Introduction

Hydrogen (H2) is regarded as one of the most promising green energies due to its advantages of cleanliness, efficiency and sustainability. Moreover, hydrogen gas is also widely used in aerospace, petrochemical engineering and many other fields. However, due to the wide concentration range in which it can explode (4–75% in air) and low ignition energy (0.02 mJ), hydrogen gas is highly explosive and dangerous during production, storage and transportation1. To improve the safety of hydrogen energy applications, the leakage risk of hydrogen gas needs to be reliably monitored. Developing high-performance hydrogen gas sensors with low-concentration detection limits, wide measurement ranges, and fast responses is highly desirable2.

Hydrogen (H2) gas sensors have been intensively studied for decades. Various types of hydrogen sensors, including metal oxide semiconductor (MOS) sensors3,4,5,6, electrochemical sensors7, work function-based sensors (e.g., Schottky diode sensors, FET sensors)8,9,10,11, and catalytic reaction sensors12, have been developed, and many efforts have been made to improve hydrogen gas detection performance. Generally, oxide semiconductor sensors and electrochemical sensors have high sensitivity, and low concentrations of hydrogen gas can be detected3,13,14,15,16,17,18. However, the applications of these sensors are restricted due to their narrow detection ranges (1–2 orders of magnitude) and slow responses (several seconds up to hundreds of seconds)17,19. In contrast, catalytic reaction-based H2 sensors are promising for flammable gas detection due to their broader detection range. Traditional catalytic reaction sensors (namely, “pellistors”), which consist of alumina beads loaded with a catalyst and a platinum thermistor for heating and temperature measurement20,21, have been hindered by their slow response time (>60 s), detection limit (~0.1%) and high power consumption (hundreds of mW)22. With the rapid development of MEMS technology, pellistor sensors have been developed into miniaturized thermoelectric detection devices, with a significant improvement in both power consumption and sensitivity23,24,25,26,27. By constructing an insulated cavity and a MEMS thermoelectric generator, a developed SiGe-based hydrogen sensor had an extended limit of detection of 50 ppm or lower25,28.

As an alternative type of thermoelectric device, MEMS thermopiles have recently received significant attention in gas sensing applications due to their high sensitivity, high signal-to-noise ratio and low power consumption29,30,31. To date, polysilicon thermopiles loaded with platinum (Pt) catalysts have been typically reported for H2 sensing, with a detection limit of ~10 ppm and a detection range of ~10 ppm-1.5% (more than three orders of magnitude)32,33. However, these characteristics are still not sufficient, as trace hydrogen at an even lower concentration of approximately 1 ppm needs to be detected in many hydrogen gas leakage detection applications.

Thermoelectric hydrogen gas sensors measure the generated heat from the oxidization of hydrogen gas, so the major factors that affect the detection limit are catalysis efficiency and sensitivity to the heat-induced temperature change in the sensing structure. By improving the catalyst efficiency, several works achieved a detection level of ~10 ppm by optimizing the catalyst26,32,33,34,35. In terms of the thermopile temperature sensing element, the reported works have generally used polysilicon thermocouples. The equivalent Seebeck coefficient, i.e., the product of the Seebeck coefficient and the thermocouple number of the thermopile, reflects the temperature detection sensitivity of the thermoelectric device. Due to the relatively low Seebeck coefficient of polysilicon36, polysilicon-based devices have not been reported for the detection of hydrogen gas at levels <10 ppm. To significantly lower the detection limit of thermoelectric hydrogen gas sensors to 1 ppm, it is necessary to replace the polysilicon thermoelectric material with a new material with a much higher Seebeck coefficient, e.g., single-crystalline silicon. It is well known that single-crystalline silicon has a Seebeck coefficient several times higher than that of polysilicon37. Unfortunately, normally, thermopile material has to be deposited on top of a thermal insulating thin film (e.g., suspended silicon nitride), and the deposition of single-crystalline silicon thermopile seems hard to be made on top of the insulating thin film. This may be the reason why single-crystalline silicon has rarely been reported as a thermopile material in previous works.

