On-chip integrated NDIR CO2 gas sensor Authors: Wenbo Wang, Yi Wang, Zhongming Lu, Yanxiang Liu, Tie Li Published: 2026 ABSTRACT This work reports an on-chip integrated non-dispersive infrared CO2 sensor, in which a microelectromechanical system infrared emitter and a thermopile detector are monolithically integrated on a silicon substrate. Portable CO2 monitoring, including wearable respiratory monitoring and portable environmental sensing, imposes stringent requirements on sensor miniaturization and rapid response. Conventional non-dispersive infrared sensors, constrained by discrete optical chambers and board-level packaging, face inherent trade-offs among compact size, fast response, detection range, and stability. To address these challenges, the optical chamber and device layout are co-optimized based on radiative transfer and thermal coupling principles to enhance infrared utilization while suppressing thermal crosstalk. Simulations show that the compact chamber achieves an effective optical path length of 11 mm with an optical coupling efficiency of 91.8%. An orthogonal optical–gas path configuration is further adopted to shorten the gas diffusion distance. The packaged sensor module remains highly compact, with dimensions smaller than 21 mm × 10 mm × 5 mm. Under a driving voltage of 2.5 V and a modulation frequency of 5 Hz, the sensor exhibits a response time of ~3 s, power consumption of 31 mW at a duty cycle of 10%, and a detection limit of 146 ppm, while maintaining stable performance over a wide temperature and humidity range. The on-chip integrated approach enables significant footprint reduction without compromising sensing performance, providing a promising pathway for portable CO2 monitoring. Introduction Carbon dioxide is not only a key metabolic byproduct of the human body but also an important indicator of indoor air quality, ventilation efficiency, and environmental comfort1,2. Variations in CO2 concentration are closely associated with human health, greenhouse gas emissions, and both industrial and residential environments, making rapid, accurate, and highly sensitive detection of CO2 critically important3–6. Among the various CO2 sensing technologies, non-dispersive infrared (NDIR) sensors exhibit inherent spectral selectivity, offering notable advantages in long-term stability, drift resistance, and environmental adaptability compared to semiconductor-based7–10, electrochemical11–14, and catalytic combustion sensors15,16. However, these advantages typically rely on sufficiently long optical path lengths, precise alignment between the emitter and detector, and effective thermal isolation. Consequently, conventional NDIR systems often employ discrete device assemblies, resulting in large sensor volume, high power consumption, and significant limitations in miniaturization and rapid response17,18. With the rapid growth of portable and wearable CO2 monitoring applications, increasingly stringent requirements have been imposed on sensor miniaturization, response speed, and batch consistency. For example, in wearable respiration monitoring, sensors are often positioned close to the face or integrated into face masks, requiring device volumes of only a few cubic centimeters, while the rapid variation of respiratory airflow demands response times below 10 s to capture exhalation–inhalation cycles. In portable environmental monitoring, sensors are frequently deployed in space-constrained environments or mobile platforms, where compact size and low power consumption are essential. In addition, large-scale deployment in smart buildings and industrial process control systems requires high device consistency and long-term stability. However, conventional NDIR sensors typically rely on relatively long optical path lengths and discrete gas chambers to ensure sufficient sensitivity and stability, which inevitably leads to relatively large device dimensions. Consequently, further miniaturization often introduces trade-offs among response speed, optical efficiency, and long-term stability. Therefore, achieving high optical efficiency, rapid response, and stable sensing performance within a compact device footprint remains a key challenge in the development of NDIR CO2 sensors. To address these challenges, extensive efforts have been devoted to optimizing NDIR sensors based on discrete or board-level integrated architectures. For instance, T. A. Vincent et al. developed a microelectromechanical system (MEMS)-based NDIR sensor with an optical path length of approximately 80 mm, achieving a CO2 detection range of 4–5% Vol; however, the system volume still reached tens to hundreds of cubic centimeters with a power consumption of about 350 mW2. Shah et al. proposed a dual-channel NDIR sensor with a detection range of 0–5% Vol, while the device size remained φ20 mm × 25 mm and the power consumption was approximately 400 mW19. Although these studies improved sensing performance by extending the optical path length or optimizing optical structures, the resulting systems still suffer from relatively large dimensions and high-power consumption, making them less suitable for portable and wearable applications. To further reduce device size and power consumption, we previously proposed two board-level integrated NDIR CO2 sensor designs, in which the infrared emitter, gas chamber, and detector were integrated on a single PCB substrate, achieving compact core unit sizes of 10 mm × 10 mm × 6.5 mm and 12 mm × 6 mm × 4 mm, with response times of approximately 14 s and 10 s and power consumption of about 30 mW20,21. Nevertheless, these board-level integrated configurations still rely on millimeter-scale discrete components, and the assembly process inevitably introduces structural deviations and alignment errors, thereby limiting further improvements in system miniaturization and device-to-device consistency. To overcome the limitations of discrete architectures, silicon-based MEMS integration has been proposed as a key approach for achieving miniaturization22. For instance, Zheng et al. fabricated a chip-level silicon integrated hollow waveguide (Si-iHWG) through silicon etching and bonding processes, integrating the optical path and gas chamber within a single structure, with an overall size of 50 mm × 25 mm × 16 mm23. Jia et al. reported a fully on-chip integrated (OCI) CO2 sensor with an optical path length of approximately 3.5 cm, a detection limit of ~750 ppm, and a response time of ~30 s24. Despite the clear trend toward integration, these approaches are constrained by low optical coupling efficiency in silicon cavities, the complexity of integrating narrow-band filters, and thermal crosstalk. Some studies have employed LED sources to replace conventional infrared emitters25, but this often compromises gas selectivity and increases system cost26. Therefore, achieving high optical coupling efficiency, rapid response, and structural stability simultaneously under stringent miniaturization remains a critical challenge for current NDIR CO2 sensors. Based on the above background, this work proposes an OCI NDIR CO2 sensor. In this design, the infrared emitter and thermopile detector are monolithically integrated on a single silicon substrate using MEMS fabrication technology, which maintains low device cost while reducing structural uncertainties and batch-to-batch variations commonly introduced by conventional discrete assembly. A compact gas chamber with high optical coupling efficiency is developed, and an orthogonal configuration of the optical and gas flow paths is introduced to shorten the gas diffusion distance, thereby enabling rapid gas exchange while maintaining efficient infrared energy utilization. In addition, thermal crosstalk between the infrared emitter and detector is mitigated through optimization of the heat distribution and device layout of the sensing chip. Through these design strategies, the proposed OCI architecture enables significant reduction in sensor footprint and response time while maintaining reliable and stable sensing performance, providing a promising approach for compact and high-performance CO2 monitoring in space-constrained applications. Materials and Methods Design In the miniaturization and low-power development of NDIR CO2 gas sensors, energy utilization efficiency, dynamic response, and structural stability have increasingly become critical factors limiting system performance. Conventional discrete NDIR sensors typically employ separately packaged emitters, optical cavities, and detectors. Such assemblies are inherently bulky, and unavoidable assembly tolerances not only limit optical coupling efficiency but also significantly increase the gas diffusion path