A High-Precision Hydrogen Sensor Array Based on Pt-Modified SnO₂ for Suppressing Humidity and Oxygen Interference

Abstract

Humidity and oxygen have significant impacts on the accuracy of hydrogen detection, especially for metal oxide semiconductor sensors at room temperature. Addressing this challenge, this study employs a screen-printed 1 × 2 resistive sensor array made from an identical 1 wt.% platinum-modified tin oxide nanoparticle material. Fabrication variability between the two sensing elements was intentionally leveraged to enhance array output differentiation and information content. Systematic hydrogen-sensing tests were conducted on the array under diverse oxygen and moisture conditions. Three distinct feature types—the steady-state value, resistance change, and area under the curve—were extracted from the output of each array element. These features, integrated with their quotient, formed a nine-feature vector matrix. A multiple linear regression model based on this array output was developed and validated for hydrogen prediction, achieving a coefficient of determination of 0.95, a mean absolute error of 125 ppm, and a mean relative standard deviation of 7.07%. The combined information of the array provided significantly more stable and precise hydrogen concentration predictions than linear or nonlinear models based on individual sensor features. This approach offers a promising path for mass-producing highly interference-resistant, precise, and stable room-temperature hydrogen sensor arrays.

1. Introduction

Hydrogen has a high calorific value upon combustion, and its only byproduct is water, which renders it one of the most promising sources of clean and renewable energy. It has various technical applications, including in fuel cells, industrial processing, and aerospace [1]. However, given the high permeability and flammability of H₂, safety issues must be given high priority during its transportation, storage, and use. Therefore, reliable sensors for monitoring H₂ gas concentrations and providing real-time leak detection are crucial for the widespread adoption of hydrogen energy.

In recent years, various types of H₂ sensors have been reported, including metal oxide semiconductor (MOS)-based resistive sensors [2], thermoelectric sensors [3], optical sensors [4], and surface acoustic wave sensors [5]. Among these, MOS sensors have gained prominence owing to their superior sensitivity, cost-effectiveness, and straightforward design, which mean that they exhibit exceptional utility in the field of portable sensing devices. Among numerous MOSs, SnO₂ is a representative material for gas-sensing, with the most extensive research and well-established theories. The gas-sensing capabilities of SnO₂ have been further enhanced through surface modification [6,7] or the application of nanostructures [8,9]. In H₂-sensing applications, MOS materials typically require modification with noble metals to facilitate the reduction of pre-adsorbed oxygen on the material’s surface by H₂, thereby increasing the sensor’s response and reducing its operating temperature. However, the incorporation of noble metals makes MOS materials highly susceptible to fluctuations in the ambient O₂ concentration and humidity [10]. Related research has reported that the reaction between pre-adsorbed oxygen and water vapor can alter the electrical conductivity and gas-sensing capabilities of the sensing material, leading to instability in the sensor signal [11]. In practical H₂ leak detection applications, the O₂ concentration and humidity in the environment may vary dynamically depending on the weather, climate, location, and time of day; during the process of H₂ leakage, there may also be uncertain fluctuations in O₂ concentration and humidity [12]. These practical issues can seriously affect the accuracy and stability of sensors.

Various strategies have been employed to mitigate the interference with MOS sensors caused by environmental factors [13,14,15]. Kim et al. reported that the humidity dependence of SnO₂ sensors was dramatically reduced by NiO doping, since NiO has a strong affinity toward water, and most water-driven species were predominantly absorbed by the NiO rather than the SnO₂ [13]. Fan et al. doped Co₃O₄ with Pr, which led to remarkable humidity independence because the redox reaction of Pr³⁺/Pr⁴⁺ facilitates the removal of surface hydroxyl groups [15]. Despite these efforts, a single MOS sensor is still unable to eliminate these interferences. In complex gas analysis, the integration of sensor arrays with algorithms provides a robust method to eliminate interference [16,17]. Oh et al. fabricated an array of five In₂O₃ sensors modified with different inert metals [16]. By integrating machine learning algorithms, the array’s sensing performance circumvented interference from humidity and temperature. However, the distinct preparation methods and varying operating temperatures for different sensing materials complicate the manufacturing process of these sensor arrays and increase the challenges in their control, resulting in disadvantages in terms of sensor miniaturization and integration on a single Si substrate. Ji et al. innovatively proposed a method to simplify array structures by using a single commercial MOS sensor with cyclic temperature modulation to emulate array effects [18]. The output of the sensor at different temperatures was reconstructed into multidimensional features, achieving “array virtualization” at the algorithmic level while reducing humidity interference. Currently, to the best of our knowledge, research on effectively reducing humidity and oxygen concentration interference for hydrogen detection through on-chip arrays on the same silicon substrate is limited.

