Introduction

Chemical gas sensors based on semiconducting oxides are widely studied for various applications in environmental monitoring, industrial safety, and health diagnostics1,2,3. In environmental monitoring, they are used to detect pollutants and hazardous gases in air, enabling assessment of air quality and early warning of harmful emissions4. In industrial settings, gas sensors help prevent accidents by providing real-time detection of flammable or toxic gases leaking in factories, mines, or processing plants5. In the medical and health domain, sensitive gas sensors allow non-invasive diagnostics–for example, analyzing exhaled breath for biomarkers of disease or monitoring indoor air for harmful chemicals6. These diverse applications demand sensors that are not only highly sensitive and selective to specific gases, but also reliable in their response.

Most research to date has focused on maximizing sensor response—for example, by optimizing sensing materials, catalysts, and device geometry to improve sensitivity and selectivity7,8. In contrast, comparatively little attention has been devoted to the intrinsic electronic noise of the sensor devices. This is a critical oversight: low-frequency noise (LFN) in the transducer of sensors can destabilize the baseline signal and limit the detection reliability9,10. In practice, slow surface reaction kinetics in chemical sensors mean that 1/f-type LFN often dominates the noise floor of the sensing signal, directly affecting the signal-to-noise ratio (SNR) and thus the minimum gas concentration (limit of detection) the sensor can discern. A recent studies emphasize that LFN has a considerable impact on sensor stability and must be characterized to ensure reliable performance.

This issue of overlooked noise extends beyond traditional resistive gas sensors to advanced thin-film transistor (TFT)-based platforms, particularly those employing amorphous oxide semiconductors (AOS). AOS TFTs using indium oxide, zinc oxide, tin oxide, indium-gallium-zinc oxide, have emerged as promising gas sensor transducers due to their inherent chemical sensitivity, surface accessibility, tunable selectivity via gate modulation, and compatibility with flexible, large-area substrates11,12,13,14. In these devices, target gas molecules interacting with the oxide semiconductor channel alter channel conductivity by donating or withdrawing electrons or influencing trap states, resulting in an amplified transistor output signal. Despite these advantages, systematic studies of low-frequency noise (LFN) in AOS TFT-based sensors remain sparse, leaving a significant knowledge gap regarding how noise may constrain practical sensor performance. Additionally, AOS TFT gas sensors typically operate at wide range of temperatures (25–150 °C) to enhance adsorption kinetics, improve reaction selectivity, and accelerate sensor recovery15,16. Compared with Si CMOS transistors, AOS channels have a much higher density of subgap states and structural disorder, resulting in larger 1/f noise. Moreover, while Si FETs show only weakly temperature-dependent 1/f noise, intentional channel heating in AOS TFTs to activate gas reactions simultaneously thermally activates additional trap states and excess LFN17. However, while extensive research has focused on temperature-dependent gas sensitivity and selectivity, the impact of elevated temperatures on LFN behavior is largely unexplored. Given that temperature profoundly influences both electrical properties (carrier mobility, trapping dynamics) and chemical interactions (surface adsorption kinetics), understanding the interplay between temperature and noise in AOS TFT sensors is crucial for achieving reliable, low-noise, high-performance sensor operation.

In this study, we present a comprehensive investigation of LFN in Si-doped ZnSnO TFT-type gas sensors under various temperature conditions. The sensor platforms are fabricated in a staggered bottom-gate configuration with an exposed oxide semiconductor channel, a structure conducive to surface adsorption-based sensing. We perform temperature-dependent LFN measurements and analyze the observed noise behavior in terms of carrier number fluctuation (CNF) model associated with trapping/detrapping dynamics, including the contribution of thermally activated deep states that become ionized at higher temperatures. The analysis reveals that as the device temperature rises, an increasing fraction of carriers exchange with deeper gap states, elevating the 1/f noise magnitude and altering its bias dependence. The knowledge gained here lays the groundwork for engineering lower-noise AOS sensor transducers by defining optimal operating biases and temperatures that maximize SNR.

Fabrication process

Figure 1(a) shows the cross-sectional schematic illustration of the fabricated SZTO TFT. The fabrication process of the SZTO TFTs is as follow (Fig. 1(b–e)): The substrate was heavily doped p-type Si with 100 nm thermal oxidation of SiO2. Prior to device fabrication, substrates underwent a cleaning process using acetone-methanol-deionized water. The amorphous SZTO thin film was deposited using a radio frequency magnetron sputter using a 2-inch Si-Zn-Sn-O target with 1 wt% Si. The channel layers were patterned through photolithography and wet-etching processes. Subsequently, an annealing process was conducted in the air at 150 °C for 2 h. The Ti (10 nm) and Al (40 nm) source/drain electrodes were deposited sequentially in the same chamber without breaking the vacuum using E-beam for Ti and thermal evaporators for Al, respectively. The photolithography and lift-off process were used for patterning the source/drain electrodes.