length, thereby constraining sensor response speed and device consistency. Even under board-level integration, these issues remain difficult to resolve fundamentally from a structural perspective. To address these structural bottlenecks, this work proposes a chip-scale OCI NDIR CO2 sensor. The design achieves co-integration of a MEMS infrared source and a thermopile detector on a single chip, while systematically reconstructing the optical and gas pathways, thereby optimizing optical coupling efficiency, gas exchange efficiency, and device uniformity at the structural level. As illustrated in Fig. 1a, the fundamental operation of an NDIR CO2 sensor can be described as an energy conversion chain of electrical–thermal–optical–thermal–electrical processes. During each operational cycle, the emitter converts electrical energy into thermal energy and emits infrared radiation. Within the gas chamber, the infrared light is absorbed by CO2 molecules, then directed via wavelength-selective filter structures and reflective paths onto the detector surface. The detector converts the incident infrared radiation into an electrical signal, which is then output to the downstream circuitry for quantitative CO2 detection. Therefore, the emitter, gas chamber, and infrared detector constitute the three critical functional units of an NDIR sensor, and their performance directly determines the system’s energy utilization efficiency and detection accuracy.Fig. 1Schematic illustration and exploded view of the OCI NDIR CO2 gas sensor.a Cross-sectional schematic of the OCI NDIR CO2 gas sensor, illustrating the NDIR absorption–based sensing principle; b Exploded view of the core components of the OCI NDIR CO2 gas sensor Schematic illustration and exploded view of the OCI NDIR CO2 gas sensor. a Cross-sectional schematic of the OCI NDIR CO2 gas sensor, illustrating the NDIR absorption–based sensing principle; b Exploded view of the core components of the OCI NDIR CO2 gas sensor Figure 1b illustrates the core structural elements of the OCI NDIR CO2 sensor. The sensor chip is mounted on a PCB substrate integrated with a thermal dissipation array to enhance heat removal during operation. A sapphire optical window is incorporated on the emitter side to effectively isolate the MEMS infrared emitter from the external environment, reducing device aging and improving long-term reliability. On the detector side, a narrowband optical filter with a center wavelength of 4.26 μm is used to selectively transmit infrared light corresponding to CO2 absorption features, ensuring high detection specificity. Additionally, a waterproof and breathable membrane covers the gas chamber to prevent water vapor and particulate contamination, maintaining long-term measurement stability. Regarding energy conversion units, a MEMS infrared source was selected as the emitter, offering a balance in size and power consumption between conventional tungsten filaments and infrared LEDs, while providing excellent batch manufacturability and long-term stability. For infrared detection, a MEMS thermopile detector based on the thermoelectric effect was employed, which requires no cooling, consumes low power, and exhibits a linear response to incident radiation, making it a mainstream choice for NDIR gas sensors27. Both the emitter and detector are fabricated using MEMS technology, providing a consistent process foundation for chip-level integration. The detailed design of the key components of the sensor chip, including the MEMS infrared emitter and thermopile detector, is provided in Supplementary Material, Section 1: Design of key components for sensor chips: MEMS emitter and detector. Simulation To evaluate the gas exchange efficiency of different gas chamber configurations, fluid dynamics simulations were performed using COMSOL Multiphysics. Three gas chamber designs (Design 1–3) with different inlet geometries were investigated, as illustrated in Fig. 2. All three designs were placed within a standardized test enclosure with identical inlet and outlet configurations to ensure comparability across designs. The turbulent flow physics interface was employed, with an inlet velocity of 0.236 m/s, derived from the experimental gas flow rate of 400 sccm through a 6 mm diameter inlet tube:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v=\frac{Q}{A}=\frac{400\times {10}^{-6}/60}{\pi \times {(0.003)}^{2}}\approx 0.236{\rm{m}}/{\rm{s}}$$\end{document}v=QA=400×10−6/60π×(0.003)2≈0.236m/sFig. 2Fluid dynamics simulation results for the three gas chamber designs.a, c, e Three-dimensional velocity distributions of Design 1, Design 2, and Design 3, respectively; b, d, f Cross-sectional velocity distributions at the Y-axis midplane (XZ plane) for Design 1, Design 2, and Design 3, respectively. Regions with local flow velocity below one-third of the maximum velocity are defined as dead zones Fluid dynamics simulation results for the three gas chamber designs. a, c, e Three-dimensional velocity distributions of Design 1, Design 2, and Design 3, respectively; b, d, f Cross-sectional velocity distributions at the Y-axis midplane (XZ plane) for Design 1, Design 2, and Design 3, respectively. Regions with local flow velocity below one-third of the maximum velocity are defined as dead zones A transient solver was used with a total simulation duration of 60 s and a time step of 0.1 s. The computational mesh comprised 1,143,891 elements. All three designs share identical overall dimensions of 16.6 mm × 5.0 mm × 3.7 mm, differing only in inlet geometry. Design 1 features six cylindrical holes on the top surface, and Design 2 features six cylindrical holes on each of two opposing lateral faces, for a total of 12 apertures; the cylindrical holes in both Design 1 and Design 2 have a radius of 0.2 mm. Design 3 features chamfered lateral apertures on two opposing faces, with a single-face open area of approximately 2.97 mm2. The simulated three-dimensional velocity distributions for all three designs are shown in Fig. 2a, c, e, and the corresponding cross-sectional velocity distributions at the Y-axis midplane (XZ plane) are shown in Fig. 2b, d, f. To quantify gas exchange efficiency, a dead zone was defined as the region where the local flow velocity is less than one-third of the maximum velocity within the chamber cross-section, as flow velocities below this threshold contribute negligibly to active gas renewal. The dead zone fractions calculated for Designs 1 through 3 are 58.58%, 19.71%, and 17.24%, respectively. The results indicate that lateral inlet placement (Designs 2 and 3) substantially reduces the dead zone fraction compared with the top-mounted inlet configuration (Design 1), with reductions of 66.3% and 70.6% for Designs 2 and 3, respectively. Among the two lateral designs, Design 3 achieves the lower dead zone fraction of 17.24%, demonstrating that chamfered apertures promote more uniform flow distribution through the optical sensing region compared with cylindrical holes. Compared with top-entry configurations, the orthogonal arrangement of the optical path and gas flow path in Design 3 effectively shortens the gas diffusion distance and reduces the gas replacement time without compromising optical coupling efficiency, thereby significantly improving the dynamic response of the sensor. Based on the optimized Design 3 gas chamber, a comprehensive simulation study was conducted to systematically evaluate the differences in optical, thermal, and mechanical characteristics between the OCI architecture and the board-level integration approach. Detailed simulation methodology is provided in Supplementary Material, Section 2: Sensor optical-thermal-mechanical simulation details. As shown in Fig. 3a(Ⅰ) and 1(b)(Ⅰ), the OCI architecture monolithically integrates the infrared emitter and detector on a continuous silicon substrate, whereas the board-level scheme mounts discrete infrared sources and detectors on a single PCB, achieving only partial miniaturization. While the OCI architecture significantly enhances system integration and structural uniformity, the continuity of the silicon substrate introduces a direct thermal conduction path between the emitter and detector, imposing stricter requirements on thermal management compared with the board-level approach.Fig. 3Multiphysics simulation results of OCI and board-level NDIR CO2 gas sensors.a (I) and (II) present the schematic illustration of the core sensing unit of the OCI NDIR CO2 sensor and the simulated chip temperature distribution at the moment of emitter deactivation, respectively; b (I) and (II) present the schematic illustration of the core sensing unit of the board-level NDIR CO2 sensor and the simulated chip temperature