In this work, a 1 × 2 resistive sensor array based on 1 wt.% Pt-modified SnO₂ (Pt-SnO₂) was prepared via a low-cost screen printing technique for the high-precision detection of H₂ under the interference of O₂ concentration and humidity fluctuations. All the sensors in the array were fabricated with the same material, on the same substrate, and in one step, and the inevitable differences between the sensor elements were exploited to form differentiated output signals. Three kinds of features were extracted. The outputs of the two elements and their quotient were combined to form a feature matrix. Finally, a highly accurate H₂ detection model was established through multiple linear regression. The sensor array construction approach and H₂ detection model proposed in this research not only significantly reduce the cost and technical difficulty of array fabrication, but also greatly improve the accuracy and stability of the quantitative analysis of H₂ concentration under the interference of the O₂ concentration and humidity level, enabling high-precision detection of H₂.

2. Materials and Methods

2.1. Sensor Array Fabrication and Experimental Methodology

Figure 1 shows the device preparation process and the research methodology used in this study. All the chemical reagents used for the preparation of sensitive materials were commercially sourced. Isopropanol ((CH₃)₂CHOH) was purchased from Xilong Technology Co., Ltd. (Shantou, China). Ethyl cellulose ((C₁₂H₂₂O₅)_n) was obtained from Sinopharm Chemical Reagent Co., Ltd. (Shanghai, China). Chloroplatinic acid solution (H₂PtCl₆, 8 wt.%, dissolved in H₂O) was purchased from Shanghai McLain Biochemical Technology Co., Ltd. (Shanghai, China). Tin oxide nanoparticles, approximately 90 nm in diameter, were purchased from Shanghai McLain Biochemical Technology Co., Ltd.

First, SnO₂ nanoparticles and H₂PtCl₆ solution were mixed in (CH₃)₂CHOH according to the calculated mass ratio to ensure that the concentration of Pt in the subsequently formed solid sensing layer was 1 wt.%. Then an appropriate amount of (C₁₂H₂₂O₅)_n was added as a binder, and stirred in an agate mortar until a uniform slurry with suitable viscosity for screen printing was obtained. Figure 2 shows an optical photograph of the fabricated array and the characterization results of the sensing materials for each sensing element within the array. In this study, two pairs of single-finger gold electrodes were formed on a Si substrate with a 285 nm thick SiO₂ insulating layer to prepare the 1 × 2 sensor array. Figure 2a shows a schematic of the sensor array without sensing layers. The width of each electrode was 100 μm. The spacing between the two electrodes in each pair was 500 μm, and the spacing between the two pairs of electrodes was 2 mm. The prepared precursor paste was then screen-printed on top of the substrate, and the area of the sensing material was about 500 μm × 500 μm. Notably, the two sensor elements in the array were formed simultaneously; these are denoted as Sensor1 (S1) and Sensor2 (S2). Next, the array was dried on a constant-temperature heating platform at 80 °C for 30 min and annealed in an atmosphere furnace at 350 °C for 1 h. Figure 2b shows an optical photograph of the sensor array with Pt-SnO₂.