Fig. 1: Fabrication process of the SZTO TFTs.
Fig. 1: Fabrication process of the SZTO TFTs.The alternative text for this image may have been generated using AI.
Full size image

a Schematic diagram and fabrication steps of the SZTO TFTs. b Fabrication process for SZTO TFTs. Cross-sectional schematic of the completed SZTO TFT structure illustrating each material layer. b Initial substrate cleaning using acetone, methanol, and deionized water. Deposition and patterning of the amorphous SZTO channel layer using RF magnetron sputtering, photolithography, and wet etching. Annealing of the patterned channel layer at 150 °C in air for 2 hours to stabilize film properties and reduce defects. Sequential deposition of source/drain electrodes consisting of Ti (10 nm) and Al (40 nm) layers using E-beam and thermal evaporation without breaking vacuum

In this study, Si-doping was employed to improve device stability and enhance temperature robustness by suppressing oxygen vacancies, which are prevalent defect states responsible for instability and performance degradation in amorphous oxide semiconductors. By reducing the density and influence of these oxygen vacancies, Si-doping significantly stabilizes the subgap DOS, enabling reliable and reproducible investigation of temperature-dependent subgap state effects on device operation and noise characteristics.

Gas sensing and LFN measurement systems

The gas-sensing performance and noise characteristics of the fabricated SZTO TFT-type sensors were systematically evaluated using a custom-designed measurement setup. Electrical characteristics and gas-response measurements were performed using a semiconductor parameter analyzer (Keysight B1500A) and a precision probe station within an enclosed test chamber equipped with a temperature-controlled chuck. The chamber temperature was precisely controlled and maintained within the desired range to simulate typical sensor operating conditions. To introduce and regulate target gas concentrations, specifically NO2, a mass flow controller (MJ Technics MFC) was utilized. The NO2 gas stream was carefully diluted with dry synthetic air (relative humidity (RH) maintained at approximately 4% at 27 °C) to achieve accurate and stable gas concentrations within the test chamber. During gas-sensing experiments, the sensor device was exposed sequentially to clean air and controlled concentrations of NO2, with sufficient stabilization time between exposures to accurately assess sensor response and recovery behavior.

For detailed investigation of LFN characteristics, as well as evaluation of signal-to-noise ratio SNR, the drain current power spectral density (PSD) of the TFT sensors was measured. The noise measurements were conducted by biasing the TFT at different gate and drain voltages corresponding to various sensing conditions. The drain current fluctuations were amplified using a low-noise current amplifier (Stanford Research Systems SR570) and subsequently recorded using a dynamic signal analyzer (Keysight 35670 A), providing detailed frequency-domain noise spectra over the range of 1 Hz–1600 Hz. This comprehensive measurement environment allowed us to systematically correlate temperature-induced noise characteristics with subgap trap state dynamics, crucially elucidating the noise performance limits and detection capabilities of the fabricated SZTO TFT sensor platform. For ensuring the reliability and statistical robustness of the measurement and result analysis, both the noise characterization and gas-response evaluations were conducted using five independent device samples.”

Results and discussion

Electrical characteristics of SZTO TFTs

The electrical characteristics of the fabricated SZTO TFTs were analyzed to evaluate device operation and underlying physical mechanisms. Figure 2(a) shows the transfer characteristics (ID–VGS) of the fabricated SZTO TFT measured at different drain-to-source bias (VDS) values. Note that the channel width (W) and length (L) are 250 and 50 μm, respectively. Figure 2(b) shows the output characteristics (ID–VDS) of the device measured at different gate-to-source (VGS) bias values. To investigate the characteristics of the SZTO, the field-effect mobility (μFE) is extracted using following equation18.

$${\mu }_{{\rm{FE}}}=-\frac{d{I}_{{\rm{D}}}}{d{V}_{{\rm{GS}}}}\times \frac{L}{W{C}_{{\rm{ox}}}{V}_{{\rm{DS}}}}$$
(1)

where Cox is the gate oxide capacitance per area. To eliminate the L-dependent contact resistance effects on μFE, the extracted μFE is normalized with 1-2RC/RTotal; RC: contact resistance and RTotal: total resistance19. The extracted μFE value of the SZTO channel is 20.1 cm/Vs. Note that the high carrier mobility of the transistor mainly stems from Si doping. Silicon incorporation in IZO profoundly influences the defect chemistry and electronic structure of the amorphous oxide, which in turn affects carrier transport. In undoped IZO, oxygen vacancies act as native donor sites that contribute free electrons but also introduce subgap states. By alloying IZO with a small fraction of Si, these oxygen-vacancy-related defects can be significantly suppressed due to the strong Si–O bond since Si has a high metal–oxygen bond dissociation energy and low electronegativity. This not only causes a positive shift in threshold voltage (Vth) but also a reduction in off-state current. Furthermore, Si doping alters the subgap DOS in the amorphous semiconductor. Oxygen vacancies in amorphous oxides introduce gap states, such as shallow donor states near the conduction band or deeper trap states, depending on their charge state. By tying up oxygen and stabilizing the oxide lattice, Si effectively lowers the density of these donor-like gap states20,21. Concurrently, the stronger bonding and improved structural order can reduce the band-tail state density by minimizing structural disorder, resulting in an increase in carrier mobility.