distribution at the moment of emitter deactivation, respectively; c Comparison of normalized temperature distributions for the discrete design, the OCI design, and the on-chip integrated design with a isolation trench; d Effect of emitter position offset on the optical coupling efficiency; e Effective optical path length of the OCI NDIR CO2 gas sensor; f Irradiance distribution on the detector surface; g Temperature distribution of the infrared emitter under operating conditions; h Temperature distribution of the MEMS emitter along the Y–Y’ axis; i Thermal stress distribution of the thin film under thermal excitation; j Temperature distribution of the PCB backplane Multiphysics simulation results of OCI and board-level NDIR CO2 gas sensors. a (I) and (II) present the schematic illustration of the core sensing unit of the OCI NDIR CO2 sensor and the simulated chip temperature distribution at the moment of emitter deactivation, respectively; b (I) and (II) present the schematic illustration of the core sensing unit of the board-level NDIR CO2 sensor and the simulated chip temperature distribution at the moment of emitter deactivation, respectively; c Comparison of normalized temperature distributions for the discrete design, the OCI design, and the on-chip integrated design with a isolation trench; d Effect of emitter position offset on the optical coupling efficiency; e Effective optical path length of the OCI NDIR CO2 gas sensor; f Irradiance distribution on the detector surface; g Temperature distribution of the infrared emitter under operating conditions; h Temperature distribution of the MEMS emitter along the Y–Y’ axis; i Thermal stress distribution of the thin film under thermal excitation; j Temperature distribution of the PCB backplane To investigate the temperature distribution now of emitter shutdown, the core-region temperature profiles of both architectures were simulated, as shown in Fig. 3a(Ⅱ) and b(Ⅱ). In the board-level configuration, the air gap between the source and detector limits lateral heat conduction along the PCB surface, resulting in a temperature profile characterized by high temperatures at both ends and a lower temperature in the middle. In contrast, the OCI architecture conducts heat continuously through the silicon substrate, with temperature gradually decaying from the emitter region toward the detector. This continuous thermal path renders the OCI structure inherently more susceptible to thermal crosstalk, shifting the thermal management strategy from gap-based structural isolation to substrate-mediated heat suppression, and necessitating targeted structural optimization to mitigate thermal interference. For NDIR sensors, the intensity of infrared radiation varies with gas concentration in accordance with the Beer–Lambert law28:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I={I}_{0}\exp (-{kcl})$$\end{document}I=I0exp(−kcl)where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{0}$$\end{document}I0 is the incident light intensity emitted by the source, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}k is the absorption coefficient of the target gas (CO2), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}c is the CO2 concentration, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l$$\end{document}l is the optical path length. However, Eq. (2) describes light attenuation purely due to molecular absorption in an ideal medium, without accounting for optical losses inherent to practical sensor systems. In a real NDIR sensor, the infrared beam must traverse an enclosed gas chamber before reaching the detector, during which a fraction of the optical power is inevitably lost due to imperfect reflectance of the chamber walls, scattering at optical interfaces, and geometric constraints on the beam path. These losses are determined solely by the physical geometry and surface properties of the gas chamber, and are independent of the gas species or concentration. To account for these practical optical losses, an optical coupling efficiency factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha \,$$\end{document}α(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0 < \alpha \le 1$$\end{document}0<α≤1) is introduced as a correction term, defined as the ratio of the optical power reaching the detector to the total optical power that would arrive under ideal lossless conditions. Incorporating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α into Eq. (2) yields the modified expression:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I={\alpha I}_{0}\exp (-{kcl})$$\end{document}I=αI0exp(−kcl) For a fixed optical path length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l$$\end{document}l and a given absorption coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}k specific to CO2, Eq. (3) indicates that the optical intensity reaching the detector is determined solely by the gas concentration \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document}c and the coupling efficiency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α. Therefore, increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α directly enhances the detected optical intensity at any given gas concentration, highlighting the critical role of coupling efficiency in sensor performance optimization. For a thermopile detector, the output voltage is proportional to the incident optical intensity24, allowing Eq. (3) to be equivalently expressed in terms of detector output voltage as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U={\alpha U}_{0}\exp (-{kcl})$$\end{document}U=αU0exp(−kcl)where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U$$\end{document}U denotes the detector output voltage, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${U}_{0}$$\end{document}U0 represents the detector output voltage in the absence of CO2. Note that when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c=0$$\end{document}c=0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U=\alpha {U}_{0}$$\end{document}U=αU0, indicating that even without gas absorption, the coupling efficiency \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α reduces the output relative to an ideal lossless system. This relationship indicates that the detector output voltage exhibits an exponential dependence on the CO2 concentration. According to Eqs. (3) and (4), for a given optical path length and gas concentration, the optical intensity reaching the detector and the corresponding output voltage are solely determined by the sensor’s optical coupling efficiency, highlighting the critical role of coupling efficiency optimization in enhancing sensor performance. Based on the above analysis of thermal conduction and optical propagation mechanisms, it is evident that enhancing optical efficiency relies on constraining the radiation path and concentrating optical energy, whereas mitigating thermal crosstalk necessitates interrupting heat conduction pathways and redirecting thermal flow away from sensitive components. In the OCI architecture, these two physical processes are coupled within the same substrate. Therefore, this study adopts an integrated design strategy for the gas chamber structure and chip layout from the perspective of co-optimization of radiative transfer and thermal coupling. At the optical level, the chamber’s reflective surfaces are optimized and the radiation path is confined to maximize the utilization of infrared energy within the effective gas absorption path. At the thermal level, isolation trenches are introduced within the chip and vertical heat-dissipation channels are constructed to suppress parasitic heat transfer from the emitter to the detector region. This strategy simultaneously improves optical coupling efficiency while effectively reducing thermal crosstalk, providing a unified design framework for highly integrated NDIR sensors. In terms of thermal management, two isolation trenches with dimensions of 1.98 mm × 0.18 mm are introduced within the chip, and a heat dissipation array is implemented on the PCB to facilitate vertical heat conduction. The simulation results shown in Fig. 3c confirm that these measures effectively suppress thermal crosstalk in the OCI architecture, maintaining the detector temperature within the safe operating range. In addition to improved thermal performance, the OCI configuration also offers inherent advantages in structural stability compared with discrete assembly or board-level integration. To quantitatively evaluate the influence of assembly misalignment on optical performance, Fig. 3d illustrates the variation of optical coupling efficiency with emitter displacement. The simulation results reveal that the optical