In this work, we develop a highly sensitive differential thermopile hydrogen gas sensor with MEMS technologies. The sensor consists of two identical, heating-enabled thermopiles, with each thermopile consisting of 54 pairs of thermocouples. More importantly, the thermocouple material is single-crystalline silicon instead of polysilicon. Due to the much higher Seebeck coefficient of single-crystalline silicon, the proposed hydrogen sensors achieve a greatly improved detection limit of 1 ppm. With the Pt NPs@Al2O3 catalyst loaded on the sensing thermopile for hydrogen gas detection, the sensors demonstrate excellent selectivity, uniformity, and long-term stability, thereby holding promise for various hydrogen gas detection applications.

Results

Sensor design and fabrication

As illustrated in Fig. 1a, the sensor is composed of MEMS differential thermopiles, with two identical thermopiles suspended on a thermally insulating diaphragm and a heating voltage applied to control the working temperature. The left thermopile for heat sensing is coated with the catalyst for selective reaction with hydrogen gas. To eliminate the influence of environmental factors, such as the thermal conductivity and flow rate of the gas, the thermopile on the right side is designed for reference and compensation. The thermopile has an area with a diameter of 640 µm and consists of 54 pairs of single-crystalline silicon thermocouples in series. Fifty-four N-type thermocouple beams and 54 P-type thermocouple beams are arranged alternately, and each pair of adjacent N-type/P-type thermocouple beams (N-type on the left and P-type on the right) are directly connected at the hot end. At the cold end, they are connected to another adjacent thermocouple pair. The single-crystalline silicon thermopile is suspended at the backside of a low-stress silicon nitride adiabatic support membrane and thermally isolated from the silicon substrate via an etched air cavity. The hot junctions of the thermocouples are uniformly distributed in a 240 μm-diameter region at the center of the suspended membrane, which is called the sensing region. Moreover, the cold junctions are connected in parallel to a silicon frame heat sink. A single thermocouple can detect the temperature difference between the hot junction and the cold junction so that this thermopile detects the average temperature difference between the sensing region and the environment. At each thermopile area, a heating resistor pattern around the sensing region is designed for heating the membrane to the desired temperature for the catalyst. Because of the suspended membrane structure of the thermopile, the heat is confined to the vicinity of the heating resistor, thereby significantly reducing the power consumption of the sensor. In between the two thermopiles, a Pt thermistor with a serpentine pattern is designed to detect the ambient temperature. In addition, a metal guard ring around the entire device is also designed for noise shielding.

Fig. 1: Structural design and working principle of the MEMS differential thermopile H2 sensor.
figure 1

a Schematic diagram of the MEMS differential thermopile-based H2 sensor, consisting of a sensing thermopile and a reference thermopile. b Schematic diagram of the H2 sensing mechanism and working principle of the sensor. The differential output signal is proportional to the temperature change due to the H2 catalytic reaction process, while environmental interferences are subtracted

Figure 1b illustrates the operation of the sensor. The catalytic sensing material is loaded in the sensing region. The sensing regions of both thermopiles are heated to the operating temperature by supplying the same heating voltage. At the operating temperature, the catalyst dissociates H2 into H atoms on the surface5,38 and generates surface hydroxyl OH groups with dissociated O2. The surface hydroxyl groups are further oxidized to produce water39. The reaction of H2 and O2 releases heat, thereby increasing the output voltage of the sensing thermopile40. In the meantime, the reference thermopile only responds to the original temperature and the common-mode disturbance caused by environmental change. The differential output of the two thermopiles reflects the specific signal induced by the selectively oxidized hydrogen gas.

