Sensitive and wafer-scale olfactory sensory neurons Authors: Wenjian Zhang, Hegeng Li, Zhixiang Hu, Lanpeng Guo, Jingqiu Chen, Huanhuan Zhang, Rongyu Mao, Hua-Yao Li, Yang Zhang, Huan Liu Published: 2026 ABSTRACT AbstractAccurate and energy-efficient gas sensing remains a key challenge for next-generation artificial olfactory systems, particularly in applications such as unmanned aerial vehicles, wearable electronics, and humanoid robotic actuators. Inspired by the inherent selectivity and low-power information-processing mechanisms of biological olfactory neurons, we report an integrated olfactory sensory neuron (OSN) that simultaneously achieves high sensitivity and selectivity, signal stability, and neuromorphic spiking output. The device combines iodine-passivated colloidal quantum dots (CQDs) as molecular receptor layers with a wafer-level high-electron-mobility transistor (HEMT) transduction platform. The intrinsically low subthreshold swing and high on/off ratio of the HEMT enable efficient amplification of gas-induced charge modulation and allow the direct generation of spike-like signals. The artificial OSN achieves an ultralow detection limit of 0.5 ppb for NO₂ and successfully differentiates NO₂ from NO via principal component analysis (PCA). These results establish a scalable approach for neuromorphic olfactory modules, providing greater degrees of freedom for AI-driven olfaction and enabling compact, high-performance sensing in robotics, environmental monitoring, and healthcare applications. Subject terms: Electrical and electronic engineering, Sensors, Electronic properties and materials Introduction Accurate detection and differentiation of gases and odors are vital across a wide range of fields, including public security, the food industry, air-quality monitoring, and medical diagnostics1–5. Conventional artificial olfaction systems, such as electronic noses, typically rely on arrays of gas sensors combined with pattern-recognition algorithms to classify gaseous stimuli4,6,7. However, these systems often suffer from large form factors and high-power consumption, largely due to the intrinsic size and structural complexity of the sensing units as well as the additional circuitry required for processing high-dimensional sensor outputs8–10. Such limitations hinder their deployment in practical scenarios requiring lightweight and energy-efficient platforms, including unmanned aerial vehicles, wearable electronics, and humanoid robots11,12. In contrast, biological olfactory neurons exhibit inherent molecular selectivity and encode environmental cues as sparse spiking signals, thus providing an ideal bio-inspired template for the next generation of artificial olfactory systems13–16. Semiconductor gas sensors are ideal for artificial olfactory systems due to their low cost, high sensitivity, and ease of integration. However, cross-sensitivity and high operating temperatures lead to significant drawbacks, such as excessive power consumption, large device footprints, complex algorithms, and limited compatibility with on-chip integration17–23. Colloidal quantum dots (CQDs) are ideal for artificial olfactory, they have many dangling bonds as ligands that bind gas molecules23. Their tailored surface chemistry and low-temperature fabrication process properties enable them to selectively bind gas molecules at room temperature24–27. Nevertheless, carrier confinement, interfacial trap states, and complex electronic coupling within CQD solids often lead to severe impedance mismatches between receptor materials and transduction units, hindering stable signal readout28. Moreover, CQD-based sensors tend to undergo surface degradation upon prolonged gas exposure, causing response attenuation and compromising long-term performance29,30. Such interfacial mismatch and structural stability issues are common in functional composite material systems, and rational molecular-level interface engineering has been proven an effective strategy to resolve the trade-off between material performance and structural compatibility in advanced electronic and energy devices31. Importantly, existing CQD sensors lack the capability to generate spike-like outputs that mimic biological OSNs, posing a fundamental barrier to building neuromorphic olfactory systems. Here, we propose an on-chip olfactory sensory neuron (OSN) capable of achieving highly sensitive and selective gas detection at room temperature. By