Fig. 2: Electrical characteristics of SZTO TFTs.
Fig. 2: Electrical characteristics of SZTO TFTs.The alternative text for this image may have been generated using AI.
Full size image

a ID–VGS of the fabricated SZTO TFT measured at VDS values of 0.1, 1.0, and 5.0 V. b ID–VDS of the device measured at different VGS values of 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, and 4.0 V. c ID–VGS of the SZTO TFT at various T values

Temperature has a pronounced impact on the I–V behavior of SZTO and related amorphous oxide TFTs because it alters the activity of oxygen vacancies and the occupancy of trap states. Figure 2(c) shows the ID–VGS of the SZTO TFT at various temperature values. With an increase in T, a negative shift of threshold voltage (Vth) and an increase in subthreshold swing are observed. Two thermally driven processes contribute to this trend. First, donor sites, such as neutral oxygen vacancies, become thermally ionized, releasing electrons into the conduction band. Any shallow donor states that are partially inactive at room temperature will be almost fully ionized at higher T. This thermal activation of oxygen-vacancy donors directly boosts the free electron concentration in the channel. Second, electrons that were trapped in subgap states gain enough thermal energy to escape into the conduction band. This thermal de-trapping means that a larger fraction of the charge is mobile, which increases the conductivity.

LFN characteristics of SZTO TFTs

Based on these findings, we investigate the LFN characteristics of SZTO TFT. Figure 3(a–c) show the drain current (ID) normalized power spectral density (PSD) (SID/ID2) versus frequency for the devices with varying T, at values of 40, 100, and 800 nA, respectively. Note that the measurements are conducted three times to ensure the reliability of the results. All spectra exhibit an approximate 1/f dependence over the from 10 to 103 Hz range, indicative of flicker noise dominated by charge fluctuation22,23. Two interesting phenomena are observed. Firstly, as T increases, the SID/ID2 of the SZTO TFT shows a noticeable increase. In particular, there is a significant increase in PSD at lower frequency ranges, resulting in a steeper slope of the PSD curve. Secondly, the differences among PSD at different T values diminish as the operation region moves into the high ID region (ID = 800 nA). To examine the underlying reason for these observations, the origin of 1/f noise in SZTO TFTs is investigated. Generally, the LFN characteristics in amorphous oxide semiconductor TFTs are explained using two primary models: the carrier number fluctuation (CNF) model and the Hooge mobility fluctuation (HMF) model24,25,26. The CNF model attributes 1/f noise to fluctuations in the number of free carriers due to trapping and detrapping events at defects near the semiconductor–gate dielectric interface. The CNF model is mathematically expressed as24:

$$\frac{{S}_{\mathrm{ID}}}{{I}_{{\rm{D}}}^{2}}={\left(\frac{{g}_{{\rm{m}}}}{{I}_{{\rm{D}}}}\right)}^{2}\frac{{q}^{2}{kT}{N}_{T}\lambda }{{WL}{{C}_{{ox}}}^{2}f}$$
(2)

where gm is the transconductance, q is the elementary charge, NT is the volume trap density, and \(\lambda\) is the tunneling attenuation coefficient of the gate oxide. In contrast, the HMF model explains 1/f noise through fluctuations in carrier mobility, primarily due to scattering mechanisms within the channel region25. In AOS TFTs, factors influencing mobility fluctuations and thus potentially contributing to 1/f noise under the HMF model include structural disorder and compositional inhomogeneity within the amorphous oxide layer, leading to localized variations in mobility and variations in carrier scattering due to fluctuating electric fields from trapped charges in subgap states27,28. According to Hooge’s empirical formulation, the normalized noise is given by25:

$$\frac{{S}_{{ID}}}{{{I}_{D}}^{2}}=\frac{{\alpha }_{H}{\mu }_{{eff}}2{kT}}{f{L}^{2}{I}_{D}}$$
(3)

where \(\alpha\)H represents Hooge’s parameter and \({\mu }_{{eff}}\) is the effective carrier mobility.