According to the Seebeck effect of the thermocouple, the differential output voltage Vdiff of the sensor is expressed as

$$V_{diff} = V_{sensing} - V_{reference} = N(\alpha _1 - \alpha _2)\Delta T_{h2}$$
(1)

where \(V_{sensing} = N\left( {\alpha _1 - \alpha _2} \right)\left( {T_{heat} + \Delta T_{h2} + \Delta T_{cm} - T_{env}} \right)\) and \(V_{reference} = N(\alpha _1 - \alpha _2)(T_{heat} + \Delta T_{cm} - T_{env})\). N is the number of thermocouple pairs, α1 and α2 are the Seebeck coefficient values of the two materials to form the hot junctions, Theat is the temperature generated by the heater, ΔTh2 is the temperature increase generated by hydrogen oxidation, ΔTcm is the temperature change caused by common-mode interference, and Tenv is the cold junction temperature, which equals the ambient temperature. The temperature change caused by common-mode interference is not negligible. Since the thermal conductivity of hydrogen gas is significantly higher than that of air, the sensor loses more heat in the hydrogen atmosphere and cause a temperature change. This is a nonspecific response to the change in gas thermal conductivity. In addition, gas flow rate variation and other factors also produce common-mode interference. Therefore, the reference thermopile for eliminating the common-mode interference is very important in the sensor.

To improve the sensitivity of the thermopile, single-crystalline silicon is selected as the thermoelectric material for constructing the thermocouples. The Seebeck coefficient of single-crystalline silicon is approximately three times that of polysilicon36,37. By using the N-type/P-type single-crystalline silicon thermocouple structure, the Seebeck coefficient of the thermocouple is 6 times that of the traditional N-type polysilicon/metal thermocouple. Because of the low thermal conductivity of the silicon nitride film41, the heat is conducted primarily through the single-crystalline silicon thermocouple beams. Therefore, the thermocouples are optimized into a spiral shape to reduce heat transfer.

Quite different from the reported thermopiles that are routinely fabricated on the surface of dielectric films using polysilicon deposition and backside etching techniques, we take a unique approach to fabricating MEMS thermopiles using the “microholes interetch and sealing” (MIS) micromachining technique based on a (111) silicon wafer42,43. Conventional (100) wafers are rarely used to fabricate single-crystalline silicon thermocouple structures with specific doping concentrations and are typically used to fabricate polycrystalline silicon thermopiles. On the other hand, although SOI wafers can be used to fabricate single-crystalline silicon thermocouple structures, the thickness of the single-crystalline silicon thermocouples is defined by the device layer. Such a process has the problems of a nonuniform thickness over the entire wafer scale and the need for double-sided alignment to release the device from the backside. In the MIS process, the single-crystalline silicon thermocouples are fabricated from (111) silicon, while the thickness of the single-crystalline silicon thermocouples and the depth of the thermal insulation cavity are defined by the depth of the RIE etch, which has good uniformity. Moreover, the designed etch holes and (111) wafers allow us to create the thermal insulation cavity just from the front side without double-sided processing. MIS technology allows the building of complex single-crystalline silicon MEMS structures with a cost-effective, single-sided process using non-SOI wafers, hence greatly improving the uniformity (e.g., device thickness) and lowering the cost of batch-fabricated MEMS thermopile sensors.

As illustrated in Fig. 2, the fabrication steps of MEMS thermopile devices start with a single-side-polished n-type 4-in (111) wafer. Boron and phosphorus ions are implanted at ~50 keV with doses of 8 × 1015 and 1 × 1016 ion/cm2, respectively. Then, low-pressure chemical vapor deposited (LPCVD) SiNx/SiO2 layers with thicknesses of 0.5 μm and 0.4 μm are patterned to form p-type and n-type thermocouple regions, respectively (Fig. 2a). Subsequently, the thickness of the single-crystalline silicon thermocouple is defined by a 4 μm-deep reactive ion etching (DRIE) (Fig. 2b). A 0.4 μm-thick LPCVD SiO2 layer is then anisotropically etched to form the sidewall structures, which protect the thermocouple beams against the final step of wet etching. A 40 μm-deep DRIE defines the depth of the thermally isolated cavity (Fig. 2c). Then, polysilicon is deposited to fill the deep trenches, and the surfaces are smoothed using chemical mechanical polishing (CMP) (Fig. 2d). CMP uses a polishing solution with high selectivity toward polysilicon against SiO2; hence, self-stopping polishing is realized when the SiO2 layer is exposed. After a 1 µm-thick, low-stress SiNx layer is deposited as the supporting layer of the thermocouples, the contact holes for hot and cold junctions and interconnects are made by RIE (Fig. 2e). All the metal interconnects are sputtered and patterned with 40 nm/100 nm/3000 nm-thick Cr/Pt/Au (Fig. 2f,g) and then covered by plasma-enhanced chemical vapor deposition SiO2 with a thickness of 0.5 μm for passivation. Finally, polysilicon is etched in 25% tetramethylammonium hydroxide (TMAH) at 80 °C, and the thermopiles are suspended over insulation cavities due to the wet etching of (111) silicon wafers. Since TMAH etches polysilicon in an isotropic manner and at a faster rate, the etching first proceeds along the vertically etched trenches filled with polysilicon, and then the single-crystalline silicon is removed horizontally. Compared with that using traditional backside etching or etching through release holes, the release time of the large MEMS thermopiles using the MIS process is shortened to ~30 min, achieving better thickness uniformity and avoiding excessive corrosion of the thermocouple lines.