integrating iodine-passivated CQDs as receptor layers with a wafer-level high-electron-mobility transistor (HEMT) platform, gas-induced surface charge reconfiguration in the CQDs is efficiently amplified and converted into measurable neuromorphic spiking signals. This architecture not only ensures compatibility with silicon microfabrication and enables low-power operation but also resolves the long-standing challenges of impedance mismatch and signal degradation in CQD sensors. Our device achieves an exceptionally low limit of detection for NO₂ (0.5 ppb) and further distinguishes NO₂ from NO through principal component analysis (PCA). These results highlight the strong potential of the proposed on-chip OSN for constructing high-performance, scalable, and biologically inspired artificial olfactory systems. Results In biological olfactory systems, olfactory receptor cells in the upper mucous membrane of the nose detect various smells. When odorant molecules bind to specific olfactory receptors located at dendritic tips of OSNs, the olfactory neuron is activated, and electric signals are sent to the olfactory bulb for signal pre-processing before being transmitted to higher regions of the brain (Fig. 1a). In this work, an on-chip OSN is proposed, which employs a gate-sensitive HEMT32–35. PbS CQDs are used as gate sensing material to function as molecular receptors (Fig. 1b). The two cross-sectional structural schematics were shown in Fig. 1b (i, ii), while the GaAs HEMT acts as a transducer to transduce and amplify the sensor signal. This artificial olfactory neuron operates like a biological olfactory receptor cell. When a gas molecule binds to the CQDs, it initiates a receptor potential, which in turn triggers an action potential of the HEMT. The CQD sensing film is spacial separated from the current-carrying two-dimensional electron gas (2DEG) channel36. Then, the source and drain electrodes function as axons (nerve fibers) transmitting the electrical signals to the signal processing unit (corresponding to the olfactory bulb) for signal pre-processing. Design and characterization of the on-chip OSN. a Schematic of the olfactory receptor neuron, illustrating the olfactory cilia and the generation of neuromorphic spiking signals. b Structural schematic of OSN and its two cross-sectional views. c Photograph of a HEMT device array on a 6-in. wafer and optical image of the OSN. d TEM image of the PbS CQDs. e Normalized absorbance of CQDs before and after iodine passivation. f, g SEM cross-section of the OSN The GaAs HEMT in this study was grown by molecular beam epitaxy (MBE) on a GaAs substrate37, with the GaAs epitaxial structure parameters shown in Table S1. A 6-inch wafer-level HEMTs were fabricated and subsequently diced into multiple units, one unit of HEMT chip measuring 7.85×8.5 mm. A top-view of the HEMT sensor chip is shown in Fig. 1c. To protect the transducer from water and oxygen in air, the surface of the chip was passivated with a SiN layer, leaving only the electrode for electrical contact exposed (detailed fabrication processes are described in the “Methods” and Fig. S1). The PbS CQDs were synthesized via a hot-injection method as reported before37. Initially, the CQDs were capped with oleic acid ligands, which allowed for good dispersibility in the octane solvent. However, these ligands impeded electronic transport among CQDs and inhibited gas molecules from interacting with the surface of the CQDs. To facilitate charge transport and promote surface gas adsorption, we employed a solution iodine ligand treatment via a phase-transfer ligand exchange strategy to remove the original oleate ligand. HRTEM characterization estimated the average diameter of the nanocrystals to be 3.43 nm, indicating that the PbS CQDs have a large surface area capable of interacting with incoming gas (Fig. 1d). XRD and XPS characterizations verified that the oleic acid ligands capped on the PbS CQDs were replaced by iodine ligands (Figs. S2 and S3), during which the crystal structure of the PbS CQDs remained unchanged, while the chemical states of Pb and S elements were altered. The optical properties of the CQDs were examined using UV-Vis Spectroscopy (Fig. 1e). A well-defined excitonic absorption peak at approximately 1060 nm was observed, indicating that quantum confinement was adequately preserved after ligand exchange in both the ink and solid thin film. The PbS CQD was deposited on the