Fig. 3: LFN characteristics of SZTO TFTs.
Fig. 3: LFN characteristics of SZTO TFTs.The alternative text for this image may have been generated using AI.
Full size image

a–c SID/ID2 versus f measured at various drain current levels (ID = 40, 100, and 800 nA) and operating temperatures (25, 50, and 75oC). The observed 1/f-type noise spectra demonstrate significant temperature dependence, particularly pronounced in the low-frequency and low-current regimes. Comparison between SID/ID2 measured at 10 Hz and (gm/ID)2 across different drain currents at (d) 25, (e) 50, and (f) 75 °C, respectively, measured in five samples. At 25 °C, the noise characteristics follow the expected trend predicted by the CNF model (solid line). However, as temperature increases to 50 and 75 °C, deviations (shaded regions) from the CNF model occur in the low-current region, indicating the emergence of thermally activated excess noise associated with subgap donor-like states

To assess whether the behavior of SZTO TFT aligns with the CNF model, the SID/ID2 sampled at 10 Hz is plotted against the ID with a square of transconductance efficiency ((gm/ID)2), as shown in Fig. 3(d–f). Note that the observed noise does not follow the expected dependence characteristic of the HMF model. At 25 °C, the SID/ID2 and (gm/ID)2 exhibit the same behavior in all ID ranges, except for the ID = 1 µA, confirming that the origin of the 1/f noise is mostly the CNF. However, with an increase in T, the deviation of SID/ID2 from (gm/ID)2 is observed in the low ID region, and this deviation becomes more pronounced in the wider ID regions. It is important to note that this excess noise behavior cannot be explained using the conventional CNF model. Specifically, this notable increase in normalized noise at elevated temperature and low current is not explained by slower (longer time constant) traps becoming activated, as thermal excitation typically decreases trap time constants by facilitating faster trapping/detrapping dynamics. Furthermore, this excess noise behavior also cannot be adequately explained by the correlated mobility fluctuation (CMF) model. The CMF model describes 1/f noise through fluctuations in carrier mobility correlated with carrier number fluctuations29,30. Essentially, it combines the CNF model and mobility variations arising from trapped charge-induced Coulombic scattering at the semiconductor–dielectric interface. In typical applications, the CMF model successfully explains noise characteristics primarily at higher current regimes, where the higher carrier density enhances the correlation between trapped charges and mobility fluctuations29. However, the excess noise observed in this study predominantly occurs at lower current levels, which contradicts the expected operational regime of the CMF model. Therefore, the observed temperature-activated excess noise at low current levels indicates the involvement of additional mechanisms beyond both CNF and CMF models, likely attributed to thermally activated access to a broader and higher-density distribution of subgap trap states, particularly donor-like states near the conduction band edge.

Excess noise in SZTO TFTs

As aforementioned, in AOS TFTs, excess noise can arise from the subgap DOS31. In AOS materials, various DOS, such as tail state and donor states, exist, which act as defects responsible for carrier trapping and detrapping processes. Notably, the subgap DOS possesses energy-dependent characteristics, making it important to analyze LFN characteristics while considering these energy variations32,33. One of the crucial aspects to be determined is the relationship between VGS and EF − EC. The data presented in Fig. 4b is obtained by varying the value of VGS, resulting in a change in the EF and, consequently, altering the energy levels of the defects contributing to the LFN. To extract this relationship, the following equation is employed34.

$${I}_{{\rm{D}}}({V}_{{\rm{GS}}})={Aq}{N}_{{\rm{C}}}{\mu }_{{\rm{eff}}}E\exp \left(\frac{{E}_{{\rm{F}}}({V}_{{\rm{GS}}})-{E}_{{\rm{C}}}}{{kT}}\right),$$
(4)

where A is the conducting channel area, NC is the DOS in the conduction band, μeff is the effective mobility, and E is the electrical field. And

$${E}_{F}({V}_{{\rm{GS}}})-{E}_{{\rm{REF}}}={k}_{{\rm{B}}}T{\int }_{{V}_{{\rm{T}}}+{V}_{{\rm{DS}}}}^{{V}_{{\rm{GS}}}}\frac{{g}_{{\rm{m}}}({V}_{{\rm{GS}}})}{{I}_{{\rm{D}}}({V}_{{\rm{GS}}})}d{V}_{{\rm{GS}}},$$
(5)

where EREF is EF at VGS = VT + VDS. Figure 4(a) shows the gm/ID versus VGS of the ZTO TFTs at various Ts. This integral effectively maps changes in gate voltage directly to shifts in the Fermi level position relative to the conduction band minimum, allowing precise identification of the energy levels of active traps involved in the LFN processes. With increased gate voltage, EF moves progressively closer to EC, enabling a more conductive state and filling shallow traps near the conduction band edge. By combining (5) and ID value at EF = EC, the relationship between the energy and VGS can be finalized. Figure 4(c) shows that the slope of the plot between the ln(ID) and EF is 1/kT, affirming the validity of the derived relationship. This consistency strongly confirms the accuracy and validity of the extraction method employed. The observed temperature-dependent shift in these plots is consistent with increased carrier thermal energy at elevated temperatures, further facilitating trap activation and affecting the carrier transport mechanisms.