Fig. 2: Fabrication process of the MEMS differential thermopile-based H2 sensor.
figure 2

a Ion implantation and dielectric film deposition. b Thermocouple patterning and thickness definition. c Diffusion and DRIE etching. d Polysilicon deposition and self-stopping polishing. e SiNx deposition and contact hole etching. f Metal sputtering and electrode patterning. g SiO2 deposition and release hole etching. h TMAH etching

Characterization of the differential thermopiles

The fabricated MEMS differential thermopile H2 sensors are characterized using scanning electron microscopy (SEM) and focused ion beam (FIB) microscopy. The Pt NPs@Al2O3 catalyst is characterized by transmission electron microscopy (TEM) and element energy dispersive spectroscopy (EDS), as illustrated in Fig. 3. The size of the whole sensor is 1 mm × 2 mm. Each sensor is wire-bonded to a PCB board (Fig. 3b,c). The Pt@Al2O3 catalyst is uniformly loaded on the sensing region of the sensing thermopile, while the reference thermopile is kept blank. Figure 3d,e shows details of the 54 pairs of thermopiles densely packed within the ~640 µm-diameter suspended SiNx film. The FIB images provide a cross-sectional view of the single-crystalline silicon thermocouples suspended under the SiNx membrane, with a pitch of 3 μm, as shown in Fig. 3f–h. A total of 20 mg of 5 wt% Pt@Al2O3 catalyst is uniformly dispersed in 1 mL of ethylene glycol and loaded onto the sensing thermopiles. As shown in Fig. 3i–k, Pt nanoparticles are uniformly grown on the Al2O3 nanosheets shown in the TEM image, with a diameter of 10–20 nm.

Fig. 3: Images of the MEMS differential thermopile-based H2 sensor.
figure 3

a SEM image of the differential thermopile H2 sensor, showing the sensing thermopile loaded with catalyst, while the reference thermopile is left blank. b, c Optical images of differential thermopiles wire-bonded to a PCB. d, e Optical images of the differential thermopiles, showing the details of hot junctions, cold junctions, thermocouples, heating resistors, sensing output, and reference output electrodes. f–h Cross-sectional view of the suspended thermopiles over the insulation cavity, with magnified views of the thermocouples underneath the supporting SiNx membrane. i, j TEM images showing the morphology of Pt nanoparticles grown on Al2O3 nanosheets. k EDS characterization showing Pt NPs@Al2O3 as the catalyst for H2 detection

Temperature response of the MEMS differential thermopiles

We first calibrate the temperature response of the MEMS differential thermopiles by heating the sensing region (device center) via the heating resistor. We measure the average temperature within the sensing region using a noncontact infrared thermal imager with a spatial resolution of 20 μm. We also validate the temperature response using finite element modeling (Fig. 4a). The 3D model has the same dimensions as the actual device, including 54 pairs of 4 μm-thick single-crystalline silicon thermocouples arranged in a spiral shape, a 1 μm-thick SiNx suspension film, and a 0.4 μm-thick metal layer for contact at the center and edges of the thermopile. A single-crystalline silicon substrate and a 40 μm-thick air cavity between the thermocouples and the substrate for thermal insulation are also considered in the model for both heat transfer through solid and convection in air. In the simulation, solid heat transfer and convective heat transfer on the upper surface of the device are considered in the heat transfer process. We first evaluate the temperature distribution within the sensing region, which shows a temperature variation of just within ±1% (Fig. 4b). We then calibrate the relationship between the heating voltage and the average temperature of the sensing region, as shown in Fig. 4c, which is in good agreement with the simulation results. The temperature sensitivity of the thermopile is obtained by detecting the output signal as a function of heating temperature, which is measured to be ~28 mV/°C, as shown in Fig. 4d. Given the resistance of a single thermopile of ~540 kΩ and a bandwidth of 200 Hz, we calculate the noise equivalent temperature difference of a single thermopile to be 0.047 mK, allowing us to resolve tiny temperature changes due to the ppm-level H2 catalytic reaction in the next experimental section.

