An integrated monolithic photonic accelerometer with on-chip fringe-counting Authors: Yuqi Meng, Qiyun Hu, Ehsan Vatankhah, Zihuan Liu, Wei Wang, Neal A Hall Published: 2026 ABSTRACT AbstractPhotonic integrated circuits have transformed optical communication and information processing, yet their potential for precision inertial sensing remains largely untapped. To date, photonic accelerometers have been constrained by readout complexity, limited dynamic range, and insufficient experimental validation. Here we present the design, modelling, and experimental characterization of an integrated photonic accelerometer that combines high sensitivity with substantially enhanced dynamic range via phase-unwrapping. The device achieves a measured noise floor of 2.0µg/Hz, with a pathway towards ~10ng/Hz through simultaneous scaling of the input optical power and the number of waveguide loops on the cantilever — together projected to bring the input-referred shot-noise-limited sensitivity to the thermal-mechanical noise floor. By incorporating ellipse-fitting-based phase-unwrapping, the sensor enables accurate phase recovery over a substantially expanded measurement range, yielding a 39 dB improvement in dynamic range compared to a conventional Mach–Zehnder interferometer. To our knowledge, this is the first photonic integrated circuit accelerometer to incorporate fringe-counting functionality in a monolithic chip, and the first systematic study of such a strategy in integrated photonic inertial sensors. The photonic hardware enabling open-loop fringe counting — two Mach–Zehnder interferometers with a nominal quarter-wavelength inter-arm path-length offset, whose actual inter-interferometer phase difference is recovered from an in-situ ellipse fit — is monolithically integrated on the silicon-nitride chip, while the ellipse-fit phase-unwrapping algorithm runs in post-processing without active biasing or feedback control. These results establish integrated photonics as a promising platform for high-dynamic-range inertial sensing and a route to chip-scale accelerometers that could surpass existing chip-based technologies. Subject terms: Electrical and electronic engineering, Optical sensors, Sensors Introduction High-performance inertial sensors are increasingly important for a wide range of applications, spanning weak seismic signal detection1,2, microearthquake monitoring, tsunami and earthquake warning systems3,4, nuclear explosion monitoring5,6 for nonproliferation enforcement, as well as motion tracking in consumer electronics such as virtual reality and mixed-reality headsets7, and drone navigation8,9. Next-generation inertial measurement units require improved sensitivity, stability, and dynamic range to enable reliable detection of both weak and large-amplitude signals. Commercial MEMS accelerometers are predominantly based on capacitive or piezoelectric transduction. These technologies have been extensively studied, and their design trade-offs are now well established9–14. By contrast, optical displacement sensing offers a promising route toward higher-performance inertial sensors because of its high sensitivity and immunity to electromagnetic interference. A variety of optical interrogation schemes have therefore been investigated using different sensor configurations. One important class relies on free-space optics, where discrete optical elements such as a vertical-cavity surface-emitting laser and grating-based beam splitter are packaged in a compact form factor for accelerometer readout15–19, often together with force-feedback operation17,19. This approach has demonstrated high sensitivity and has also shown commercial viability20. Fiber-optic inertial sensors have likewise been widely explored, typically using interferometric or resonant configurations, including Mach–Zehnder interferometers21–23, Michelson interferometers24–26, fiber Bragg gratings27–30 and Fabry–Pérot cavities31,32. More recently, micro-opto-electromechanical systems (MOEMS)-based inertial sensors have emerged as an active area of research, driven in part by the rapid development and broader accessibility of integrated photonics fabrication platforms. On-chip photonic inertial sensors for acceleration detection have been demonstrated using both interferometric and cavity optomechanical approaches such as MRR33–36 or photonic crystal (PhC) cavities37–40. Among these, Mach–Zehnder interferometer (MZI)-based implementations41–44 offer a practical advantage in that they do not rely on high-Q optical cavities, thereby simplifying optical interrogation and potentially reducing readout cost and complexity. Compared with conventional optical sensors that often require discrete optical components and labor-intensive assembly, on-chip photonic sensors offer important advantages including CMOS-compatible fabrication, wafer-scale manufacturability, and compact footprint45. In this work, we leverage four concrete advantages of the silicon-nitride photonic integrated circuit (PIC) platform that