Fig. 4: Effects of subgap DOS.
Fig. 4: Effects of subgap DOS.The alternative text for this image may have been generated using AI.
Full size image

a (gm/ID) versus VGS at different operating temperatures for the SZTO TFTs. b Extracted energy difference between the EF and EC as a function of VGS at various temperatures, demonstrating temperature-dependent shifts in the Fermi level position. c Logarithmic plot of drain current (ln(ID)) versus EF, confirming a consistent slope of 1/kT, validating the extraction method for determining EF – EC. d Schematic representation illustrating the subgap DOS distribution, identifying donor states (red curve) and tail states (purple curve) relative to the conduction band edge. e Extracted NT ersus EF – EC at different temperatures measured in five samples. The significant increase in NT at higher temperatures arises from thermally activated donor states, dominating the excess noise contribution in the low current (low-bias) region, whereas tail states predominantly influence higher current (high-bias) regimes

Figure 4(d) shows the schematic of DOS versus EF-EC for the SZTO TFTs concerning the LFN measurement. As demonstrated in Fig. 4(d), both tail states and donor states are present near the EC. Depending on the operation region of the SZTO TFTs, the different types of subgap DOS affect the 1/f noise of the device. In the low ID region, the LFN is mainly generated at the energy level relatively far away from the EC. Thus, the excess noise in this region can be generated from the donor states, whereas in the high ID region, it predominantly stems from the tail states. Figure 4(e) shows the NT with respect to energy at various Ts. At 25 °C, the NT exhibits a constant value at the energy level 0.1 eV below the EC. As the VGS increases and the energy level is formed closer to EC, excess noise is generated, indicating the influence of the tail states. Interestingly, as the T increases, a significant increase in excess NT is observed at the energy level 0.1 eV below the EC. In amorphous oxide semiconductors, such donor-like subgap states are widely attributed to oxygen-vacancy–related defects and their complexes, which introduce shallow donor levels within ~0.1–0.3 eV below the conduction band edge35. Si incorporation is known to strongly bind oxygen and suppress broad deep defect bands, but it does not completely eliminate oxygen vacancies; instead, it narrows their energy distribution and shifts the dominant donor-like states closer to Ec.

The reason for the presence of the effects of donor states only at higher Ts should be elaborated upon because the number of subgap DOS is relatively T-independent35. Donor-like subgap traps in SZTO TFTs remain largely inert at low temperature but become dynamically active at elevated temperature, thereby amplifying the LFN through thermally driven trapping–detrapping processes. These donor-like states lie near the conduction band and have finite activation energies for electron emission into the band. At low temperatures, they stay filled/neutral with electrons “frozen” in the traps. However, at higher temperatures, electrons gain enough thermal energy to escape to the conduction band, leaving behind positively charged donor centers. The thermal release of a trapped electron over its activation barrier is characterized by an emission time constant τ that follows an Arrhenius law (τ ∼ τ0exp(Ea/kT))36. Deep donor-like states (larger Ea) thus have long τ at low T, effectively contributing negligible noise when the device is at low temperature. As T rises, however, τ shortens dramatically for these deep traps, enabling frequent carrier exchange with the band. Each capture or emission event is a stochastic process in which electrons are intermittently captured and emitted with exponentially distributed waiting times. Each such event perturbs the channel charge, producing a fluctuation in drain current. In a device with many traps distributed in energy near the band edge, the broad spectrum of τ values leads to a superposition of Lorentzian fluctuations across a wide range of frequencies36. The deep-level traps generate the slow fluctuations, so when these donor-like states become thermally active at high T, they chiefly boost the low-frequency end of the noise spectrum. Consequently, the noise power spectral density rises most prominently at low frequencies (10–100 Hz), where traps with τ on the order of 0.01–0.1 s induce significant slow current oscillations. By contrast, higher-frequency noise (around 1000 Hz) is dominated by shallower, fast traps that were already active at lower T, so that portion of the spectrum shows a comparatively smaller change. These findings shed light on the intricate interplay of different subgap DOS in the SZTO TFTs. Elucidating such characteristics not only provides fundamental insights into the device physics of amorphous oxide semiconductors but also offers critical guidelines for reliability enhancement in thermally demanding applications, such as displays and next-generation NAND and DRAM, where effective thermal management is crucial37,38.

Gas-sensing properties of SZTO TFT-type gas sensors

Now, we investigate the gas-sensing properties and effects of LFN on the SNR of the SZTO TFT-type sensors. The chemoresistive behavior of the SZTO‑TFT to NO2 gases is first quantified by measuring the change of ID under controlled flows of NO2 gas. The measurements were carried out at temperatures chosen to accelerate both adsorption and desorption while remaining below the point at which irreversible drift becomes appreciable. NO2 concentrations of 100, 250, 500 and 1000 ppb were supplied through a calibrated mass‑flow system; the total flow rate and absolute humidity (4% RH at 27 °C) were held constant to eliminate secondary influences. After the exposure of NO2 gas, the transfer curves are continuously measured to anlayze the transient gas-sensing properties. The reproducible positive increase of the Vth is observed, whose magnitude increased monotonically with the analyte concentration. The sensing mechanism of the SZTO TFT-type gas sensor can be explained by an electron‑withdrawal mechanism: oxidizing NO2 gas molecules physisorb on the semiconductor surface and capture conduction‑band electrons according to

$${{\rm{NO}}}_{2}+{e}^{-}\to {{{\rm{NO}}}_{2}}^{-}$$
(6 )

while pre-adsorbed oxygen species participate through

$${{{\rm{O}}}_{2}}^{-}+{{\rm{NO}}}_{2}+2{e}^{-}\to {{{\rm{NO}}}_{2}}^{-}+2{{\rm{O}}}^{-}$$
(7)