sensor chip via a single spin coating process. Cross-sectional SEM analysis (Fig. 1f, g) shows a film thickness of approximately 300 nm and clearly delineates the boundaries between the successive nanoscale material layers. To demonstrate the elemental distribution of the OSN, energy-dispersive X-ray spectroscopy (EDS) was employed to characterize the elemental composition in the cross-sectional micro-region. EDS elemental mapping of Pb, S, N, Si, Ga, and As confirmed the structure of the OSN, corresponding to the cross-sectional structure shown in Fig. 1bii and Fig. S4. We obtained a highly uniform sensing film (Figs. S5 and S6), indicating that the CQDs are suitable for simple, low-cost, and scalable applications. The HEMT gas sensor operates in a depletion mode due to the nature of the AlGaAs/InGaAs heterostructure. The ID-VG curve of the HEMT transducer with a gate voltage applied on the extended gate electrode is depicted in Fig. 2a. This device exhibits channel conduction at a gate voltage of 0 V. Its subthreshold swing (SS) can be as low as 75 mV/dec, approaching the ideal SS of 60 mV/dec38. Furthermore, the device exhibits an on/off ratio of up to 106, indicating strong gate modulation of the channel current. These characteristics highlight the device’s potential for application in highly sensitive gate-controlled field-effect transistor gas sensors. The IG-VG curve demonstrates that the HEMT transducer is a Schottky barrier gate field-effect transistor (Fig. S7). In this study, the electronic properties of the sensing film were coupled to the gate potential change by utilizing the Schottky gate of GaAs HEMT. When the gate voltage is applied to the control gate electrode, a potential drop occurs in both the PbS sensing film and the Schottky barrier diode, resulting in a negative shift of the threshold voltage and an increase in the subthreshold swing of the HEMT to 250 mV/dec (Fig. 2a). Figure 2b presents the output characteristic curves of the OSN, showing a saturation current exceeding 4 mA at a gate voltage of 0.5 V which demonstrates a marked improvement in the signal current intensity of the device. The simplified equivalent electrical circuit of the HEMT gas sensor is presented in Fig. 2c. The device comprises two parts: the transistor and a resistive voltage divider, which includes the CQD film resistance RS and the RGS originates from the equivalent resistance of the Schottky barrier of the gate. As the control gate and the extended gate are deposited on the barrier, there is a parasitic resistance. R0 between the control gate and the extended gate due to the Schottky contact. The CQD sensing film serves as a gas-driven switch to control the application of control-gate voltage on the HEMT. Due to the resistive voltage divider, a change in the resistance of the sensitive layer alters the potential of the extended gate electrode by ΔVEG, the drain current ID of the HEMT will change according to Eq. 1, where W and L are the width and length of the gate, μn is electron mobility, C0 is the gate capacitance, and Vth is the threshold voltage of the HEMT39. . Mechanism and simulation of the OSN. a ID–VG characteristics of the OSN with gate voltage applied on the extended gate and control gate. b ID–VD characteristics of the OSN. c Equivalent electrical circuit of the OSN. d Schematic band diagram of the OSN and the 2DEG. e Two-dimensional doping profile of the simulated GaAs HEMT. f ID-VG and IG-VG characteristics simulation. g TCAD simulation of electron current density distribution in the HEMT. h Simulated electron current density from the cutline profiles, the cutline is shown in (g). i Simulated ID–VCG characteristics of the HEMT with different gate contact resistance The potential on the extended gate (VEG) is crucially dependent on the coupling ratio of R0 and RGS. According to Eq. 2, to evaluate the impact of the Rs value and R0 value on the sensitivity of the on-chip OSN, a numerical simulation was performed. In this simulation, the Schottky gate resistance RGS is estimated to be about 30 MΩ when negatively biased according to the IG-VG curve (Fig. S7), which reflects the actual operational status of the OSN and will be further discussed later. We assumed that the variation of RS was estimated to be 0.5RS when exposed to NOX, and the initial voltage of the extended gate was fixed at –0.3 V through the modulation of VCG. Subsequently, we plotted the sensitivity. ΔVEG