coupling efficiency decreases significantly as the positional deviation of the infrared emitter increases, highlighting the advantage of the OCI architecture in maintaining precise relative alignment between the emitter and detector and mitigating performance degradation caused by assembly errors. From the optical design perspective, the OCI architecture employs a compound parabolic concentrator as the reflective surface of the gas chamber to match the ±45° divergence angle of the MEMS infrared emitter. As shown in Fig. 3e, this configuration enables a compact optical layout while maintaining an effective optical path length of 11 mm. As illustrated in Fig. 3f, the simulation results show that the maximum irradiance on the thermopile surface reaches 8.81 × 108 mW/m², with the infrared radiation primarily concentrated in the central region of the thermopile sensitive area. Assuming an output optical power of 1 mW from the MEMS infrared emitter, integration of the irradiance distribution over the thermopile surface yields an optical coupling efficiency of 91.8%, approximately 9–18 times higher than that of conventional discrete NDIR gas sensors29,30. The actual optical cell coupling efficiency after processing is approximately 87%, which is close to the simulation result. For detailed test results, please refer to Figure S2 in the supplementary materials. Following the optical chamber design, a comprehensive thermal simulation of the sensor was conducted. As shown in Fig. 3g, the MEMS emitter center temperature reaches approximately 820 K under the designed driving conditions, ensuring sufficient infrared emission intensity for CO2 detection, while the temperature of the sensor surface and the region surrounding the thermopile is effectively controlled near ambient conditions. To further characterize the spatial thermal distribution, the temperature profile along the Y–Y’ axis through the center of the MEMS emitter is extracted and presented in Fig. 3h. The results reveal a highly localized thermal distribution: the temperature rises sharply from near-ambient levels to approximately 820 K within the suspended membrane region, and drops steeply back to ambient within a short distance beyond the membrane boundary. This steep thermal gradient confirms that the suspended membrane structure provides effective thermal confinement, concentrating the Joule heating within the active emitter area while limiting heat dissipation toward the surrounding silicon substrate through the membrane itself. To further evaluate the mechanical reliability of the suspended membrane structure under this high-temperature operation, a thermal stress simulation was performed based on the electrothermal simulation results. As shown in Fig. 3i, when the MEMS emitter center temperature reaches approximately 820 K, the maximum von Mises stress in the suspended membrane region is 553 MPa, which remains well below the fracture strength of silicon nitride (10.8–11.7 GPa31), providing a safety factor ranging from 19.53 to 21.16. This confirms that the suspended membrane structure maintains sufficient mechanical integrity under the designed operating conditions. Furthermore, the 820 K hotspot is spatially confined to the suspended membrane region, as evidenced by Fig. 3h, while the rigid bulk silicon substrate remains near ambient temperature due to the thermal isolation provided by the suspended membrane structure. As silicon retains its mechanical properties well below its melting point of 1687 K and its plastic deformation temperature of approximately 900 K, the structural integrity of the bulk silicon chip is not compromised by the localized high-temperature hotspot. Additionally, the pulsed modulation scheme employed in this work ensures that the emitter operates at elevated temperature only during the active phase of each modulation cycle, rather than continuously, which substantially reduces cumulative thermal fatigue accumulation compared with DC operation. Finally, as shown in Fig. 3j, the maximum temperature at the interface between the PCB substrate and the backing plate is approximately 329 K, satisfying the safe operating temperature requirements of standard electronic components. Fabrication To realize the above OCI structure, a fabrication process compatible with MEMS batch manufacturing was designed, as illustrated in Fig. 4a. The main steps of this process are as follows:Substrate preparation: a 4-inch, 400 ± 10 μm thick, N-type (100) silicon wafer was selected as the substrate;Dielectric deposition: 0.3 μm SiO2 and 0.5 μm SiNx layers were deposited via low-pressure chemical vapor deposition (LPCVD) to form dielectric films;Polysilicon deposition: a 0.5 μm polycrystalline silicon (Poly-Si) layer was deposited using LPCVD and doped via ion implantation (50 keV, 9 × 1015 ions/cm2);Thermocouple leg fabrication: the P+Poly-Si layer was patterned to form thermocouple legs, followed by the growth of a 0.12 μm SiO2 insulation layer and sputtering of 0.3 μm Au for the thermocouple electrodes, which was subsequently patterned;Multilayer insulation: a SiO2/SiNx/SiO2/SiNx stack with thicknesses of 0.05 μm/0.15 μm /0.05 μm/0.15 μm was prepared;Resistor fabrication: Ti/Pt layers of 0.04 μm/0.4 μm were sputtered to form the resistive elements, with Ti serving as the adhesion layer;Black platinum deposition: platinum-black was electroplated onto the Pt resistors in both the MEMS emitter and thermopile detector regions to enhance performance;Electrode exposure: the SiO2/SiNx multilayer was etched to expose the thermopile detector electrodes;Reflective layer: 0.03 μm Cr/0.3 μm Au was sputtered as a reflective coating, with Cr as the adhesion layer;Via formation: vias were patterned and etched on the front side;Backside etching and release: dry etching was performed through the backside vias to form thermal dissipation holes, followed by release of the MEMS emitter and thermopile detector areas.Fig. 4Fabrication process and device characterization of the OCI NDIR CO2 gas sensor.a Schematic illustration of the fabrication process of the OCI NDIR CO2 gas sensor; b Photograph of the fabricated 4-inch wafer and the diced sensor chips; c, d Optical micrographs of the as-fabricated MEMS infrared emitter and thermopile detector, respectively Substrate preparation: a 4-inch, 400 ± 10 μm thick, N-type (100) silicon wafer was selected as the substrate; Substrate preparation: a 4-inch, 400 ± 10 μm thick, N-type (100) silicon wafer was selected as the substrate; Dielectric deposition: 0.3 μm SiO2 and 0.5 μm SiNx layers were deposited via low-pressure chemical vapor deposition (LPCVD) to form dielectric films; Dielectric deposition: 0.3 μm SiO2 and 0.5 μm SiNx layers were deposited via low-pressure chemical vapor deposition (LPCVD) to form dielectric films; Polysilicon deposition: a 0.5 μm polycrystalline silicon (Poly-Si) layer was deposited using LPCVD and doped via ion implantation (50 keV, 9 × 1015 ions/cm2); Polysilicon deposition: a 0.5 μm polycrystalline silicon (Poly-Si) layer was deposited using LPCVD and doped via ion implantation (50 keV, 9 × 1015 ions/cm2); Thermocouple leg fabrication: the P+Poly-Si layer was patterned to form thermocouple legs, followed by the growth of a 0.12 μm SiO2 insulation layer and sputtering of 0.3 μm Au for the thermocouple electrodes, which was subsequently patterned; Thermocouple leg fabrication: the P+Poly-Si layer was patterned to form thermocouple legs, followed by the growth of a 0.12 μm SiO2 insulation layer and sputtering of 0.3 μm Au for the thermocouple electrodes, which was subsequently patterned; Multilayer insulation: a SiO2/SiNx/SiO2/SiNx stack with thicknesses of 0.05 μm/0.15 μm /0.05 μm/0.15 μm was prepared; Multilayer insulation: a SiO2/SiNx/SiO2/SiNx stack with thicknesses of 0.05 μm/0.15 μm /0.05 μm/0.15 μm was prepared; Resistor fabrication: Ti/Pt layers of 0.04 μm/0.4 μm were sputtered to form the resistive elements, with Ti serving as the adhesion layer; Resistor fabrication: Ti/Pt layers of 0.04 μm/0.4 μm were sputtered to form the resistive elements, with Ti serving as the adhesion layer; Black platinum deposition: platinum-black was electroplated onto the Pt resistors in both the MEMS emitter and thermopile detector regions to enhance performance; Black platinum deposition: platinum-black was electroplated onto the Pt resistors in both the MEMS emitter and thermopile detector regions to enhance performance; Electrode exposure: the SiO2/SiNx multilayer was etched to expose the thermopile detector electrodes; Electrode exposure: the SiO2/SiNx multilayer was etched to expose the thermopile detector electrodes; Reflective layer: 0.03 μm Cr/0.3 μm Au was sputtered as a reflective coating, with