Both processes collectively contribute to the depletion of conduction electrons within the SZTO channel, effectively shifting its electrostatic balance toward electron deficiency, as shown in Fig. 5(a). In TFT operation, the Vth is defined as the minimal gate voltage required to accumulate sufficient electrons at the semiconductor–dielectric interface, forming a conductive channel between the source and drain electrodes. As the electron density within the semiconductor is reduced due to NO2 gas adsorption, a greater positive gate voltage becomes necessary to replenish the electron population to levels adequate for channel conduction. Consequently, the Vth of the TFT shifts positively.

Fig. 5: Gas Sensing Properties of SZTO TFTs.
Fig. 5: Gas Sensing Properties of SZTO TFTs.The alternative text for this image may have been generated using AI.
Full size image

a Schematic illustration of the gas-sensing mechanism in SZTO TFT sensors to NO2 gas. Adsorption of NO2 molecules and interaction with conduction electrons in the channel lead to electron depletion, causing a positive shift in Vth. b Transient sensing responses of SZTO TFT gas sensors at different NO2 concentrations (100–1000 ppb) at a fixed operating temperature of 75 °C. The sensor exhibits slow recovery (~700 s), which is significantly improved by applying short negative Vrecov (−10 V, 100 ms duration, repeated 10 times). c Gas-sensing responses of SZTO TFT sensors measured at various operating temperatures (25–75 °C) and NO2 concentrations. The magnitude of response increases with temperature due to enhanced adsorption kinetics

Furthermore, systematic analyses of the SZTO TFT gas-sensing performance as a function of gas concentration and operating temperature were conducted. Figure 5(b) shows the concentration-dependent response of the SZTO TFT-type sensor at a fixed operating temperature of 75 °C. The magnitude of the positive shift in the Vth increases monotonically with rising NO2 concentrations ranging from 100 to 1000 ppb, consistent with Langmuir-type adsorption kinetics. Notably, the recovery time of the sensor at this temperature is relatively slow (~700 s), posing a practical limitation. To address this issue, a recovery bias scheme (Vrecov) was introduced at the gate terminal. Specifically, while keeping the drain terminal grounded (VDS = 0 V), short negative gate pulses (Vrecov = −10 V, pulse duration = 100 ms) were repeatedly applied 10 times to facilitate rapid sensor recovery. After 10 cycles of Vrecov application, complete sensor recovery is achieved, which is attributable to the negative recovery bias effectively repelling adsorbed NO2 molecules from the SZTO channel surface. This bias-induced recovery approach significantly enhanced the recovery speed and reproducibility of sensor performance, thereby enabling more energy-efficient operation suitable for low-power sensor applications.

Figure 5(c) presents the NO2 gas-sensing responses at varying operating temperatures for different NO2 concentrations. The sensor response magnitude increases with temperature, which is attributed to the acceleration of adsorption kinetics and enhanced chemisorption coverage of NO2 gas molecules on the SZTO channel surface. It is important to note the magnitude of threshold voltage shift induced by gas adsorption remains relatively constant across different operating regions (subthreshold, linear, and saturation). Consequently, if gas response is evaluated solely in terms of the threshold voltage shift, the sensor can operate effectively regardless of the selected operating region. Alternatively, when the response is evaluated based on current variation (ΔID/ID), the subthreshold region provides the highest sensitivity due to its steep transfer characteristics. Nevertheless, this assessment of sensor performance based purely on response characteristics neglects the critical influence of noise. As clearly demonstrated by our preceding analyses of noise characteristics, significant variations in LFN arise across different operating regions under elevated temperature conditions. Therefore, selecting the optimal operating region for high-temperature sensor operation mandates a comprehensive evaluation that explicitly incorporates both gas-sensing response and temperature-dependent noise behavior to accurately determine the region that maximizes the SNR.

Effects of excess noise on SNR

In gas sensors, the sensing performance is ultimately governed not only by the magnitude of the gas-induced response but also by the noise floor that limits the smallest detectable signal. To mitigate noise, previous studies have mainly pursued material-level strategies, such as employing sensing channels with intrinsically lower 1/f noise9. While these approaches effectively reduce the baseline noise floor, they often diminish gas response by weakening charge transfer or limiting the active interaction area, leading to an inherent trade-off between low noise and high sensitivity17. Moreover, even though previous studies have investigated LFN and SNR characteristics of semiconductor-based gas sensors, few have explicitly addressed the temperature dependence of these properties in their performance evaluations. In previous studies, analyses conducted at room temperature have been directly extrapolated to elevated operating temperatures without accounting for the possibility of significant temperature-induced changes in the noise characteristics9,17. However, such an approach is not appropriate when excess LFN originating from subgap DOS, which exhibits a highly sensitive temperature dependence, is present—as clearly revealed by our preceding LFN measurements (Figs. 3 and 4). Therefore, in this study, we aim to incorporate temperature-sensitive noise characteristics into a comprehensive SNR analysis.