goes beyond a generic interferometric readout. First, a compact chip-scale footprint eliminates the discrete optical alignment required by free-space and fiber-based homodyne systems. Second, CMOS-compatible silicon-nitride processing offers wafer-scale manufacturability and a path to substantial per-die cost reduction. Third, monolithic co-fabrication of the sensing waveguides and the cantilever on a single die gives intrinsic mechanical and thermal stability between the two MZIs, which is essential for maintaining the nominal quadrature relationship between the two MZIs that underpins our fringe-counting readout, with any residual fabrication-induced phase deviation absorbed by the ellipse-fit calibration. Fourth, the platform offers a clear path toward co-integration with on-chip laser dies and photodetectors for a fully self-contained sensor module. Prior PIC-based accelerometer demonstrations on photonic crystal37–40, microring33,46, and silicon-nanowaveguide45 platforms have established the feasibility of photonic transduction and contributed key performance benchmarks for PIC-based inertial sensing — including high sensitivity via cavity-enhanced transduction and demonstrations of frequency-modulated resonant readout — but to our knowledge none have demonstrated open-loop, bias-free, fringe-counting operation on a PIC. Despite this progress, an important challenge in MZI-based photonic inertial sensors remains insufficiently addressed: open-loop operation with the ability to simultaneously achieve dynamic range and low minimum detectable acceleration. In conventional MZI sensors, operation at or near quadrature (i.e., the point of maximum slope of the interference curve) is typically required to maintain linear response. As a result, bias drift can directly affect the scale factor, while large input acceleration may drive the sensor into a nonlinear regime or even across multiple fringes, leading to phase ambiguity. To address this limitation, we employ two interferometers with a fixed phase offset and use an ellipse-fitting-based phase-unwrapping method to recover the phase and enable phase measurements beyond the conventional linear operating range. Although phase unwrapping is a well-established technique in fiber-optic25,26 and free-space47,48 interferometric metrology, its translation to integrated photonic inertial sensors has been limited by three practical factors. First, achieving the required quadrature relationship between two interferometers on a PIC demands sub-wavelength control of relative path lengths, which has only recently become routine on low-loss silicon-nitride platforms49. We introduce a nominal λ/4 geometric offset and absorb any residual fabrication deviation via the ellipse fit. Second, many prior PIC accelerometer efforts have focused on employing high-Q optical-cavity transduction (photonic-crystal37–40 and microring33 resonators), where the linewidth itself sets the maximum measurable phase excursion and active biasing is required to hold the device at a Lorentzian slope — making open-loop operation challenging. Methods based on high-Q optical cavity resonances do not necessarily lend themselves to absolute phase measurements, and therefore phase unwrapping. In this work, we design and have fabricated a silicon nitride integrated photonic chip, and we then subsequently apply standard MEMS post-processing steps, including deep reactive ion etching and reactive ion etching, to define a cantilever-based optomechanical structure directly on the photonic platform. We developed a sensitivity model that accounts for both effective index variation and waveguide elongation, leading to a comprehensive description of the optomechanical coupling. The device is experimentally characterized using scanning laser Doppler vibrometry to determine its mechanical response under known large-signal excitation. The outputs from the two MZI channels are then processed using ellipse fitting and quadrature open-loop phase unwrapping. In addition, the sensor noise is modeled and compared with experimental measurements, indicating that the high-frequency noise floor approaches the shot-noise-limited sensitivity (input-referred shot noise), while the elevated noise at low frequencies is primarily attributed to environmental disturbances. To the best of our knowledge, this work represents the first demonstration of a homodyne MZI inertial sensor implemented on an integrated photonic platform with open-loop quadrature phase unwrapping, providing a practical route toward high-dynamic-range photonic inertial sensing. Materials and methods Design and simulation In brief, the accelerometer operates as follows. Input acceleration applied to the base of the chip causes the silicon-nitride cantilever to deflect; this deflection strains the silicon-nitride waveguides routed along the top surface of the cantilever, producing a change in optical path length and effective refractive index of each sensing arm. The resulting optical phase shift between each