as empirical LOD determined with repeated ppb-level measurements, crosssensitivity analysis, influence of varying humidity levels, and drift effect due to contact degradation.

Experimental Section
Fabrication of Suspended CNT Gas Sensors by Dry Transfer: Figure 1a schematically illustrates the operation of the suspended CNT gas sensor.To fabricate this device, the device structures were patterned with standard surface micromachining, and the electrodes (40 nm Pd on 1 nm Cr) were deposited by e-beam metal evaporation.Both source-drain pitch and gate distance are 2 m.Ultraclean, suspended nanotubes were grown by chemical vapor deposition (CVD) at 850 °C between Si/SiO 2 cantilevers and then mechanically transferred onto the source and drain electrodes of the device with a piezo-driven micromanipulator.The cantilevers containing suspended nanotubes and the device were aligned by the manipulation of the piezo stages.This mechanical transfer does not expose the nanotubes to any processing chemical and therefore maintains the cleanliness of CVD-grown nanotubes. [32]The nanotubes typically have a diameter of 1.5-2.5 nm, as determined by radial breathing modes in Raman spectra.Further details of the fabrication process are described in a recent work. [33]ontact Improvement via Electrode Etching with Ar+ and Thermal Annealing: The study reported that it was essential to lower the contact resistance at the CNT-electrode interface for the self-heating operation.If the contact resistance was too high, significant Joule dissipation on contacts prevents the suspended channel from heating up.Therefore, the electrode surface was cleaned with Ar + etching directly before the dry transfer of the CNT (25 W, 1 min, Von Ardenne CS 320 S).After the transfer, the device was annealed at 250 °C for 1 h for further contact improvement. [33]With the help of these processes, the ON-resistance of the device was 97 kΩ (Figure 1b, at V sd = 0.1 V, V gs = -10 V) and the output characteristic (I d -V sd ) shows current peaking followed by a negative differential conductance (NDC) regime (Figure 1c), an indication of substantial Joule heating of the suspended CNT channel. [17]as Sensing Measurements: For gas sensing measurements, synthetic air (20% O 2 /80% N 2 in volume) was used as a diluent gas.Humidity was controlled by diverting a gas line through a bubbler filled with deionized water and a thermostat (Julabo F-34).Temperature (Pt-1000) and humidity sensors (Honeywell HIH-4000 Series) were installed at the exhaust of the gas chamber.To approach real-life operation conditions, all the measurements were performed at 23 ± 1 °C and 45 ± 1.5% R.H. Electrical characterizations were performed with LabVIEW-controlled DAQ boards (National Instruments NI 6289 and NI BNC-2110).In the recovery mode, bias conditions in the NDC regime (V sd = 1.5 V, V gs = −10 V) were applied constantly for 1 min.In the readout mode, transfer characteristics were recorded every 250 ms with a pulsed V gs sweep [34] and a constant V sd of 0.1 V.The measurements constantly alternate between the two modes.
Raman Spectroscopy: After the gas measurement campaign, Raman spectroscopy was performed to characterize the effect of repeated selfheating in NO 2 on the structural quality of the nanotube.Raman spectra were obtained from the center of the suspended nanotube with a confocal Raman microscope inVia from Renishaw.The nanotube was exposed to a laser beam with an excitation wavelength of 515 nm, which was focused with an objective with 50× magnification.The area imaging scans were performed with a power density of 30.2 W mm −2 and a measurement time of 60 s per spot.
Statistical Analysis: In the gas sensing measurements, device current (I d ) was recorded.An exponential fit to the baseline transients was calculated in a measurement without NO 2 .In measurements with NO 2 , at each concentration, the first data point was subtracted from all data points of the transient, to obtain a waveform representing ΔI d .Then, the baseline exponential fit was subtracted from the transient at a defined NO 2 concentration.The resulting waveform was fitted by a Langmuir-model-based exponential fit function to determine the sensor signal.To indicate the level of electronic noise in the measurements, all fit coefficients were presented with ±  values, both in the manuscript text and figures.All data analysis and fittings were performed with MATLAB.