The SNR evaluation explicitly accounts for both the measured gas-induced signal and the intrinsic electronic noise inherent in the device. Since the gas response of the SZTO TFT sensor is relatively slow, the primary noise contribution limiting sensor performance is the fluctuation in threshold voltage due to LFN. Thus, the SNR is determined by defining the sensor’s signal as the change in threshold voltage (ΔVth) induced by exposure to NO2 gas, while the noise (σVth) is derived from the integrated PSD of the gate voltage fluctuations arising from LFN. The magnitude of the σVth is accurately computed using the gate-referred voltage noise spectral density (SVG), which is defined as the SID normalized by the square of the gm2 39,40:

$${S}_{{VG}}=\frac{{S}_{{\rm{ID}}}}{{{g}_{{\rm{m}}}}^{2}}$$
(8)

By integrating SVG across the measurement bandwith, the σVth is obtained by using following relation41,42,43:

$${\sigma }_{{\rm{Vth}}}=\sqrt{{\int }_{{f}_{1}}^{{f}_{2}}{S}_{{VG}}\left(f\right){df}}$$
(9)

where f1 and f2 are the low and high cutoff frequencies for the measurement. Using this calculated threshold voltage noise, the SNR of the sensor is subsequently expressed as the ratio of the measured gas-induced ΔVth to the LFN-induced σVth:

$${\rm{SNR}}=\,\frac{{V}_{{\rm{th}}}}{{\sigma }_{{\rm{Vth}}}}=\frac{{V}_{{\rm{th}},{\rm{response}}}-{V}_{{\rm{th}}}}{\sqrt{{\int }_{{f}_{1}}^{{f}_{2}}{S}_{{VG}}\left(f\right){df}}}$$
(10)

Figure 6(a) shows the constituent parameters determining the SNR—specifically, the gas-induced ΔVth (signal) and the σVth induced by low-frequency 1/f noise—plotted on the y1 and y2 axis, respectively, while the resulting SNR is represented on the y3 axis. Measurements are conducted at a fixed NO2 concentration of 1 ppm and an operating temperature of 75 °C, spanning the subthreshold, saturation, and linear regions of TFT-type gas sensor. The dashed blue horizontal line serves as a baseline reference, denoting the SNR achievable in the ideal scenario without excess LFN in the elevated temperature. Importantly, it becomes evident that significant deviations from this ideal condition arise across specific operational regions. Substantial degradation of the SNR occurs within the subthreshold and saturation regions, attributable to pronounced subgap DOS-induced excess noise, as prominently illustrated by the shaded red region in Fig. 6(a). By contrast, the linear region exhibits comparatively minimal excess noise contributions, thereby maintaining a significantly higher SNR.

Fig. 6: SNR optimization method.
Fig. 6: SNR optimization method.The alternative text for this image may have been generated using AI.
Full size image

a Dependence of SNR on the operating region of SZTO TFT sensors at a fixed NO2 concentration (1 ppm) and temperature (75 °C). ΔVth (signal, y1-axis) and σVth (noise, y2-axis) are plotted along with the resulting SNR (y3-axis). The dashed blue line represents the ideal SNR without excess noise at elevated temperature. Significant degradation of SNR occurs in subthreshold and saturation regions due to subgap DOS-induced excess noise (highlighted in red), whereas the linear region exhibits minimal excess noise and highest SNR. b SNR versus NO2 concentration plots for the three operating regions, used to determine LoD, defined as SNR = 3. LoD varies by nearly an order of magnitude depending on the operational region chosen: linear region achieves lowest LoD (0.36 ppb), whereas the saturation and subthreshold regions exhibit substantially higher LoDs of 0.97 ppb and 2.28 ppb, respectively

In Fig. 6(b), these insights are further contextualized by plotting the measured SNR against varying NO2 concentrations (100–1000 ppb) for each distinct operating region. This plot allows precise extraction of the LoD—defined as the NO2 concentration at which SNR equals 3—as indicated by the intersection of dashed extrapolation lines. Remarkably, despite utilizing identical sensor devices, sensing materials, and experimental conditions, the LoD demonstrates substantial variability solely based on the chosen operational region. Specifically, the linear region achieves an exceptionally low LoD of approximately 0.36 ppb, whereas the saturation and subthreshold regions exhibit considerably higher LoD values of approximately 0.97 ppb and 2.54 ppb, respectively. This nearly tenfold disparity in LoD unequivocally emphasizes that sensor performance metrics, particularly at elevated operating temperatures, cannot be evaluated purely by response magnitude alone. Rather, it is essential to explicitly account for the region-specific excess noise characteristics and their profound influence on the sensor’s practical detection limit. By rigorously incorporating temperature-dependent noise characteristics into the SNR analysis, we reveal the critical importance of operating-region selection as a previously underappreciated yet decisive factor for achieving optimal sensor sensitivity and reliability.