sensing arm and a common reference arm is read out as the intensity of the corresponding MZI output, and two such MZIs designed with a nominal λ/4 path-length offset (with the actual inter-MZI phase difference recovered from an in-situ ellipse fit) provide a quadrature pair from which the unwrapped optical phase — and hence the input acceleration — is recovered via ellipse-fit-based phase unwrapping in post-processing. The sensor is based on two MZIs integrated onto the same cantilever beam. The 3D CAD in Fig. 1a presents the sensor beam and the location of fiber array attachment. Light enters the chip via a fiber array connected to a grating coupler. The light is split into three parts: one reference arm and two independent sensing arms. The mask CAD in Fig. 1b, c highlights the location of sensing and reference arms. Each sensing arm is multiple turn and runs approximately 1/3 the length of the cantilever six times. The reference arm resides over the bulk silicon chip. Each sensing arm runs along the top surface of the complaint cantilever beam. Figure 1d is an abstraction showing the split of power into each arm. The reference arm output is split into two, and each sensing arm output is combined with a reference arm signal to complete the two independent MZIs. The light from each MZI exits the chip via a grating coupler. When input acceleration is applied to the base of the photonic chip, the resulting mechanical deformation of the cantilever induces an optical phase shift between each interferometer’s sensing and reference arm. The phase shift results in a change in optical output power, which is detected by a fiber-coupled photodiode off-chip. Figure 1e is a photograph of the sensor chip showing the etched mechanical cantilever beam. Figure 1f is a micrograph of the waveguide section of the dual MZI. The two MZIs are almost identical, except that MZI #2 Sensing Arm is made λ/4 longer than MZI #1 Sensing Arm to induce a static phase offset of π/2 in MZI #2’s interference curve relative to MZI #1. The effect is to achieve a quadrature relationship between the two independently operating MZIs on the same cantilever beam. Concept and design of the photonic MEMS Z-axis accelerometer with dual Mach–Zehnder interferometer (MZI) readout. a 3D CAD view showing the etched cantilever beam and tip mass, (b) Top GDS CAD view of the sensor chip, (c) zoom of the GDS CAD view showing the Dual MZI, (d) Abstraction illustrating optical paths and power splitting, (e) a Photograph of the fabricated Z axis beam device, (f) Micrograph of MZI waveguide structure (scale bar: 100 µm; grating coupler pitch ~127 µm) The output photodiode voltage from a single MZI can be written as: where neff is the effective index, λ is the effective wavelength of the waveguide core mode, and L is the total length of the sensing waveguide loops on the cantilever beam. Vo is the peak photodiode voltage achieved. The bending cantilever is shown in Fig. 2a, and the core of the waveguide is shown in Fig. 2b. As the waveguide deforms, both its physical length and effective refractive index change, denoted ΔL and Δneff. The effective index variation arises from two contributing mechanisms: the stress-optic effect and the modification of the waveguide core dimensions induced by deformation. The change of the effective index per input acceleration can therefore be written as: where a is the input acceleration, σ is the average stress inside of the waveguide along the y−axis of the cantilever in Fig. 1c, n is the refractive index of the waveguide, and wwg and hwg are the width and height of the waveguide as labeled in Fig. 2b. Numerical simulation of the optomechanical sensing mechanism. a The COMSOL CAD model of the sensor (color axis: strain tensor, yy-component, ×10⁻⁷), (b) The simulated Electric field distribution in the waveguide core, (c) The different contributing factors to the device sensitivity (x-axis: contribution to total phase response, %) The change of the waveguide core dimension per input acceleration is simulated in COMSOL, and the change in the effective index with respect to the change of waveguide core dimensions is simulated using Lumerical Optical FDE model with foundry-specific SiN refractive index data. The stress optic coefficient46,50 for SiO2 is −4.22×10−121/Pa and for the SiN is 3.6×10−121/Pa. Based on these simulations, the optomechanical coupling parameters are summarized in Table 1 and Table 2 below: Table 1Mechanical coefficientsCoefficientsValueUnit∂wwg∂a−3.54×10−14m/ms2∂hwg∂a−1.2×10−14m/ms2∂σSiN∂a1.87×104Pa/ms2∂σSiO2∂a5.23×103Pa/ms2dLda9.88×10−10m/ms2 Mechanical coefficients Table 2Optomechanical coupling coefficientsCoefficientsValueUnit∂neff∂wwg9.41×1041/m∂neff∂hwg7.23×1051/m∂neff∂σSiN3.6×10−121/Pa∂neff∂σSiO2−4.22×10−121/Pa∂neff∂ncore0.64∂neff∂nclad0.46 Optomechanical coupling coefficients Based on these parameters, the contributions of the aforementioned mechanisms are evaluated and their relative impacts on the total optical phase shift are summarized in the