and control gate voltage VCG against Rs for each R0 value (Fig. S8). Our findings indicate that a higher R0 value can enhance the ΔVEG value. However, when R0 exceeds 1 GΩ, the control gate voltage surpasses 10 V, rendering the gate vulnerable to breakdown under high voltage bias. In this study, the R0 value is approximately 200 MΩ, and therefore, the control gate voltage can operate within a reasonable range lower than 2 V (Fig. S9a), thereby ensuring the sensitivity and stability of the OSN during long-term operation (Fig. S9b). Figure 2d shows the band diagram of the OSN. Since the CQD film is located between the control gate and the extended gate, it does not change the band structure of the Schottky heterojunction and the AlGaAs/InGaAs heterojunction. In this study, the CQD film adopted a flat band model40, and acts as a larger resistance to weaken the VCG effect on the channel. When the gate contact resistance decreases, a more negatively biased gate voltage is applied to the HEMT. We conducted further investigations on the on-chip OSN using Silvaco TCAD simulation, and introduced a gate-sensitive film into the device simulation by defining different gate contact resistances (Rcontact) in the ATLAS simulator (Fig. 2e). The simulated structure from bottom to top is the semi-insulating GaAs substrate, InxGa1-xAs channel layer, AlxGa1-xAs barrier layer, and electrode contact. The molar fractions of In and Al in the InxGa1-xAs channel layer and AlxGa1-xAs barrier layer are 0.2 and 0.3, respectively. The δ-Si doping model in the AlxGa1-xAs barrier layer is approximated by a Gaussian distribution, and the doping center is located 4 nm away from the InxGa1-xAs channel to maximize the simulation accuracy. The gate work function is set to 4.6 eV to match the experimental results. Based on this model, the electrical characteristics of the HEMT are simulated, and the device is a depletion-mode HEMT with a threshold voltage of −0.4 V and a Schottky gate contact (Fig. 2f). Figure 2g presents the simulated 2D e-current density distribution in the HEMT, with the gate voltage biased at –2 V and the drain biased at 0.5 V. When the gate contact resistance changes from 3 × 108 to 2.5 × 108 Ω/μm, a significant decrease in e-current density of the channel is observed. This negative bias depletes the 2DEG channel and decreases the e-current density (Fig. 2h). The simulation results also reveal a clear sensing mechanism of the OSN when the sensing film resistance increases/decreases upon target gas exposure, the threshold voltage of the OSN will shift towards a negative/positive direction (Fig. 2i). The sensor response was defined by ΔI=I−I0, where I0 and I denote the drain current of the OSN in air and target gas, respectively. The typical ID-VCG transfer curves of the OSN in 0–5 ppm NO2 are illustrated in Fig. 3a. As PbS CQD is a p-type semiconductor in this work, the sensing film resistance RS will decrease when exposed to NO2, thereby causing more VCG bias to be applied on the HEMT transducer according to (Eq. 1). In this regard, the CQD sensing film functions as a gas-driven switch that regulates the application of control-gate voltage on the GaAs HEMT. Quantification of the “gas-to-electrical” signal transduction efficiency was achieved through measurements of the gas adsorption-induced changes in surface potential (Fig. S10 and Note S1). The HEMT device exhibits superior charge transport properties, with a transducer function factor (yPbS=9.30×1022V−1s−1m−2) that is 11.6 times greater than those reported for previously studied PbS TFT devices (yPbS=8.04×1021V−1s−1m−2)21. A positive shift in threshold voltage can be observed upon NO2 exposure when the control gate of the HEMT gas sensor is negatively biased, and a substantial amplification of the signal can be achieved when the control gate voltage is around –2 V. As a depletion-mode FET, its drain current is effectively modulated by the VCG, with sensing performance evaluated within the VCG range of −1 to 1 V (Fig. S11). At a forward-biased VCG of 1 V, NO₂ exposure modulates the current by altering the quantum-dot film resistance and 2DEG concentration, though the current change remains relatively small due to device saturation. In contrast, within the subthreshold region (VCG ≤ −1 V), minute gate potential variations lead to significant drain-current fluctuations, resulting in markedly higher sensitivity. The typical ID-VD output curves