Figure 2c,d display magnified local SEM images of the sensing layers of S1 and S2, respectively. The corresponding distribution patterns of O, Sn, and Pt, along with their EDS spectra, are presented in Figure 2e–h, respectively. These characterization results indicate that the Pt atoms were homogeneously distributed in both sensing elements, with an approximate concentration of 1 wt.%. Crucially, although the two elements were fabricated simultaneously, variations in the screen printing process led to differences in Pt loading within their sensing layers: one exhibited 0.80 wt.%, while the other showed 1.09 wt.%. Furthermore, differences in hydrogen-sensing performance characteristics between the two elements, as presented later in this work, originated primarily from this compositional discrepancy.

To achieve highly accurate H₂ concentration prediction in the presence of O₂ concentration and humidity interference, a systematic gas-sensing test was conducted on the fabricated sensor array. The experimental design follows the method of controlling variables, with the H₂ concentration (C_H) ranging from 500 ppm to 2500 ppm (a step of 500 ppm), the O₂ concentration (C_O) ranging from 10% to 30% (a step of 55%), and the relative humidity (RH) ranging from 30% to 55% (a step of 5%). Afterward, according to the process shown in Figure 1, appropriate features were extracted from the raw data and then divided into training and validation sets for the establishment of the H₂-sensing model. A detailed description of the experimental procedure and data processing can be found in the Results and Discussion Section.

2.2. Measurement Setup

Figure 3 shows a schematic diagram of the experimental setup for the H₂-sensing tests under the interference of O₂ and RH. The gas samples were prepared with 0.5% H₂ in N₂, pure O₂ (≥ 99.999%), and pure N₂ (≥ 99.999%). The target gas refers to a gas sample containing H₂ at a specific concentration that was used to measure the H₂-sensing characteristics of the sensing array, whereas the reference gas contained no H₂, allowing the sensing array to enter or recover to its initial state. Throughout the entire test process, the total gas flow into the test chamber was consistently controlled at 200 sccm.

In the experiments involving humidity interference, a dry mixture of H₂ and N₂ (dry H₂/N₂ in Figure 3) was formed by controlling MFC1 and MFC2, while MFC5 controlled the flow of dry pure N₂ (dry N₂ in Figure 3) gas to be the same as the flow of the dry H₂/N₂ gas. MFC3 and MFC4 were controlled to form a dry mixture of O₂ and N₂ (dry O₂/N₂), which was then introduced through a bubbler (Sichuan Shubo (Group) Co., Ltd., Chongzhou, China) to carry water vapor (humid O₂/N₂). Switch 1 regulated the process, blending humid O₂/N₂ with dry H₂/N₂ to create a 200 sccm humid H₂/O₂/N₂ mixture to serve as the target gas, or with dry N₂ for a 200 sccm humid O₂/N₂ mixture to serve as the reference gas. Switch 2 worked in conjunction with switch 1 to control whether the target gas sample or the reference gas entered the test chamber. In this way, the O₂ concentration and RH level in both the target and reference gases were kept consistent. The RH level of the gas introduced into the test chamber was calibrated by a humidity meter, as shown in Figure 3. In the experiments without humidity interference, the H₂/N₂ mixture controlled by MFC3 and MFC4 was not introduced into the bubbler, but was instead mixed directly with dry H₂/N₂ to form the dry target gas, or with dry pure N₂ to form the dry reference gas.

The sensor array was placed on a heating plate in the test chamber, and its operating temperature was maintained at room temperature (27 °C) using a temperature controller to prevent potential temperature fluctuations. During the response phase, a target gas mixture containing certain concentrations of H₂, O₂, and N₂ (balance gas) at a certain RH level was introduced into the test chamber. In the stabilization and recovery phases, interference gases with the same O₂ concentration and RH level as the target gas were introduced. The electrical characteristics of the fabricated sensor array were tested using a Keithley 2636B analyzer (Tektronix Technology (China) Co., Ltd., Shanghai, China). The computer directly controlled the source meter, temperature controller, and humidity meter.