chart shown in Fig. 2c. It is observed that the effects associated with waveguide height and width variations contribute opposite direction compared to the stress-optic and elongation effects. Overall, the elongation-induced phase change is found to be more than five times larger than that resulting from effective refractive index variations. Accordingly, the optomechanical model is simplified by incorporating the effective refractive index change into a single scale factor to the length change, yielding: Fabrication of the sensor The photonic chip was fabricated using a standard silicon nitride (SiN) waveguide process. Here, SiN denotes the LPCVD stoichiometric silicon-nitride waveguide layer, and SiO₂ denotes the silicon-dioxide cladding/buffer layers. Following fabrication, post-processing of the photonic chip was carried out in university cleanroom facilities. In Fig. 3, processing steps a–e were performed at The University of Texas at Austin cleanroom, while steps f–g were completed at the Texas A&M University cleanroom. Fabrication process flow of the photonic MEMS accelerometer. a Backside photoresist mask patterning. b Define backside hard mask using RIE. c Topside PR mask patterning. d Topside hard mask etching. e Topside PR mask patterning for DRIE. f Topside DRIE. g Bottom side DRIE for structure release and defining cantilever. h Top view of the cantilever, the sensing waveguide loops extend 2 mm The chip includes a 3.5 μm silicon dioxide (SiO2) layer on the backside, which serves as the backside hard mask. The front-side total cladding thickness is approximately 5 μm, and the silicon handle layer has a thickness of 725 μm. As the first processing step, a 10 μm-thick photoresist (PR) layer was patterned on the backside of the chip. Backside reactive ion etching (RIE) was then performed to remove the 3.5 μm SiO2 layer, thereby defining the hard mask for a downstream deep reactive ion etching (DRIE) step. Next, a PR mask was patterned on the front side of the chip, followed by RIE to etch approximately 5 μm of the front-side SiO2 cladding layer. An additional PR layer was spun and patterned on the top surface at this stage to protect the front-side structures from continued etching. The chip then underwent a timed DRIE process from the front side to define the thickness of the cantilever beam. Subsequently, the chip was flipped to expose the backside, and a final DRIE step was performed to release the cantilever structure. After completion of the etching processes, the PR was removed using acetone followed by plasma ashing. The resulting cantilever beam thickness was measured to be approximately 167 μm using Dektak XT profilometer. Homodyne detection approach In our previous work, a thermo-optic modulator (TOM) was required to bias a single MZI to quadrature in order to achieve maximum sensitivity and linearity, thereby limiting the dynamic range and imposing strict thermal stability requirements for maintaining operation at the quadrature point. In this work, two independent MZIs are designed such that their outputs have an intrinsic phase difference of 90∘ with respect to each other. Input acceleration induces a common phase modulation in both interferometers. Therefore, two output signals from the photodiodes Vpd1, Vpd2 can be written as: where ϕt is modulated by input acceleration. Because of fabrication imperfections the two channels have different DC offsets and extinction ratio, and the intrinsic phase difference is not perfectly 90∘. (Vpd1, Vpd2) can more generally be written as: When plotted against each other in x-y mode, their trajectory forms an ellipse rather than an ideal circle. To extract the amplitude factor (a1,a2), DC bias (b1,b2) and relative phase difference (ϕ12=ϕ1−ϕ2), an ellipse fitting method based on the least squares method is used. Once the fit is performed and parameters are extracted, the ellipse can be mapped to a unit circle, thus achieving an orthogonal representation of the two signals (I=cosϕt,Q=sinϕt). Through phase unwrapping the dynamic phase change caused by input acceleration ϕt can be solved. Compared with single-channel readout, which relies on the direct mapping of intensity variations from a single interferometer, the quadrature method transforms the measurement from the voltage domain into the phase domain, enabling unambiguous phase recovery across multiple optical fringes. This substantially expands the sensor’s usable dynamic range while simultaneously eliminating the need for phase tuning using TOM. Experimental method As shown in Fig. 4a, a C-band DFB laser is used as the optical source. The output light first passes through a Polarization Controller (PC) and is then coupled via an epoxy-bonded fiber array to the grating couplers (GCs) on the chip. Light entering the chip is equally split into two independent MZI. The output light from the two MZIs are individually detected by Thorlabs Photodetectors (DET01CFC) and then amplified using transimpedance amplifiers with a gain of 105VA. The chip