of the OSN in the range of 0–5 ppm NO2 were further tested while the control gate voltage was fixed at –2 V (Fig. 3b). The OSN demonstrates current responses to NO2 at the microampere scale. In contrast to conventional resistive PbS-I devices25, whose GΩ-range resistance typically limits output currents to the nA level, the current magnitude in this device has been enhanced by a factor of 106. To obtain higher signal current and signal-to-noise ratio, we set the drain voltage at 2 V for the subsequent gas sensing measurement. Sensing performance of the On-chip OSN. ID-VCG transfer curves (a) and ID-VD output curves (b) of the OSN in 0–5 ppm NO2. c, d Time-resolved ID measurement of the OSN upon exposure to NO2: c across concentrations from 62.5 to 25 ppm under different gate voltages; d at a fixed concentration of 1.25 ppm, and (e) and 0–12.5 ppm NO2. f The relationship between the sensor response (ΔID) and the NO₂ concentration (0–12.5 ppm), the inset is the linear fitting of the response from 62.5 to 0.5 ppm NO₂. g Response of the OSN sensor to different target gases. h Sensing performance of the OSN towards 5 ppm NO2 under relative humidity of 0–90% We further investigated the gas response of the OSN under different gate voltages. Figure 3c shows the time-resolved ID of the OSN upon exposure to NO₂ at concentrations ranging from 62.5 ppb to 25 ppm, with the gate voltage set to –2 V, –2.5 V, and –3 V, and a constant drain voltage of 2 V. The OSN operates in different current regimes depending on the applied gate bias. As the gate voltage becomes more negative, the drain current decreases significantly, and the relative change in current becomes more pronounced. The device exhibits the strongest response at a gate voltage of –2 V, which was therefore selected for all subsequent tests. Figure 3d shows the measured drain current of an OSN in response to 1.25 ppm NO2, with the control gate and drain electrode biased at –2 V and 2 V, respectively. The sensor exhibits fast response and recovery properties. The dynamic response of the sensor towards NO2 gas with concentrations ranging from 62.5 ppb to 12.5 ppm is shown in Fig. 3e. We extracted the sensor response ΔI for different concentrations of NO2, as shown in Fig. 3f. Physicochemical interactions between OSN output and NO2 concentration are consistent with the law of mass action again proving that the OSN’s ability to detect trace NO2. The sensor shows a linear response in the low concentration. We fitted the linear regime from 62.5 to 500 ppb and obtained a slope of 72.2 with a fitting quality (R2) of 0.96. Owing to the high signal-to-noise ratio and baseline stability, the limit of detection (LOD) was estimated to be 0.5 ppb from the sensor’s signal processing performance. Detailed calculation parameters are provided in Table S2. To further evaluate the performance of the OSN, the NO2 sensing performance reported in the literature is listed in Table S3. The OSN exhibits a higher response at relatively low gas concentrations, with an actual LOD as low as 62.5 ppb. As mentioned in Fig. 3g, the response of the OSN towards 5 ppm CO, H2, acetone, SO2, NH3, 1.25 ppm HCHO, triethylamine, ethanol, acetaldehyde, acetic acid, ethyl acetate, acetone, p-xylene, Phenol, NO, and NO2. The results demonstrate that the OSN exhibits selectivity to both NO and NO2. The selectivity test real-time response curve of the OSN as shown in Fig. S12. Since the sensitivity and stability of CQDs are crucial factors for the gas sensor performance. Therefore, we conducted further investigations into the ambient stability of the OSN (Fig. 3h). We observed that the sensor’s response slightly increased as the humidity of the environment gas increased. The highest response, approximately 140 μA towards 5 ppm NO2, was obtained under a relative humidity of 70%. Notably, we observed minimal response loss even at an extremely high humidity level of 90% RH, and the baseline fully recovered to its initial value after humidity exposure. Hence, the OSN exhibited good long-term stability and humidity stability over a wide range of 0–90% RH. Benefiting from the inherent stability of the device and the excellent environmental stability of the OSN, the sensor maintains favorable long-term stability even after prolonged exposure to high humidity and high-concentration NO2, as shown in Fig. S13. During an extended test spanning 80 h under continuous conditions of 70% RH and 