Cr as the adhesion layer; Reflective layer: 0.03 μm Cr/0.3 μm Au was sputtered as a reflective coating, with Cr as the adhesion layer; Via formation: vias were patterned and etched on the front side; Via formation: vias were patterned and etched on the front side; Backside etching and release: dry etching was performed through the backside vias to form thermal dissipation holes, followed by release of the MEMS emitter and thermopile detector areas. Backside etching and release: dry etching was performed through the backside vias to form thermal dissipation holes, followed by release of the MEMS emitter and thermopile detector areas. Fabrication process and device characterization of the OCI NDIR CO2 gas sensor. a Schematic illustration of the fabrication process of the OCI NDIR CO2 gas sensor; b Photograph of the fabricated 4-inch wafer and the diced sensor chips; c, d Optical micrographs of the as-fabricated MEMS infrared emitter and thermopile detector, respectively The fabricated 4-inch silicon wafer and individual diced devices are shown in Fig. 4b, exhibiting intact device structures and well-defined morphologies, thereby validating the stability and feasibility of the designed fabrication process. The MEMS infrared emitter and thermopile detector have device areas of 1.8 mm × 0.78 mm and 1.2 mm × 1.2 mm, respectively, with corresponding released regions of 0.9 mm × 0.9 mm and 1.05 mm × 1.05 mm. To further enhance the system’s energy utilization efficiency, a platinum-black coating was simultaneously applied to the functional surfaces of the MEMS micro-heater and the thermopile detector, as shown in Fig. 4c, d. The highly porous structure and large specific surface area of Pt-black significantly reduce the effective reflectivity of the device surfaces, thereby enhancing mid-infrared radiation emission at the source side and improving absorption efficiency of incident infrared radiation at the detector side. Benefiting from this radiation–absorption synergistic enhancement mechanism, both the thermoelectric response and output voltage amplitude of the detector are increased without additional power consumption, leading to a significant improvement in the system signal-to-noise ratio. The final on-chip integrated CO2 sensor chip has overall dimensions of 12 mm × 3 mm × 0.4 mm, providing a solid structural and process foundation for achieving low-power, high-sensitivity, and fast-response CO2 gas detection. Signal Modulation and Digital Processing The MEMS infrared emitter was driven by a square wave modulation signal at a frequency of 5 Hz with a duty cycle of 10%, corresponding to an average power consumption of approximately 31 mW. The modulation frequency was selected to balance noise performance and power efficiency: operating below 5 Hz would place the signal within the 1/f noise-dominated region, degrading SNR, while a higher frequency would require an increased duty cycle to maintain the same light-on duration per cycle, resulting in higher power consumption. The demodulation was implemented using a synchronous peak-to-peak sampling approach. Within each modulation cycle, the detector output is sampled during both the light-on phase (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{peak}}$$\end{document}Vpeak) and the light-off phase (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{trough}}$$\end{document}Vtrough), and the peak-to-peak value is computed as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{ptp}}={V}_{{perk}}-{V}_{{through}}$$\end{document}Vptp=Vperk−Vthrough Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${V}_{{trough}}$$\end{document}Vtrough captures ambient background radiation, detector dark current, and circuit bias noise in the absence of the modulated emitter, the subtraction operation inherently cancels these common-mode interference components, eliminating the need for a dedicated hardware reference channel while still achieving effective background suppression. The demodulated signal is subsequently processed through a sliding average filter before CO2 concentration calculation. This modulation–demodulation scheme effectively suppresses common-mode background interference without requiring a dedicated hardware reference channel, thereby simplifying the sensor architecture while maintaining reliable signal quality across the target detection range. To ensure measurement accuracy across a wide temperature range, a thermistor-based temperature drift compensation mechanism was implemented, utilizing piecewise linear interpolation across four calibration temperature setpoints ( − 20 °C, 0 °C, 20 °C, and 50 °C). Detailed descriptions of the signal processing algorithm and temperature compensation methodology are provided in Supplementary Material, Section 4: Modulation, Demodulation and Temperature Compensation Algorithm Details Modulation frequency selection. Results and discussion Result Following the fabrication of the sensor chip, the three designed gas chamber structures were fabricated using computer numerical control (CNC) precision machining of copper (Cu) substrates, enabling accurate realization of the designed aperture geometries and internal cavity dimensions. Following machining, the inner surfaces of each gas chamber were mechanically polished to a surface roughness of 0.1–0.2 μm to minimize scattering losses, and a 2 μm gold (Au) thin film was subsequently deposited on all interior surfaces via electroplating to enhance infrared reflectivity. The infrared reflectance of the electroplated Au surface was characterized using a Bruker Vertex 70 V vacuum FTIR spectrometer equipped with a Bruker A513 reflectance accessory, which supports incident angles ranging from 13° to 83°. Measurements were performed at an incident angle of 14°, and the average reflectance over the wavelength range of 2.5–25 μm was determined to be approximately 82.66%, approaching the ideal specular reflection condition assumed in the optical simulation. The fabricated gas chambers are shown in Fig. 5a, c, e for Designs 1 through 3, respectively.Fig. 5Fabricated gas chamber structures and dynamic response characteristics of the three designs.a, c, and e Photographs of the fabricated gas chambers for Design 1, Design 2, and Design 3, respectively, with interior surfaces coated with a 2 μm Au film via electroplating; b, d, and f Dynamic response curves of the sensor integrated with Design 1, Design 2, and Design 3, respectively Fabricated gas chamber structures and dynamic response characteristics of the three designs. a, c, and e Photographs of the fabricated gas chambers for Design 1, Design 2, and Design 3, respectively, with interior surfaces coated with a 2 μm Au film via electroplating; b, d, and f Dynamic response curves of the sensor integrated with Design 1, Design 2, and Design 3, respectively To evaluate the influence of gas chamber geometry on sensor dynamic response, each of the three fabricated gas chambers was integrated with the OCI sensor chip and characterized under identical testing conditions. The measured 90% response times for Designs 1 through 3 are 19 s, 8 s, and 3 s, respectively, as shown in Fig. 5b, d, f. These results are in good agreement with the fluid simulation findings presented in Fig. 2, confirming that the dead zone fraction is a reliable predictive metric for gas exchange efficiency: a smaller dead zone fraction corresponds to faster gas renewal and shorter response time. Design 3, which achieves the lowest dead zone fraction of 17.24%, yields the shortest response time of 3 s, while Design 1, with the largest dead zone fraction of 58.58%, exhibits the slowest response of 19 s. Based on the combined fluid simulation and experimental characterization results, Design 3 was identified as the optimal gas chamber configuration, demonstrating the best overall performance in terms of flow uniformity, dead zone minimization, and dynamic response speed among all evaluated designs. Accordingly, Design 3 was adopted as the gas chamber for the sensor presented in this work. With the sensor chip and gas chamber fabrication complete, a systematic packaging design was subsequently implemented to integrate all components into a fully functional sensor module, as shown in Fig. 6a. A sapphire optical window and a 4.26 μm bandpass filter were mounted vertically above the sensor chip with their planes perpendicular to the chip surface, enabling optical transmission along the direction parallel to the chip surface. This configuration preserves spectral selectivity while significantly reducing the planar footprint. The high level of system integration achieved in this work is best illustrated by the dimensional progression from chip to fully functional