is mounted on a shaker (Wilcoxon F4), and excited by a controlled signal generated by the DAQ of a SLDV (scanning laser doppler vibrometer) model PSV-500 from Polytec. The PSV-500 was used to noninvasively probe the surface vibration profile of the entire chip in response to periodic test signals, while the MZI outputs are simultaneously recorded. Figure 4b shows the PSV-500 LDV laser head as well as the shaker test platform. Figure 4c, d are photographs of the chip at different zoom levels. Figure 4d make visible mechanical structures other than the Z-axis, but those are not relevant to the scope of this article. Experimental setup for vibration testing and characterization of the packaged sensor. a Experimental setup for sensor characterization: A: Fiber coupler 1550 nm laser source, B: Polarization controller, C: Shaker Platform, D: Accelerometer chip, E: Scanning laser doppler vibrometer from Polytec, F: photodiode, G: Transimpedance amplifier, H: PSV-500 data acquisition system, (b) photograph of the packaged sensor on the shaker platform under the SLDV for vibration testing, (c) close-in photograph of the packaged sensor, (d) micrograph of the sensor captured using the microscope and camera of the LDV Several experimental parameters are noted here, including the optical-loss budget and the inter-MZI phase difference, both of which determine the achievable noise floor. The grating coupler was designed for a per-facet insertion loss of ~10 dB at 1550 nm, yielding an anticipated loss equal to 20 dB for each MZI. The on-chip silicon-nitride waveguide propagation loss is < 0.5 dB/cm based on foundry-provided characterization data. Each MZI employs approximately 2 cm of total waveguide, therefore approximately 1 dB of propagation loss. All total, one would therefore anticipate 21 dB total loss for each MZI. The measured end-to-end optical insertion loss of the packaged sensor, however, is higher than this design target: the end-to-end insertion loss of MZI #1 is 33.6 dB, and that of MZI #2 is 31.9 dB. We attribute the discrepancy primarily to the fiber-array packaging, in which the bonding step did not bring the fiber facets into ideal direct contact with the grating-coupler surface, leaving a small unintended air gap. It is also possible that the grating coupler itself is underperforming relative to its design (a known sensitivity of grating couplers to small fabrication offsets in etch depth and grating pitch). We start with 120 mW laser power and end with approximately 33 µW laser power at the photodetector for each MZI. The modest ~1.7 dB difference between the two MZIs is consistent with small channel-to-channel variation in the 1 × 2 splitter ratio and in the output grating couplers. In the future, we will migrate from grating couplers to edge couplers (2 dB to 3 dB insertion loss vs. 20 dB end-to-end), to reduce the per-facet insertion loss and increase the on-chip optical power available. The intrinsic phase difference between the two MZIs was designed to be π/2; the ellipse-fit value extracted from the measured (Vpd1, Vpd2) trajectory is ϕ₁₂ = 125°. The deviation from the nominal 90° is attributed to fabrication variation in the waveguide path-length difference and is fully absorbed by the ellipse-fitting calibration step. Results and discussion Large signal response A large signal test is performed to obtain the five elliptic parameters. To achieve a robust elliptical fit, the induced phase signal ϕt of the MZIs should traverse a large range, and this is achieved by applying a sinusoidal tone to the shaker near the cantilever’s fundamental resonance frequency at 1315 Hz. Figure 5a presents an LDV-scanned deflection profile of the beam vibrating at its fundamental mechanical mode, and Fig. 5b presents the photodiode voltage waveforms from the two MZIs. For each waveform presented, the static offset voltage is subtracted, and the signal is scaled such that its range is −1, + 1 yielding: Resonant response and quadrature optical phase readout of the sensor. a SLDV scan image of the cantilever beam at its resonance 1315 Hz. b Photodiode 1 and photodiode 2 output under large amplitude drive and c the corresponding elliptic circle, (d) the measured tip mass displacement and unwrapped optical phase plotted on the same graph, e FFT of photodiode 1 and 2 voltage output, (f) FFT magnitude of the unwrapped optical phase ϕt; The annotated values 1.61906 rad and 0.00444 rad correspond, respectively, to the magnitude of the fundamental tone at 1315 Hz (the signal) and to the largest harmonic at 2630 Hz (the dominant distortion harmonic); their ratio yields the < 1% THD value The subscript n denotes normalization. Figure 5c presents the two signals in xy mode (i.e., Vpdn2 vs. Vpdn1), providing the anticipated ellipse. We note that the photodiode waveforms in Fig. 5b remain proper trigonometric functions of the instantaneous optical phase, Vpdn1=cos[ϕ(t)] and Vpdn2=cos[ϕ(t)+ϕ12], as noted by Eqs. (4)–(7); their visually unusual shape under large-amplitude