3. Results and Discussion

3.1. H₂-Sensing Properties of the Elements

First, a systematic H₂-sensing test was carried out on the sensor array to collect experimental data, following the process described in the Experimental Section. Throughout the test, all the gas mixtures were controlled to be exposed to the sensor array for 250 s for H₂ response, and the sensor array was then given sufficient time to fully recover after the response phase. Figure 4 shows an example of how each element of the sensor array responded to H₂ concentrations from 500 ppm to 2500 ppm (in steps of 500 ppm) at a fixed O₂ concentration of 15% and an RH of 55%.

Figure 4a,b show the dynamic H₂ response curves of S1 and S2. The operating temperature was maintained at 27 °C. Both sensor elements were biased at 0.5 V. The gas response was calculated via ΔI/I_a, where ΔI = I_g − I_a, I_a is the current of the sensor in the reference gas and I_g is the current in the target gas mixture. The results show that at room temperature, the current response of both sensors S1 and S2 increased significantly with increasing H₂ concentration, and the responses reached 10.27 and 22.93 at 2500 ppm H₂, respectively, which are comparable to the good H₂ responses of existing H₂ sensors [19,20,21,22]. These good H₂ responses can be attributed to the modifying effect of the noble metal. Compared with pure SnO₂, the modification of Pt effectively enhances the sensor’s response to H₂ by promoting the dissociation of H₂ molecules into H atoms and promoting reduction reactions on the surface of SnO₂ nanoparticles [23]. Figure 4c shows the linear fit between the gas response (ΔI/I_a) of both sensor elements and the H₂ concentration (ppm). The slope for S1 is 0.00437, with an R² of 0.99381, whereas the slope for S2 is 0.00978, with an R² of 0.99565. The promising sensitivity and linearity of the sensors provided precise data for subsequent experiments. Moreover, the above results indicate the differences between S1 and S2. Even if the materials and processes used for fabricating the sensor array are identical, there will inevitably be individual differences in each sensor element [12]. For example, differences in film thickness, roughness, etc., can cause individual components to exhibit different gas sensitivities.

3.2. Impact of RH and O₂

Figure 5 shows the dynamic response curves of S1 (a) and S2 (b) in the array to a fixed H₂ concentration of 1000 ppm when the O₂ concentration and RH were varied. Figure 5c,d plot the gas response values (ΔI/I_a) at 300 s extracted from Figure 5a,b for both sensor elements. Since the sensing materials are the same, the effects of RH and O₂ tend to affect both elements in a similar way. The response of both sensor elements decreases when the O₂ concentration increases at a constant RH; on the other hand, an increase in the RH also reduces the current response when the O₂ concentration is fixed. Therefore, it is difficult for either sensor element to yield accurate H₂ concentration prediction results based on the response value alone.

The sensing principle of the Pt-SnO₂ sensor, shown in Figure 6, is used to briefly illustrate the H₂-sensing mechanism and the influence of O₂ and water molecules on the H₂-sensing characteristics of the sensor. As shown in Figure 6, the two electrodes of each sensing element are connected to each other by SnO₂ nanoparticles embedded with Pt. When the sensor is exposed to air at room temperature, O₂ is adsorbed onto the surface of the SnO₂ nanoparticles in the form of O₂⁻ due to the presence of oxygen vacancies in the sensing material, as shown in Equation (1) [24].

$${\mathrm{O}}_2\left(\mathrm{ads}\right)+{\mathrm{e}}^{-}\to {\mathrm{O}}_2^{-}\left(\mathrm{ads}\right)$$

(1)

During this process, electrons transfer from SnO₂ nanoparticles to the adsorbed oxygen species, depleting electrons near the surface, forming a space-charge layer, and causing the conduction band to bend upward from E_C^b to E_C^s. This phenomenon occurs between each nanoparticle, forming a barrier potential of |Vs| that impedes electron migration and determines the overall resistance of the sensing material. When the sensor is exposed to the reducing gas H₂, H₂ molecules dissociate into H atoms via Pt catalysis, react with the adsorbed O₂⁻ on the SnO₂ surface to form water, and release free electrons, as shown in Equation (2) [25].

$$4\mathrm{H}(\mathrm{ads})+{\mathrm{O}}_2^{-}(\mathrm{ads})\to 2{\mathrm{H}}_2\mathrm{O}(\mathrm{gas})+{\mathrm{e}}^{-}$$