We identify the optimal operating condition for the SZTO TFT gas sensor as the linear regime with VGS biased above Vth at an elevated temperature, where the NO2 response is enhanced while the thermally activated excess 1/f noise remains limited, resulting in the highest SNR and lowest LoD. More generally, our approach provides a practical design guideline for TFT-based gas sensors: (i) measure LFN and gas response as a function of VGS and temperature, (ii) convert the measured PSD into SVfb to evaluate σVth, and (iii) derive SNR and LoD over the accessible bias–temperature space. By constructing such SNR-based operation maps, the bias and temperature conditions can be systematically optimized for a given material stack and device geometry, without sacrificing the intrinsic response of the sensing channel.

In light of the recent trends toward device miniaturization and enhanced reliability, LFN has emerged as a critical technological bottleneck in advanced semiconductor devices, including logic44,45, memory46,47, neuromorphic computing48,49,50, stochastic computing41,51,52 and sensor technology53,54. In this context, this study holds significance in that we systematically investigate the impact of LFN within the most prominent gas-sensing platform, correlating noise behavior with gas response. This integrated analysis framework provides a deeper understanding of how intrinsic noise mechanisms fundamentally constrain sensor performance, thereby offering valuable guidance for the development of next-generation, high-reliability gas-sensing platforms

Reliability of SZTO gas sensors

We further investigate the long-term reliability, humidity dependence, and repeatability of the SZTO TFT gas sensors under repeated operation. First, long-term reliability was evaluated by monitoring the gas response and LFN of representative devices over an extended period of six months (Fig. 7(a)). During this period, the baseline electrical parameters as well as the response and normalized noise level at a representative frequency remained within a narrow variation window. These results indicate that neither the sensing reaction nor the associated trapping/detrapping processes induce significant irreversible degradation in the SZTO channel, confirming good operational robustness of the TFT gas sensor platform.

Fig. 7: Reliability of SZTO TFTs.
Fig. 7: Reliability of SZTO TFTs.The alternative text for this image may have been generated using AI.
Full size image

a Long-term reliability of gas-response monitored over 300 days, showing only minor variation. b NO2 response as a function of relative humidity, indicating modest humidity influence and a gradual decrease in response at higher RH. c Dynamic NO₂ sensing during repeated exposure–recovery cycles, demonstrating highly reproducible baseline levels and peak responses

Second, we examined the effect of ambient humidity on sensing performance and noise by measuring the NO2 response at different relative humidity levels (Fig. 7(b)). Because the devices are operated at high temperature, the overall influence of humidity on the electrical characteristics and noise spectra is modest. Nevertheless, a noticeable reduction in gas response is observed as the humidity increases, which we attribute to partial screening of charge-transfer processes at the oxide surface.

Finally, the repeatability of the SZTO TFT gas sensors was assessed through multiple consecutive NO2 exposure and recovery cycles under identical operating conditions (Fig. 7(c)). The dynamic response curves show highly reproducible baseline levels and peak response amplitudes from cycle to cycle, demonstrating that the sensors can maintain stable and repeatable operation over repeated use. Taken together, these long-term, humidity-dependent, and repeatability results confirm that the proposed SZTO TFT gas sensor exhibits sufficient stability and reliability for practical applications, while still enabling the detailed LFN- and SNR-based analysis developed in this work.

Conclusions

In this study, we systematically investigated the temperature-dependent LFN and its effects on gas-sensing characteristics of SZTO TFT-type gas sensors. We identified a thermally induced excess LFN originating from donor-like subgap states. This excess noise markedly degraded the SNR, highlighting the nuanced interplay between electronic noise, operational bias, and temperature in defining sensor performance. Notably, despite employing identical sensor platforms, sensing materials, and gas conditions, our study revealed nearly an order-of-magnitude variability in the LoD, purely depending on the chosen transistor operating region. Specifically, the linear region offered the lowest LoD (approximately 0.36 ppb), compared with significantly higher LoD values in saturation (0.97 ppb) and subthreshold (2.54 ppb) regions. These results compellingly underscore that gas sensor performance cannot be reliably predicted based solely on sensor response magnitude or simplified noise analyses conducted at ambient temperatures. Instead, accurate predictions of practical performance necessitate a holistic approach, explicitly accounting for thermally induced variations in subgap state dynamics and associated noise characteristics. This research not only provides deeper fundamental understanding of noise processes in amorphous oxide semiconductors but also offers critical guidance for optimizing operational conditions and materials engineering strategies.