module. The bare sensor chip measures 12 mm × 3 mm × 0.4 mm, monolithically integrating the infrared emitter, gas chamber, and thermopile detector within a single silicon die. After incorporating the optical filter and chamber assembly, the core sensing unit expands to 16.6 mm × 5.0 mm × 3.7 mm. The signal processing circuitry — encompassing signal amplification, thermistor-based temperature compensation, ADC sampling, and MCU-based real-time data processing — was integrated on a compact PCB to complete the fully functional sensor module. The complete module, shown in Fig. 6b, is contained within a volume of less than 21 mm × 10 mm × 5 mm, approximately one-fifth that of commercially available NDIR CO2 sensors such as the MH-Z19C, CM1107H, and S88, while delivering full functionality from raw infrared signal acquisition to calibrated CO2 concentration output.Fig. 6Packaging, circuit design, and overall dimensions of the OCI NDIR CO2 gas sensor.a Package design, PCB design and component layout of the sensor; b Optical image of the packaged sensor device with overall dimensions.; c Peak wavelength of blackbody radiation as a function of temperature based on blackbody radiation theory; d Radiative temperature of the infrared emitter in the OCI NDIR CO2 sensor chip as a function of the applied driving voltage Packaging, circuit design, and overall dimensions of the OCI NDIR CO2 gas sensor. a Package design, PCB design and component layout of the sensor; b Optical image of the packaged sensor device with overall dimensions.; c Peak wavelength of blackbody radiation as a function of temperature based on blackbody radiation theory; d Radiative temperature of the infrared emitter in the OCI NDIR CO2 sensor chip as a function of the applied driving voltage According to blackbody radiation theory, the spectral radiance of an emitter increases with temperature across all wavelengths, with the peak emission wavelength shifting toward shorter wavelengths at higher temperatures. As illustrated in Fig. 6c, the peak emission wavelength approaches 4.26 μm — corresponding to the characteristic absorption band of CO2 — as the emitter temperature approaches 681 K. As the operating temperature increases toward and beyond 681 K, the infrared radiance within the CO2 absorption band increases accordingly, resulting in stronger detector signals. However, the target operating temperature should not be determined solely by spectral matching criteria, but should instead be selected based on a balanced consideration of signal intensity, signal-to-noise ratio, and power consumption requirements for the intended application. To characterize the operating temperature of the designed MEMS infrared emitter under the target driving voltage of 2.5 V, the emitter temperature was experimentally measured using a universal stereo microscope. Prior to measurement, the surface of the MEMS emitter was coated with carbon black to enhance emissivity and improve the accuracy of thermal radiation observation. Subsequently, a gradually increasing driving voltage was applied, and the corresponding temperature response was recorded at room temperature, as shown in Fig. 6d. The results indicate that under a driving voltage of 2.5 V, the MEMS infrared emitter reaches a stable operating temperature of approximately 787 K. The long-term thermal stability of the electrodeposited Pt-black coating used to enhance infrared emissivity warrants careful consideration, as thermally induced sintering of its nanoporous structure could degrade surface emissivity over time. According to the well-established sintering behavior of metals, significant sintering of platinum becomes pronounced near the Tammann temperature, defined as approximately half of the absolute melting point32. Given a platinum melting point of 2041 K33, the corresponding Tammann temperature is approximately 1020 K, substantially higher than the experimentally measured operating temperature of 787 K. The pulsed modulation scheme employed in this work further reduces the cumulative thermal exposure of the Pt-black coating compared with continuous DC operation34, thereby mitigating the risk of long-term thermal degradation under the designed operating conditions. To evaluate sensor performance, a complete CO2 gas testing system was constructed (Fig. 7a). Test gases with different CO2 concentrations were prepared using a gas blending method, in which two mass flow controllers regulated the flows of CO2 and high-purity N2 to achieve the desired mixtures. The sensor was placed in a temperature- and humidity-controlled chamber and connected to external signal processing circuitry for real-time data acquisition and analysis (all subsequent tests, except for temperature characteristics, were conducted at room temperature.). During testing, the CO2 concentration was stepped every 3 min while maintaining a total gas flow rate of 400 mL min⁻¹. The thermopile detector output was amplified by 2200×, and the raw voltage response to CO2 concentration steps is shown in Fig. 7b, demonstrating a clear and stable dynamic response over the 0–10% CO2 range. This experimental result also provides a quantitative basis for validating the selected emitter operating temperature of 787 K. The baseline detector output voltage was measured as 2.19 V under pure N2 atmosphere, and decreased to 2.16 V, 2.11 V, and 1.79 V at CO2 concentrations of 500 ppm, 5000 ppm, and 10%, respectively, yielding signal voltages of 0.03 V, 0.08 V, and 0.40 V at the corresponding concentration levels. The baseline noise level, characterized under pure N2 conditions, was approximately 0.7 mV. Accordingly, the SNR was calculated as 43 (32.7 dB), 114 (41.1 dB), and 571 (55.1 dB) at 0.05%, 0.5%, and 10% CO2, respectively. These results demonstrate that a MEMS infrared emitter operating at 787 K with a driving voltage of 2.5 V provides sufficient optical power to ensure robust and reliable CO2 detection with an excellent signal-to-noise ratio across a wide concentration range, from near-atmospheric levels to high concentrations. Within the 0–1% CO2 range, the sensor exhibits a sensitivity of 14.41 μV/ppm. Under these conditions, the limit of detection was calculated as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${LOD}=3{\sigma }_{{noise}}/S$$\end{document}LOD=3σnoise/Swhere, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma }_{{noise}}$$\end{document}σnoise denotes the standard deviation of the sensor output voltage under zero target gas conditions, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S$$\end{document}S represents the sensor sensitivity. In the noise characterization test, the sensor was continuously measured under a pure nitrogen environment for 10 min, yielding a noise voltage of 0.70 mV, corresponding to a limit of detection of approximately 146 ppm.Fig. 7Experimental setup and performance evaluation of the OCI NDIR CO2 gas sensor.a Experimental system established for sensor performance characterization; b Amplified output voltage of the sensor with a gain of 2200; c Dynamic CO2 concentration response after processing by the embedded algorithm; d, e Sensor response characteristics measured at −20 °C and 50 °C, respectively; f Humidity interference test results of the proposed sensor; g Humidity interference test results of the commercial MH-Z19B sensor for comparison Experimental setup and performance evaluation of the OCI NDIR CO2 gas sensor. a Experimental system established for sensor performance characterization; b Amplified output voltage of the sensor with a gain of 2200; c Dynamic CO2 concentration response after processing by the embedded algorithm; d, e Sensor response characteristics measured at −20 °C and 50 °C, respectively; f Humidity interference test results of the proposed sensor; g Humidity interference test results of the commercial MH-Z19B sensor for comparison Embedded algorithms were applied to the raw output to generate real-time CO2 concentration readings (Fig. 7c). Table 1 summarizes the test results, in which each step was calibrated against standard CO2 concentrations and the maximum reading error was calculated to evaluate measurement accuracy. Results indicate that within the 0–10% CO2 range, the sensor achieves an accuracy better than ±(50 ppm + 3% rdg), comparable to or exceeding that of current commercial CO2 sensors.Table 1Annotations of each standard concentration point in Fig. 7b, c and the corresponding detection accuracyMarkABCDEFGHIJKCO2 standard value (ppm)050001000020000300004000050000600007000080000100000Maximum reading error/2.5%1.66%1.14%0.84%0.98%0.86%0.78%0.76%0.81%0.74% Table 1Annotations of each standard concentration point in Fig. 7b, c and the corresponding detection