resonant excitation is a direct consequence of the fact that φ(t) is itself a strongly modulated, non-monotonic function of time that traverses multiple full fringes within each shaker cycle. That the (Vpdn1,Vpdn2) trajectory falls on a clean ellipse in Fig. 5c and that the unwrapped phase in Fig. 5d is a clean sinusoid that tracks the LDV-measured tip displacement confirms that the underlying mapping is the expected cosine and sine pair. With the ellipse parameters obtained, the sensor is now calibrated and the unwrapped phase signal ϕt can be obtained for all subsequent measurements. Next, we apply to the shaker a tone burst waveform 4 s in duration and with 50% duty cycle, comprising approximately 2600 excitation cycles which is sufficient for the cantilever vibration to achieve steady-state vibration. We capture unwrapped phase from the sensor while simultaneously capturing the tip displacement of the cantilever using the LDV. Figure 5d presents both signals. The unwrapped phase is expected to be proportional to the tip displacement, and this is indeed the case. Both signals capture the transient ramp-up of the cantilever displacement in response to the tone burst. The phase unwrapping using the ellipse fitting method faithfully captures transient cantilever vibrations that induce a phase change of 7.2 rads peak-to-peak, more than 2π. The FFT of each individual photodiode signal is presented in Fig. 5e. Pronounced harmonic distortion is evident in each individual interferometer signal (i.e., Vpd1 and Vpd2). In contrast, the FFT of the unwrapped phase in Fig. 5f shows negligible distortion. Using the data markers in Fig. 5f, the harmonic distortion ratio (THDF, defined as the incoherent sum of higher-order harmonic amplitudes referred to the fundamental amplitude) is <1% for the phase excursion equal to 7.2 rad peak-to-peak. For a conventional MZI sensor such as that reported in [43,45], 1% THD corresponds to a signal amplitude equal to 0.08 rad peak-to-peak. The corresponding improvement in dynamic range in this work compared to other works is therefore 20×log10(7.2rads/0.08rads)=39dB. The current demonstration is limited by the maximum vibration amplitude available from the shaker. Phase unwrapping at even larger amplitudes should in principle be feasible and lead to more impressive improvements. These results therefore highlight the substantial dynamic range enhancement and harmonic distortion suppression enabled by the proposed technique. Sensitivity and noise characterization Next, we apply a broadband signal to the shaker while simultaneously measuring base acceleration using the LDV and the unwrapped phase from the sensor. We normalize the unwrapped phase to base acceleration at each frequency. A broadband chirp is applied to the shaker for this purpose, (1 Hz–5 kHz). Sensitivity is defined as the phase change per unit acceleration, expressed in rad/m/s2. Figure 6 presents the measured transfer function alongside the simulation prediction, showing excellent agreement. The resonance quality factor Q is the only fitted parameter and is determined by calculating the ratio of the measured peak magnitude to the flat-band value; the measured Q is approximately 1000. The values of all other input parameters (e.g., mechanical resonance frequency and the optical sensitivity in the flat band) are unfitted and directly taken from the design stage. The sensitivity of the sensor remains flat before resonance at approximately 1315 Hz and is 6.4×10−3rad/m/s2. The small imperfections in the measurement at 800 Hz and 900 Hz result from normalization of the captured phase to the input base excitation. The shaker has narrowband resonances and antiresonances in the measurement band, and normalization is prone to small errors at these frequencies. Frequency response and noise performance of the accelerometer. Panel (a) is the frequency response of the sensor using the homodyne unwrap approach, and (b) is the simulated and measured input-referred noise (MDA minimum detectable acceleration, computed as the square root of the measured input-referred acceleration power spectral density) The sensor’s noise performance was characterized under ambient conditions with all active excitation signals disabled. Figure 6b plots the input-referred acceleration noise (minimum detectable acceleration, MDA, defined as the square root of the measured input-referred acceleration power spectral density), which is obtained by converting unwrapped phase noise (rad/Hz) to acceleration using the calibrated phase sensitivity of 6.4×10−3rad/(m/s2). The remaining curves in Fig. 6b are simulated. The low-frequency rise in measured input-referred noise reflects ambient acceleration noise and possibly laser RIN. The rise is not due to the device transfer function, which is flat below resonance. The static Vpd1 and Vpd2 values at rest are used to compute the standard deviation or RMS value of each signal’s shot noise. The two uncorrelated time-domain noise waveforms are created, and the