(2)

Therefore, H₂ reduces the thickness of the depletion layer on the surface of SnO₂, decreasing the |Vs| and increasing the current. When the sensor is exposed to a humid atmosphere, water molecules adsorb onto the SnO₂ surface, occupying oxygen adsorption sites and reducing the concentration of available adsorption sites [26]. At higher humidity levels, a water film forms on the surface of the sensing material, physically blocking the adsorption of both H₂ and O₂. This phenomena of water vapor poisoning affects the sensor’s performance [27]. On the other hand, when O₂ and H₂ coexist in the environment, a combustion reaction may occur under the catalytic action of Pt, which consumes the H₂ that has been dissociated and adsorbed on the surface of the sensing material to a certain extent, thereby reducing the sensitivity of the sensor. Moreover, when the O₂ concentration and RH level fluctuate at the same time, the chemical reactions on the surface of the sensing material are further complicated.

3.3. Data Analysis

In the experimental design of this study, the H₂ concentrations were set from 500 ppm to 2500 ppm with steps of 500 ppm, the O₂ concentrations from 10% to 30% with steps of 5%, and the RH levels from 30% to 55% with steps of 5%. A total of 150 (5 × 5 × 6) sets of data were obtained. The dynamic response, as shown in Figure 4 and Figure 5, consisted of three phases: the stabilization phase (0–55 s), the response phase (55–305 s), and the recovery phase (after 305 s). Three distinct features were extracted from the stabilization phase and the response phase of the sensor array, including the steady-state value (SS), resistance change (VR), and area under the curve (AUC), as shown in Figure 7. In this study, the SS represents the steady-state resistance of the sensor in the stabilization phase, which equals the average resistance R_a from 0 to 55 s. The VR is defined as R_a/R_g, where R_g represents the sensor’s resistance after the reaction with the target gas, which is calculated as the average resistance between 280 s and 300 s (corresponding to the stage between the two dotted lines in Figure 7). The AUC is the area enclosed between the sensor resistance value and the time axis from 55 s to 300 s.

Based on the above description, three features were extracted for sensors S1 and S2, respectively, to obtain two 3 × 150-dimensional raw feature data matrices. Among them, the 120th group of data (corresponding to a H₂ concentration of 500 ppm, an O₂ concentration of 10%, and an RH of 40%) was obviously abnormal, and was excluded before further data analysis and processing. The remaining 149 sets of data were normalized via Z-score normalization for each feature dimension, resulting in two feature data matrices of 3 × 149 dimensions. Twenty sets of data under four nonextreme O₂ concentration and humidity conditions—15% O₂ + 40% RH, 15% O₂ + 50% RH, 25% O₂ + 40% RH, and 25% O₂ + 50% RH—were selected as the validation set. The remaining 129 sets of data were used as the training set for quantitative analysis modeling. The sensing model was built via multiple linear regression using Equation (3) below:

$$C_{\mathrm{H}}={\alpha }_0+{\displaystyle {\sum }_{i=1}^n{\alpha }_{i}Va{r}_i}$$

(3)

It was fitted with different inputs to determine the fitting coefficients α₀ to α_n, where C_H is the H₂ concentration, Var_i is the i-th dimensional feature (SS, VR, or AUC), and n is the total number of feature dimensions involved in the modeling.

Table 1. Comparison of the quantitative analysis performances of different models.

Number	Inputs	n	R2	MAE (ppm)	RSDm (%)
1	M1	3	0.90	200	14.45
2	M2	3	0.73	278	32.01
3	M1/M2	3	0.16	546	6.22
4	[M1, M2]	6	0.91	190	10.88
5	[M1, M2, M1/M2]	9	0.95	125	7.07
6	[M1, M12, M13]	9	0.94	135	10.57
7	[M2, M22, M23]	9	0.69	241	96.10