accuracyMarkABCDEFGHIJKCO2 standard value (ppm)050001000020000300004000050000600007000080000100000Maximum reading error/2.5%1.66%1.14%0.84%0.98%0.86%0.78%0.76%0.81%0.74% Annotations of each standard concentration point in Fig. 7b, c and the corresponding detection accuracy With the temperature compensation algorithm applied, Fig. 7d, e and 7(e) show the sensor performance at −20 °C and 50 °C, yielding maximum relative errors of 4.47% and 4.11%, respectively, indicating reliable sensing performance across the tested temperature range. Humidity interference is another critical metric for assessing practical reliability35, as water vapor exhibits broadband infrared absorption that can overlap with the target gas absorption band and introduce systematic measurement errors. To evaluate this, measurements were conducted at 10 discrete relative humidity levels ranging from 4.9% RH to 87.1% RH at room temperature, with a standard CO2 gas mixture of 500 ppm supplied throughout the test and a dwell time of 5 min at each humidity point. For comparison, the commercial MH-Z19B sensor was tested under identical humidity conditions with a dwell time of 7 min per point to accommodate its response time of approximately 120 s. As shown in Fig. 7f, g, the maximum deviation in the mean output of the proposed sensor across the 10 humidity points was 155.66 ppm, while that of the MH-Z19B was 83.87 ppm. The superior humidity immunity of the commercial sensor can be attributed to its dual-channel optical design, which enables common-mode rejection of humidity-induced interference by simultaneously acquiring reference and measurement signals. Incorporating a dual-channel optical architecture to further enhance humidity robustness represents one of the key directions for future development of the proposed sensor. To evaluate the long-term stability of the designed sensor, continuous operation was conducted under a pure nitrogen environment over a period of 7 days. The output voltage and corresponding CO2 concentration readings are presented in Fig. 8a, b, respectively. As shown in Fig. 8a, the sensor chip exhibits excellent output stability throughout the test duration, with a relative standard deviation (RSD) of only 0.061%. Figure 8b further shows that the maximum CO2 reading under pure nitrogen does not exceed 150 ppm, with a 7-day mean value of 26.28 ppm, confirming the outstanding baseline stability of the sensor. It is noteworthy that this experiment also validates the operational reliability of the MEMS infrared emitter under continuous cycling conditions, corresponding to over 3 × 106 modulation cycles (5 Hz × 60 × 60 × 24 × 7), demonstrating that the emitter maintains stable output without degradation. Furthermore, the negligible drift in detector response over this period confirms that the thermal and optical coupling between the emitter and detector remains highly stable following system equilibration.Fig. 8Performance characterization of the OCI NDIR CO2 gas sensor.a Output voltage and b corresponding CO2 concentration readings recorded over 7 days under a pure nitrogen environment; c Selectivity test results for CO2, methane, and ethane at 3% nominal concentration; d Sensor start-up behavior after power-on; e Dynamic response during CO2 concentration switching between 0 and 100,000 ppm; f Output RSD at different CO2 concentrations; g Response consistency among three chips selected from different wafer regions; h Beer–Lambert law fitting curves for the three sensor chips Performance characterization of the OCI NDIR CO2 gas sensor. a Output voltage and b corresponding CO2 concentration readings recorded over 7 days under a pure nitrogen environment; c Selectivity test results for CO2, methane, and ethane at 3% nominal concentration; d Sensor start-up behavior after power-on; e Dynamic response during CO2 concentration switching between 0 and 100,000 ppm; f Output RSD at different CO2 concentrations; g Response consistency among three chips selected from different wafer regions; h Beer–Lambert law fitting curves for the three sensor chips The gas selectivity of the sensor was subsequently evaluated using three test gases — CO2, methane, and ethane — each at a nominal concentration of 3%, with a gas introduction duration of 10 min per gas. As shown in Fig. 8c, the averaged sensor readings for CO2, methane, and ethane are 30,497.18 ppm, 1,734.33 ppm, and 1,233.17 ppm, respectively. The substantially lower cross-responses to methane and ethane demonstrate that the sensor possesses good selectivity toward CO2, attributable to the spectral specificity of the mid-infrared optical filter. The primary absorption bands of methane and ethane are located at approximately 3.3 μm and 3.4 μm, respectively, which lie outside the transmission window of the 4.26 μm bandpass filter, thereby effectively suppressing their contributions to the detector signal. This behavior confirms that the optical filtering mechanism effectively suppresses off-band gas absorption interference, thereby enhancing detection specificity under mixed-gas conditions. In terms of start-up behavior, the sensor was powered on under a nitrogen environment and monitored until stable readings were achieved. The results indicate that the sensor reaches operational readiness within approximately 30 s of power-on (Fig. 8d), which is well suited for practical deployment scenarios requiring rapid initialization and low-latency operation. Compared with conventional NDIR systems that often require a relatively long stabilization period, the present device exhibits a noticeably shorter warm-up duration, which can reduce waiting time before measurement and improve usability in real-time monitoring scenarios. This fast warm-up characteristic is attributed to the low thermal mass of the MEMS infrared emitter, which reaches its target operating temperature within a short time after power-on. For dynamic response characterization, the CO2 concentration was alternately switched between 0 and 100,000 ppm under a constant gas flow. As illustrated in Fig. 8e, the sensor achieves 90% of the full-scale concentration change within 3 s, indicating a fast response capability suitable for most practical application scenarios. The measured response time includes not only the gas diffusion and exchange process within the gas chamber, but also the delay introduced by the digital filtering and signal processing procedures used to improve output stability. Such temporal performance is advantageous for applications involving continuously changing gas environments, including respiratory monitoring, environmental safety assessment, and portable personal protection devices. The repeatability and signal stability of the sensor were further evaluated through the relative standard deviation of the output signals measured at different CO2 concentrations, as summarized in Fig. 8f. The standard deviation of the output voltage was found to be approximately 0.5–0.7 mV, corresponding to a concentration fluctuation of less than 300 ppm. The relatively small variation among repeated measurements demonstrates good consistency and stability of the sensing system during operation. This performance further confirms the effectiveness of the OCI architecture and the vertically coupled cross-flow gas chamber design in maintaining reliable gas sensing performance with reduced signal fluctuation during repeated measurements. Detailed descriptions of the gas sources, sensor calibration procedures, and experimental testing setup are provided in Supplementary Material, Section 5: Other details. To assess inter-device variability and verify wafer-level batch uniformity, three sensor chips (No. 1, No. 2, and No. 3) were randomly selected from the top, middle, and bottom regions of the fabricated wafer and individually packaged for testing. As shown in Fig. 8g, the RSD among the three devices was calculated across the higher concentration range of 4–10% CO2, yielding a maximum inter-device RSD of 0.69%. This result confirms that the device-to-device performance variation across the wafer is minimal, demonstrating the excellent batch fabrication uniformity enabled by the monolithic OCI process. The output voltage of each device was further fitted to the Beer–Lambert law, as presented in Fig. 8h, yielding coefficients of determination R2 of 99.45%, 99.35%, and 99.46%, respectively, confirming strong agreement with the theoretical model and further substantiating the process stability and design robustness of the proposed OCI-based NDIR sensor. The consistency of the fitted R² values across all three devices, despite being sampled from spatially separated regions of the wafer, provides additional evidence that the