Fig. 4: Temperature response of the MEMS differential thermopiles.
figure 4

a Finite element modeling of the MEMS thermopiles considering electrical heating, heat transfer in the device, and convection in air. b Modeled temperature distribution in the sensing region of the thermopile, showing good uniformity (variation within ±1%). c Measured averaged temperature within the sensing region vs. the heating voltage, which is in good agreement with the simulation results. d Measured output voltage of a single thermopile vs. the averaged heating temperature within the sensing region, showing a temperature sensitivity of ~28 mV/°C

Hydrogen sensing performance of MEMS differential thermopiles

As shown in Fig. 5, we test the H2 detection of our MEMS differential thermopiles in air. The sensors are mounted in a test chamber filled with an H2/air mixture at a flow rate of 200 sccm. The gas flow is controlled by the high-end-MEMS intelligent gas distribution system with high-precision mass flow meters (MFC), gas mixing units, and gas switching units. The flow control accuracy is calibrated to <1 sccm using an Agilent flowmeter ADM-G6691A. A DC power supply is used to maintain an optimized working temperature. To determine the optimized heating temperature, we switch the supplied gases to the MEMS differential thermopiles between air and a 1% H2/air mixture, and the differential output (Vdiff) signals are recorded at operating temperatures from 50 °C to 240 °C, as shown in Fig. 5a. We observe that Vdiff first linearly increases with the heating temperature in the range of 50–120 °C and then tends to saturate in the range of 120–240 °C. Therefore, we determine an optimized working temperature of 120 °C for the Pt@Al2O3 catalyst for the H2 reaction considering the balanced output signal and power consumption (Fig. 5b, c). In addition, excessive heating can result in larger thermal noise and reduce the catalyst’s lifetime44. We thus measure the power consumption of a single thermopile to be ~40 mW. In the subsequent experiments, the operating temperature of the sensors is set to 120 °C.

Fig. 5: H2 sensing performance of MEMS differential thermopiles.
figure 5

a–c Sensor response (Vdiff) to 1% H2 at different operating temperatures from 50 to 240 °C. An optimized working temperature of 120 °C is determined, with a power consumption of 40 mW. d–f Differential output vs. different hydrogen concentrations, showing a linear response within an H2 concentration range of 1 ppm to 2%. g Measured sensor response and recovery time (t90) of 1.9 and 1.4 s, respectively. h Zoomed-in view of the sensor response to continuously reduced H2 concentration down to 1 ppm

We then characterize the sensor response (Vdiff) over a range of H2 concentrations from 1 ppm to 2%. As shown in Fig. 5d, e, the amplitude of the differential output is linearly proportional to the H2 concentration in the range of 1 ppm-2%, with an output of 902 mV at 2% H2 and 15 µV at 1 ppm H2. It can be inferred that the 15 µV output is caused by a 0.53 mK temperature increase, validating the high-sensitivity design of the differential thermopile H2 sensors. These results show that the sensor has excellent detection sensitivity across a concentration range that is more than four orders of magnitude. The response and recovery times (t90) when the MEMS differential thermopile sensors are exposed to 0.1% hydrogen are measured to be 1.9 s and 1.4 s, respectively, as shown in Fig. 5g. These values validate the rapid response and recovery of our differential thermopile H2 sensors, which are on par with the state-of-the-art of thermoelectric gas sensors. Figure 5h shows magnified sensor response curves when continuously supplying H2 gases with reduced concentrations from 80 ppm to 1 ppm, suggesting that our differential thermopile sensors can easily detect 1 ppm H2.