Figure 1 .
Figure 1.Suspended carbon nanotube gas sensor.a) Schematic illustrations of accelerated gas desorption from the self-heated, suspended CNT channel (left) and re-adsorption of gas molecules as the CNT cools down (right).b) Hysteresis-free transfer characteristic (I d -V gs ) of our device in ambient conditions (23 °C, 45% R.H.) obtained with a pulsed V gs sweep.[34]c) Output characteristic (I d -V sd ) of the device showing a peak I d at V sd of ≈0.8 V followed by a negative differential conductance (NDC) regime.The bias conditions of V sd = 1.5 V and V gs = −10 V (marked with red circle), which belong to the NDC regime, are used to induce the nanotube self-heating effect for accelerated sensor recovery in our gas measurements.
Then, ΔI d can be expressed as a polynomial containing individual responses ΔI d,env and ΔI d,NO 2 , higher-order terms such as ΔI 2 d,env and ΔI 2 d,NO 2

Figure 3 .
Figure 3. Device current (I d ) measured a) near-threshold voltage, V gs = 0.5 V, and b) at a larger gate bias, V gs = -3 V, from the same gas measurement with repeated 1 min self-heating pulses and increasing concentrations of NO 2, which are represented by dashed black lines.a) At V gs = 0.5 V, I d undershoots to approximately the same level after 1 min of nanotube self-heating at all NO 2 concentrations.Blue represents I d in the readout mode with low biases and red represents recovery mode.b) In contrast, at V gs = −3 V, self-heating pulses do not recover I d completely.Green represents I d in the readout mode and red represents recovery mode.

Figure 4 .
Figure 4. Baseline transients in ambient conditions (23 °C, 45% R.H.) without NO 2 .a) Current record from a measurement with repeated 1 min selfheating pulses.Blue represents I d in the readout mode with low biases and red represents recovery mode.At the end of each self-heating pulse, I d in the readout mode undershoots and then returns to the pre-heating marked with a gray dashed line.b) Five I d responses directly after self-heating pulses (corresponding to the highlighted parts in (a)) plotted on the same axes.These baseline transients show strong reproducibility.Black solid line represents an exponential fit to the data points in the first minute (number of data points: 240).The exponential growth of I d can be explained by the re-adsorption of environmental species.Insets: exponential fits to the data points in the first 3 and 5 min.

Figure 5 .
Figure 5. NO 2 gas measurements with repeated 1 min self-heating pulses.a) Two measurements with increasing concentrations of NO 2 .b) Transient ΔI d,NO 2 responses directly after self-heating pulses (corresponding to the highlighted parts in (a), 1st measurement).The initial slope of the ΔI d,NO 2 transient increases as NO 2 concentration increases.To plot ΔI d,NO 2 waveforms, the baseline exponential fit found in the absence of NO 2 (ΔI d,env , Figure 4b) has been subtracted from ΔI d waveforms.

Figure 6 .
Figure 6.Comparison of a,b) nonlinear transient sensing and c,d) linear transient sensing with varying readout times of 1-5 min.a,c) Determination of the initial rate R of ΔI d,NO 2 transient via Equation (3) a) and of the initial slope S via linear fits c) with data points in the first 1, 3, and 5 min of 1 ppm NO 2 transient (number of data points: 240 for 1 min fits, 720 for 3 min fits, 1200 for 5 min fits).Initial rate R does not vary significantly with the readout time (a), while a linear model underestimates the initial slope S of the concave ΔI d,NO 2 transient (c) as the readout window becomes longer.b,d) Sensor response to a range of NO 2 concentrations when b) initial rate R and d) initial slope S are calculated with 1 min (left), 3 min (middle), and 5 min (right) readout windows.Dots represent means and error bars represent standard deviations across measurements (sample size: 5 at 0 ppb, 2 at other concentrations); red dashed lines are linear sensor response functions.The 1 ppm NO 2 data point from the second measurement (Figure5a) is considered an outlier and is marked by a blue rhombus.Sensitivity, the slope of a red dashed line, is not affected significantly by readout time for rate R (b), but it decreases noticeably for slope S as readout time becomes longer (d).

Figure 7 .
Figure 7. a)  noise and b) extrapolated limit of detection (LOD) as a function of readout time for the linear (red) and nonlinear (blue) transient sensing methods.
readout speed.Recent advances show impressive progress toward ppb-level sensing, but their readout and recovery times are still longer than what is required for practical applications.In addition, for their best performance in readout speed and LOD, other nanostructure-based sensors require 2 to 4 orders of magnitude more power than our CNT-based sensors.

Figure 8 .
Figure 8. Characterization of nanotube structural quality after gas measurement.a) Raman image of the carbon nanotube gas sensor reported in this work, superimposed on an optical microscopy image.The Raman image was created with a filter for carbon nanotube's characteristic G mode. b) Raman spectra from the suspended channel obtained after our measurement campaign including ≈3 h of self-heating.The absence of D mode suggests that no structural damage has been introduced by repeated selfheating.The full-range Raman spectra can be found in Supporting Information (Figure S2, Supporting Information).

Table 2 .
Comparison of selected NO 2 sensors based on nanostructures.Values not directly reported in the cited works have been estimated from plots and are marked with asterisks.Unless specified otherwise, LOD assumes a 3 confidence interval.For works that use steady-state sensing, t readout and t recovery indicate the time for the signal to reach 90% and 10% of its steady-state value, respectively.