5 ppm NO2, the baseline current exhibited almost no drift. To discriminate and calibrate these gases using a single OSN. We recorded the drain current ID as a function of gas concentration and time, as depicted in Fig. 4a. It can be observed that the sensing responses of a single OSN towards NO and NO2 are very similar and cannot be easily distinguished. To address this issue, we demonstrated a proof-of-concept artificial olfactory system based on a single OSN. The first step in the general framework of the e-nose is feature extraction, as the sensing signal output by the gas sensor is a high-dimensional time-series array that reflects various parameters such as the species and concentration of the target gas, as well as environmental parameters41. Algorithm-assisted classification of nitrides. a Time-resolved ID measurement of the OSN upon exposure to NO2 and NO at concentrations of 62.5 ppb to 25 ppm. Feature extraction from the time-resolved ID curve (b) and derivative of ID curve (c). d Principal component loadings for five parameters (F1-F5) in PC1 and PC2 were calculated via Singular Value Decomposition, projecting the five-dimensional signals to two principal components, PC1 and PC2. These loadings were then applied to the original data to derive PCA results. e Principal component loadings for sensors F1-F5 in PC1 and PC2. f Two-dimensional principal subspace of the sensor response data. g Heatmap of the five features versus gas concentration. h Predictive performance of the random forest regression model The response curve of the OSN to 5 ppm NO and NO2 is presented in Fig. 4b, where the response is depicted as a change in drain current over time. The response process can be divided into three regions, namely, baseline, response, and recovery. The baseline period corresponds to the time when the chamber is exposed to pure N2, the response phase occurs during exposure to the target gas, and the recovery phase begins when pure N2 flows into the test chamber again. From this response curve, three response features can be extracted, namely, the maximum change in ID (ΔI), area of response (response curve and baseline surrounded between ta and tb, Ares), and the area of recovery (response curve and baseline surrounded by tb and tc, Arec). Furthermore, we calculated the derivative of ID and plotted the pattern in Fig. 4c, from which the differential response maximum (Dermax) and differential response minimum (Dermin) were extracted. These features are associated with the response and recovery speed of the sensor. The response features were extracted from both the ascending (from 62.5 ppb to 25 ppm) and descending (from 25 ppm to 62.5 ppb) cycles. To visualize the gas sensing patterns, we plotted 3D gas sensing patterns in the feature space using ΔI, Ares, Arec, Dermin, and Dermax in Fig. S14. As expected, the OSN of the plots overlapped at low concentrations due to the similar response of the OSN towards NO and NO2 since only one type of receptor material was used. It was found that the two groups of NOX could be more easily distinguished with the addition of Dermin. The improved classification accuracy can be attributed to the difference in the adsorption kinetics of the two gases, which can be explained by the Langmuir adsorption isotherm theory and is correlated with the intrinsic gas sensing mechanisms42. To further distinguish odor signals between NO and NO2, we conducted a classification study based on PCA. PCA served to project the five-dimensional sensor data into two principal components through Singular Value Decomposition (SVD). (Fig. 4d, e) Loading analysis revealed that all factors contribute similarly to PC1, while PC2 is primarily influenced by F4 (0.57) and F5 (0.57). This differential parameter contribution pattern indicates successful capture of NOX chemical signatures through our OSN design. The sensor data are analyzed to extract features that can be evaluated as a whole to eliminate redundancy and to arrive at a description of the target gas by PCA. Figure 4f shows the PCA plot of the first two principal components of the response features, which shows clearly grouping and separation of NO and NO2 at different concentrations. Based on the feature extraction method, we extracted five features from the test curve for each gas at each concentration and performed normalization. The resulting heatmap is shown in Fig. 4g. As illustrated, the values of each feature