shows the quantitative analysis performance of seven multiple linear regression models established with different types of data as inputs. The 3 × 129-dimensional feature data corresponding to elements S1 and S2 are denoted as M₁ and M₂, respectively. Models 1 and 2 took M₁ and M₂ as inputs, respectively. Model 3 took as input the 3 × 129-dimensional quotient data obtained by dividing M₁ by M₂, which helped to eliminate multiplicative noise in the signal. Model 4 took as input the 6 × 129-dimensional feature data combining M₁ and M₂. Model 5 took as input the 9 × 129-dimensional data combining the inputs of Models 1, 2, and 3. For each data point within M₁ and M₂, square and cube transformations were applied to generate the matrices M₁², M₁³, M₂ ², and M₂³. The concatenation of M₁, M₁², and M₁³ subsequently constituted a 9 × 129-dimensional data matrix, which was employed as the input for Model 6. Analogously, the concatenation of M₂, M₂², and M₂³ formed a 9 × 129-dimensional data matrix, which was utilized as the input for Model 7.

The accuracy of the validation results was assessed using the coefficient of determination (R²) and the mean absolute error (MAE), whereas the stability of the validation results was evaluated using the relative standard deviation (RSD). A comprehensive evaluation of the quantitative analysis performance of the model was subsequently conducted. Table 1 lists the R² and MAE values between the predicted H₂ concentration results and the reference values for 20 groups of data in the validation set. RSD_m represents the mean of the RSD values for multiple sample predictions at each H₂ concentration within the range of 500~2500 ppm in the validation set.

Compared with Model 1 and Model 2, Model 1 has a larger R², a smaller MAE, and a lower RSD, indicating that the quantitative analysis performance of the model based on the feature data of S1 is significantly better than that based on S2. Although Model 3 has the minimal RSD among all models, indicating the best stability, the division operation results in the loss of differential information, leading to a catastrophic reduction in accuracy (the R² is much lower than 0.5 and the MAE exceeds 500). Model 4 achieves comparable accuracy to Model 1 while having a lower RSD, indicating that modeling with the combined data from the two sensor elements results in better stability. Model 5 achieves the highest R², the smallest MAE, and the second-best RSD next to Model 3, indicating that the comprehensive utilization of the differentiated information from two sensor elements can simultaneously enhance the accuracy and stability of quantitative analysis, leading to better analysis results. Model 6 outperforms Model 1 in all the performance indicators, indicating a nonlinear relationship between the H₂ concentration and the extracted features. The accuracy indicators of Model 6 are close to those of Model 5, but there is still a significant gap in stability compared with Model 5, indicating that complex nonlinear models also fail to achieve the stability improvement brought by the comprehensive utilization of different sensor information; instead, such complex models may worsen analytical performance. All the performance indicators of Model 7 are lower than those of Model 2, which implies that when the performance of a single sensor is inherently poor, a complex nonlinear model may worsen the analytical performance.

Figure 8 shows a comparison between the values predicted by Models 5, 6, and 7 and the reference values for the five H₂ concentration levels in the validation set. The error bars represent the standard deviation between the predicted values of the validation samples at each H₂ concentration, providing a more intuitive view of the improvement in prediction stability obtained with Model 5.

4. Conclusions

In summary, this study developed a 1 × 2 resistive sensor array based on Pt-SnO₂ nanoparticles for high-precision H₂ detection at room temperature, effectively mitigating interference from fluctuations in the RH and O₂ concentration. The sensor elements were fabricated using the same material via the screen printing technique in a single step, which reduced the process difficulty and production cost. The differential outputs of the sensing elements originated from inevitable variations in the sensitive layer during the fabrication process. Three kinds of features were extracted from the outputs of both sensing elements, including the steady-state value, resistance change, and area under the curve. A feature matrix was then created to combine the features from the two sensing elements and their quotient, and a multivariate linear regression model for H₂-sensing was established. The comprehensive information of the prepared arrays yielded stable and accurate H₂ concentration predictions, outperforming the results from linear and nonlinear models established based on individual sensor elements with the same features. This study provides a new approach to addressing the influence of O₂ and humidity on the H₂-sensing process and the mass industrial production of H₂ sensor arrays.