Moreover, we evaluate the selectivity, repeatability, uniformity, and stability of our MEMS differential thermopile H2 sensor. To verify the specificity of the sensor to H2, the sensors are tested in various combustible gases, such as carbon monoxide (1%), methane (1%), ethane (1%), and common VOCs, such as ethanol (1%), acetone (1%), and toluene (1%). Please note that these results are compared with the sensor response in 0.1% H2, as shown in Fig. 6a. Overall, our sensors exhibit excellent selectivity against many combustible gases and VOCs. The sensors are also repeatedly tested at 0.1% H2, as shown in Fig. 6b, showing good repeatability. Thanks to the MIS-based MEMS fabrication, the differential thermopile also features good uniformity in terms of the sensor response. We randomly pick 3 sensors and measure their differences in Vdiff amplitude at 5000 ppm H2 varying only within ±2.5%, as shown in Fig. 6c. In addition, we evaluate the sensor stability by keeping a given sensor in an ambient environment and measuring its sensor response (Vdiff) to 1% H2 gas every week for over 2 months. It can be inferred from Fig. 6d that the sensor performance is not degraded in ambient conditions and remains stable with a fluctuation within ±2.5%.

Fig. 6: Selectivity, repeatability, uniformity, and stability of the MEMS differential thermopile H2 sensors.
figure 6

a Measured sensor response to various combustible gases, such as carbon monoxide (1%), methane (1%), ethane (1%), and common VOCs, such as ethanol (1%), acetone (1%), and toluene (1%), compared with the sensor response to 0.1% H2. b Repetitive testing of the same sensor in 0.1% H2. c Sensor responses of three randomly selected devices and their responses to the same H2 concentration varying only within ±2.5%. d Stability of the same sensor every week over 2 months, showing a fluctuation only within ±2.5%. e, f Comparison of the sensor performance of our MEMS differential thermopiles and other reported H2 gas sensors in the literature, where τr is the response time, τd is the recovery time, DR is the detection range with the unit of order of magnitude, and L is the feature size of the device

Overall, our MEMS differential thermopile H2 sensors demonstrate a good detection limit of ~1 ppm, with a fast response and recovery time of only a couple of seconds, across a wide linear detection range from 1 ppm to 2% H2 concentration (more than four orders of magnitude). Compared with state-of-the-art thermoelectric or thermopile-based H2 sensors12,23,25,26,27,28,32,33,35, our devices show an order of magnitude better detection limit, an order of magnitude larger detection range, and comparable response and recovery times, as shown in Fig. 6e, f.

More importantly, as shown in Fig. 6f, semiconductor sensors13,15,17,18 typically have a better than <1 ppm-level detection limit, while the response time is ~10 s, and the detection range is 2–3 orders of magnitude. Work function-based sensors (e.g., Schottky diode/FET sensors)8,9,10,11 show a wide linear detection range of up to four orders of magnitude, while the response time is still typically >20 s. In contrast, our MEMS differential thermopile sensors possess outstanding yet balanced device features and sensing performance (in terms of detection limit, detection range, response and recovery time, and device size) compared to their aforementioned counterparts.

Conclusion

In summary, we design and fabricate a new MEMS differential thermopile H2 sensor. The sensor consists of two identical temperature-controlled thermopiles, which detect the temperature change due to the catalytic reaction of H2 on the sensing thermopile. By using single-crystalline silicon with a large Seebeck coefficient and high-density thermocouples, the thermopiles exhibit a temperature sensitivity of 28 mV/°C and sub-mK-level temperature resolution. The sensors demonstrate an outstanding yet balanced performance with a detection limit of 1 ppm, a wide linear detection range of 1 ppm-2% (more than four orders of magnitude), and a fast response and recovery time of 1–2 s. Moreover, the sensors also have good selectivity to H2, repeatability, and long-term stability. Our MEMS differential thermopile sensors hold promise for trace detection and early warning of H2 leakage in a wide range of applications.