vary systematically with changes in the concentrations of the two gases. Features F1, F4, and F5 exhibit an overall positive correlation with gas concentration, whereas features F2 and F3 show an overall negative correlation. To evaluate the capability of the sensor data in predicting gas concentrations, we trained a random forest regression model using features extracted from the raw data, the train-test split ratio was 1:1. Furthermore, to provide a more comprehensive assessment of the model’s generalization capability, we have additionally reported the root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics. The results show that for the Random Forest (NO) model, the RMSE is 2.19, MAE is 1.39, and R² is 0.92; for the Random Forest (NO2) model, the RMSE is 2.49, MAE is 1.43, and R² is 0.90. These error distribution metrics indicate that although both models achieve high R² values, Random Forest (NO) demonstrates slightly better concentration prediction accuracy than Random Forest (NO2) (Fig. 4h). These results demonstrate that the model based on this sensor data can predict the concentrations of both gases with high accuracy. Discussion We have successfully developed an on-chip OSN that mimics the structure and functionality of the human OSN. The OSN is fabricated using PbS CQDs as the sensing material for the gate region of an HEMT chip. The CQDs are deposited on the chip through a single spin-coating process without any additional treatment. Due to the HEMT transducer function, the read-out signal current varies at the microampere level, with the current magnitude amplified by a factor of 106, and the device can detect concentrations as low as 0.5 ppb. Moreover, the device exhibits superior long-term stability over a wide range of relative humidity from 0 to 90%. As a proof of concept, the proposed OSN modules can accurately discriminate between the species and concentrations of NO and NO2 through the PCA using adsorption kinetics features extracted from the original response curves. This OSN can be manufactured with high reliability and is compatible with a range of sensitive materials, which makes it a promising candidate for the development of high-performance on-chip artificial olfactory systems. Materials and methods Materials PbS CQDs were synthesized following a previously reported method37: Reagents including lead oxide PbO (99.9%), oleic acid (OA) (99%), and 1-octadecene (ODE) (90%) were purchased from Aladdin, while bis(trimethylsilyl) sulfide (TMS, 95%) was obtained from Sigma. Initially, the lead oleate was prepared by combining PbO (1.8 g), OA (6 mL), and ODE (20 mL) in a three-neck flask under vacuum conditions for 8 h. The mixture was then heated to 90 °C to facilitate the dissolution of PbO. Subsequently, the temperature was raised to 120 °C to initiate the reaction. A solution of TMS (280 μL) in 10 mL of ODE was prepared separately. Upon reaching the target temperature of the lead precursor, the TMS/ODE solution was rapidly injected into the flask under a nitrogen atmosphere. The reaction proceeded for 30 s before the flask was immersed in an ice bath to quench the reaction. After cooling to ambient temperature, the resulting product was subjected to centrifugation and washed twice with acetone. Finally, the purified CQDs were dispersed in octane at a concentration of 50 mg mL−1 for further use. Synthesis and passivation of PbS CQDs The phase-transfer ligand exchange was conducted following established protocols as described in ref. 43. Initially, PbI2 was dissolved in dimethylformamide (DMF). Subsequently, 5 mL of CQD solution in octane (5 mg mL−1) was combined with 5 mL of the ligand solution within a 50 mL centrifuge tube under ambient conditions. The mixture was gently shaken for 1 min, facilitating the complete transfer of CQDs into the DMF phase. The resulting exchanged solution was subjected to three successive washes with octane. Thereafter, 2.5 mL of toluene was introduced to induce precipitation of the CQDs via centrifugation. Following a 10-min vacuum drying period, the CQDs were redispersed in a butylamine/DMF solvent mixture (8:2 v/v) at a concentration of 300 mg mL−1. Fabrication of the OSN The GaAs HEMT in this paper was grown by molecular beam epitaxy (MBE) on a GaAs substrate. The device structure comprised an undoped GaAs buffer layer with a thickness of 500 nm, followed by a 15 nm thick In0.2Ga0.8As channel layer, a 4 nm thick Al0.3Ga0.7As