optical path geometry and coupling efficiency are reliably reproduced across the full wafer area. Furthermore, Table 2 benchmarks the proposed OCI NDIR CO2 sensor against several representative commercial products and recently published research prototypes. The proposed sensor demonstrates clear advantages in terms of power consumption, response time, and overall form factor. Considering the high cost-effectiveness of MEMS-based emitters and infrared detectors, the OCI-based NDIR sensing strategy shows promising potential for portable gas sensing applications.Table 2Performances of cost-effective NDIR CO2 gas sensorsManufacturerModelPrinciple*RangeAccuracyPower consumptionFrequency and duty cycleLODResponse timeDimensions(mm3)n.e.t36IRNEX-P 20 mmIA0–5%±5% FS135 mW––≤30 s (t90)φ20 × 16.8(1)TelAire37T6713IA0–0.5%± (30 ppm + 3% rdg)100 mW1/5 Hz @ //<180 s (t90)30 × 15.6 × 8.6(1)Sensirion38SCD41PA0.04–4%± (40 ppm + 5% rdg)49.5 mW––60 s (t63)10.1 × 10.1 × 6.5(1)Senseair39S88IA0.04–1%± (40 ppm + 3% rdg)–––<30 s (/)33.9 × 19.8 × 8.7(1)Alphasense40IRC-ATIA0–5%1.5% FS~150 mW3 Hz @ /–<40 s (t90)φ20 × 16.5(1)Winsen41MH-Z19CIA0.04–1%± (50 ppm + 5% rdg)–––120 s (t90)32.9 × 19.7 × 8.6(1)CUBIC42CM1107HIA0–5%10% rdg<250 mW1 Hz @ /––33 × 22 × 13.1(1)NiCeRa43CO2-4K-3M-PIA0.04–0.4%± (200 ppm + 5% FS)125 mW1 Hz @ //<180 s (t90)65 × 20 × 10.6(1)ZyAura44ZG09IA0–1%± (50 ppm + 3% rdg)<170 mW//~60 s (/)32.2 × 20.2 × 13.7(1)Qinlin Tan45/IA0-4.8%/////Φ10 × 10(2)Xiaoning Jia24/IA0-10%///750ppm~2.8 s (/)/Zhangxiong Chen46/IA0-0.005%/30 mW5 Hz @ 60%263ppb~2.5 s (/)180 × 100 × 10(2)Liyang Feng21/IA0-10%/33 mW1 Hz @ 10%/10 s (t90)13.7 × 13.1 × 6(1)This workSIMIT-OCI-SIA0–10%± (50 ppm + 3% rdg)31 mW5 Hz @ 10%146ppm~3 s (t90)20.7 × 9.2 × 4.8(1)* The principle here refers to the principle of optical-electrical conversion within the sensor, including infrared absorption (IA) and photoacoustic (PA) types. (1) Including the processing circuit. (2) Gas chamber sizeThe bold values denote the device performance from this work, highlighted to facilitate comparison with the other devices in the table Table 2Performances of cost-effective NDIR CO2 gas sensorsManufacturerModelPrinciple*RangeAccuracyPower consumptionFrequency and duty cycleLODResponse timeDimensions(mm3)n.e.t36IRNEX-P 20 mmIA0–5%±5% FS135 mW––≤30 s (t90)φ20 × 16.8(1)TelAire37T6713IA0–0.5%± (30 ppm + 3% rdg)100 mW1/5 Hz @ //<180 s (t90)30 × 15.6 × 8.6(1)Sensirion38SCD41PA0.04–4%± (40 ppm + 5% rdg)49.5 mW––60 s (t63)10.1 × 10.1 × 6.5(1)Senseair39S88IA0.04–1%± (40 ppm + 3% rdg)–––<30 s (/)33.9 × 19.8 × 8.7(1)Alphasense40IRC-ATIA0–5%1.5% FS~150 mW3 Hz @ /–<40 s (t90)φ20 × 16.5(1)Winsen41MH-Z19CIA0.04–1%± (50 ppm + 5% rdg)–––120 s (t90)32.9 × 19.7 × 8.6(1)CUBIC42CM1107HIA0–5%10% rdg<250 mW1 Hz @ /––33 × 22 × 13.1(1)NiCeRa43CO2-4K-3M-PIA0.04–0.4%± (200 ppm + 5% FS)125 mW1 Hz @ //<180 s (t90)65 × 20 × 10.6(1)ZyAura44ZG09IA0–1%± (50 ppm + 3% rdg)<170 mW//~60 s (/)32.2 × 20.2 × 13.7(1)Qinlin Tan45/IA0-4.8%/////Φ10 × 10(2)Xiaoning Jia24/IA0-10%///750ppm~2.8 s (/)/Zhangxiong Chen46/IA0-0.005%/30 mW5 Hz @ 60%263ppb~2.5 s (/)180 × 100 × 10(2)Liyang Feng21/IA0-10%/33 mW1 Hz @ 10%/10 s (t90)13.7 × 13.1 × 6(1)This workSIMIT-OCI-SIA0–10%± (50 ppm + 3% rdg)31 mW5 Hz @ 10%146ppm~3 s (t90)20.7 × 9.2 × 4.8(1)* The principle here refers to the principle of optical-electrical conversion within the sensor, including infrared absorption (IA) and photoacoustic (PA) types. (1) Including the processing circuit. (2) Gas chamber sizeThe bold values denote the device performance from this work, highlighted to facilitate comparison with the other devices in the table Performances of cost-effective NDIR CO2 gas sensors * The principle here refers to the principle of optical-electrical conversion within the sensor, including infrared absorption (IA) and photoacoustic (PA) types. (1) Including the processing circuit. (2) Gas chamber size The bold values denote the device performance from this work, highlighted to facilitate comparison with the other devices in the table Discussion Unlike previously reported board-level integrated NDIR CO2 sensor designs, the present work adopts a fundamentally different integration strategy that addresses several inherent limitations of discrete component assembly. Rather than optimizing individual components within a board-level framework, the OCI approach redefines the sensor architecture from the ground up through the monolithic co-integration of the sensing chip and gas chamber structure. From a system architecture perspective, monolithic on-chip integration fundamentally eliminates the assembly-induced geometric deviations inherent in board-level discrete assembly. As demonstrated by the optical simulation in Fig. 3, lateral displacement between the emitter and detector directly degrades the optical coupling efficiency. By lithographically defining the relative positions of all optical elements within a single fabrication process, the OCI architecture removes this source of variability and renders the optical alignment insensitive to assembly tolerances. From a dynamic performance perspective, the OCI sensor operates at a modulation frequency of 5 Hz — five times higher than the 1 Hz employed in the previous board-level design — while achieving a response time of approximately 3 s, representing a 70% improvement over the previously reported 10 s. In addition, despite the stringent thermal isolation requirements imposed by the on-chip co-integration of the infrared emitter and thermopile detector, the sensor maintains a compact overall module size of 21 mm × 10 mm × 5 mm, demonstrating that effective thermal crosstalk suppression and system miniaturization can be simultaneously achieved within a monolithic design. It is acknowledged that the transition from a dual-channel board-level design to a single-channel OCI architecture involves an inherent trade-off in long-term stability, as the absence of a reference channel limits real-time compensation for emitter aging and environmental drift. Nevertheless, the lithographically defined optical geometry and monolithic fabrication consistency of the OCI design provide inherent stability advantages that mitigate this limitation to a considerable extent. To assess the practical stability of the sensor, a continuous operation test was conducted over a period of seven days, during which the sensor output remained stable without significant drift, demonstrating acceptable long-term performance under the designed operating conditions. Incorporating a dedicated reference channel to further enhance interference rejection and long-term drift compensation remains a key direction for future work. Conclusions This work presents an OCI NDIR CO2 sensor that addresses the limitations of conventional NDIR CO2 sensors in terms of size, response speed, and system stability. By integrating a MEMS infrared emitter and a thermopile detector on a single chip and constructing a micro-scale gas chamber with an effective optical path of 11 mm and a coupling efficiency of 91.8%, the sensor achieves a synergistic optimization of optical efficiency, device uniformity, and gas exchange performance. Combined with optimized fabrication and packaging processes, a compact sensor chip with dimensions of 12 mm × 3 mm × 0.4 mm was realized. After integration with the optical channel and signal processing circuitry, the core sensing unit measures 16.6 mm × 5.0 mm × 3.7 mm, and the overall module is confined within 21 mm × 10 mm × 5 mm, demonstrating significant miniaturization compared to conventional NDIR sensors. System-level testing shows that, under a 2.5 V drive voltage and 5 Hz modulation frequency, the sensor achieves a measurement accuracy better than ±(50 ppm + 3% rdg) within the 0–10% CO2 range, a rapid response time of approximately 3 s, and a long-term stability RSD of 0.061% over 7 continuous days of operation, while maintaining consistent measurement performance and output stability across wide temperature and humidity ranges. These results confirm the effectiveness of the OCI and vertical cross-flow optical–gas channel design in enhancing optical coupling efficiency, reducing gas exchange time, and improving system robustness. The proposed OCI NDIR CO2 sensor combines high reliability with ultra-compact size, fast response, and high-performance detection, demonstrating promising potential for portable gas monitoring, smart home systems, industrial process control, and wearable applications, while providing a practical design reference for the next generation of miniature high-performance gas sensors. Supplementary information Supplementary Material On-chip integrated NDIR CO2 gas sensor Supplementary Material On-chip integrated NDIR CO2 gas sensor