spacer layer, a Si δ-doping layer, a 25 nm thick Al0.3Ga0.7As barrier layer, a 25 nm thick AlAs etch stop layer, and a 50 nm thick GaAs cap layer. To improve the anti-interference ability of the HEMT, plasma enhanced chemical vapor deposition (PECVD) was used to produce a SiN Passivation Layer, and the areas of the electrodes were exposed by using reactive ion etching (RIE) and buffered oxide etch (BOE) solution. Halide ligand-exchanged PbS CQDs were then spin-coated onto the sensor chip at 2000 rpm for 30 s, followed by drying at ambient. Gas sensing measurement The sensors were evaluated within a vacuum chamber probe system (AES-4TH, Beijing Elite Tech Co., Ltd). Electrical properties were characterized using a semiconductor parameter analyzer (Agilent, B1500A). A custom-made gas mixing system was used to regulate the incoming gas at a flow rate of 0.5 L/min. Humidity control of the gas was achieved by passing the carrier gas through a vessel containing 100 mL of water. Different relative humidity levels are achieved by adjusting the flow ratio of air passing through the bubble chamber. For high-humidity conditions, a heating sleeve and appropriate insulation are employed to increase the number of water vapor molecules. By mixing humidified carrier gas with dry carrier gas, the relative humidity of the gas stream can be controlled within the range of 0–100% RH. Real-time monitoring and adjustment of humidity are performed using a probe station. Due to imperfect chamber sealing or measurement errors, the relative humidity displayed on the probe station screen was 0.3% RH even under dry air exposure. Material characterization The size and morphology were characterized by a High-Resolution Transmission Electron Microscope (HRTEM, Tecnai G2 20). A field emission scanning electron microscope (Gemini SEM 300, Zeiss, Oberkochen, Germany) was used to obtain SEM images. Optical Characterizations were carried out by UV-Visible/NIR spectroscopy (UV-Vis Spectroscopy, Shimadzu UV-3600 plus). The chemical composition and elemental valence states were analyzed by X-ray Photoelectron Spectroscopy (XPS, K-Alpha, Thermo Scientific). The crystal structure and phase composition were characterized by Cu-Kα radiation (λ = 1.5418 Å) X-ray diffraction (XRD, SmartLab SE, Rigaku, Japan). The surface morphology and roughness were investigated by Atomic Force Microscopy (AFM, Edge, Bruker). Device simulations Device simulations were conducted utilizing the Silvaco TCAD software. The employed model was grounded in the drift-diffusion framework, wherein Poisson’s equation and the continuity equations for both electrons and holes were solved self-consistently through a finite element method. The doping concentrations of the Si δ-doping layer were 1019 cm−3. To account for the resistance of the gate-sensitive film, varying gate contact resistances were implemented within the ATLAS simulator. Supplementary information Supporting Information Acknowledgements This work was supported by the National Natural Science Foundation of China (No. 61922032, No. 62271222) and the Hubei Provincial Natural Science Foundation of China(No. 2022CFA035). We thank funding support from National Key R&D Program of Hubei (2024BCB042), and Interdiciplinary Research Program of HUST No.5003190015. We also thank "Chutian Talent Program" Science and Technology Innovation Team of Hubei and the Analytical and Testing Center of Hust for support. Author contributions H.L. conceived and supervised the project. W.J.Z. and Z.X.H. designed the experimental protocol and performed the tests. L.P.G. synthesized the materials, and Y.Z. fabricated the devices. H.G.L., Y.Z., and H.L. analyzed the experimental data. J.Q.C. participated in the investigation and contributed to mechanism analysis. R.Y.M. and H.H.Z. performed formal analysis and contributed to the interpretation of characterization data. Z.X.H. carried out the simulations. W.J.Z., Z.X.H., H.G.L., Y.Z., and H.L. co-wrote the manuscript. All authors discussed the results and reviewed the final manuscript. Conflict of interest The authors declare no competing interests. Footnotes These authors contributed equally: Wenjian Zhang, Hegeng Li, Zhixiang Hu Contributor Information Yang Zhang, Email: zhang_yang@semi.ac.cn. Huan Liu, Email: huan@hust.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41378-026-01333-x. References Associated Data Supplementary Materials Supporting Information