phase noise resulting from the two waveforms is computed using Eq. (8). Details of the shot noise simulation can be found in the supplemental materials. The phase noise is, in-turn, converted to acceleration using the calibrated sensitivity from Fig. 6a. Finally, we take the FFT of the computed time domain acceleration noise to present the spectrum shown in Fig. 6b. At low frequencies, the measured input-referred noise is higher than the simulated shot-noise floor and shows a pronounced low-frequency rise, indicating dominance of ambient acceleration noise or possibly laser RIN. As the frequency increases, this gap progressively narrows, and the measured noise approaches the simulated shot-noise floor, indicating that the noise floor of the sensor is fundamentally limited by shot noise. In our experiments, the optical power at the photodetector is 33 μW. The corresponding shot noise per Fig. 6 is 2.0μg/Hz, in the flat band region. Theoretically, if the output power were 1 mW, the input referred shot noise would be 363 ng/Hz. The thermal mechanical noise floor is independent of optical system properties and is also included for reference. The noise is approximately 10ng/Hz in the flat band, considering the 19 mg effective mass of the cantilever. Discussion In this work, the design, fabrication, and characterization of a novel optomechanical accelerometer are presented. The PIC combines a flexural mechanical structure with the high sensitivity enabled by optical detection. Two MZIs with different static phase offsets were integrated on the same mechanical cantilever beam, enabling a homodyne detection scheme with elliptic fitting and phase unwrapping. This approach eliminates the need to bias the sensor at a fixed operating point and overcomes the trade-off between noise floor and maximum measurable acceleration. A substantial improvement in dynamic range has been demonstrated with this sensor, compared to conventional MZIs. Among the features of the present device, the aspects that are uniquely enabled by photonic integration — as opposed to being generic features of any interferometric readout — are the following. Monolithic on-chip integration of the two MZI sensing arms and their reference arm on the same die provides sub-wavelength control of the inter-arm path-length offset, which sets the nominal quadrature condition between the two MZIs to within a fabrication tolerance that is then absorbed by the ellipse-fit calibration on every acquisition. Co-fabrication of two MZIs on a single silicon-nitride die also ensures that the two interferometers share a common thermal and mechanical environment, so that environmental fluctuations are largely common-mode and do not corrupt the quadrature relationship. We emphasize that the open-loop fringe-counting approach demonstrated here is complementary to — rather than supplanting — the high-Q resonant transduction approaches33,38–40 that have driven much of the prior PIC-accelerometer literature; the two architectural families occupy different points in the sensitivity, dynamic-range, and readout-complexity design space. Table 3 compares the present work with previously reported photonic accelerometers in terms of sensing performance and readout characteristics. While several prior platforms have demonstrated high sensitivity and low noise floors, their range of measurable acceleration amplitudes often remains constrained by the optical transduction or readout scheme. In the present work, the use of monolithically integrated on-chip fringe-counting readout substantially relaxes this constraint, enabling low-distortion phase recovery over a much larger acceleration measurement range. In this work, the experimentally demonstrated gains in linearity and THD, while significant at 39 dB, are limited by the available shaker amplitude rather than by the optical readout itself. Table 3 therefore underscores that the key contribution of this work lies in establishing an integrated photonic accelerometer architecture in which sensitivity and usable acceleration range can be improved simultaneously, rather than being traded against one another through readout limitations or by mechanical design parameters. Table 3A comparison of the principal performance metrics, including those achieved in this workWorkTechnologyMDA μg/HzFundamental Frequency kHzOpen-loop operation: measurement range decoupled from optical-transduction linearityLi et al.38PhC Cavity2072.1NoGe et.al.41DSMZI4.293.4NoKrause et.al.39PhC Cavity1027.5NoFlores et.al.40PhC Cavity17.0241.5NoBramhavar et.al.33MRR616.3NoPrevious Work42MZI0.51.18NoThis WorkHomodyne dual MZI2.01.3Yes, 39 dB imp. over single MZI A comparison of the principal performance metrics, including those achieved in this work When the sensor is driven at its mechanical resonance with a large-amplitude excitation, the resulting peak-to-peak optical phase oscillation exceeds one wavelength and spans more than one optical fringe. Across this large phase range, the sensor readout is shown to faithfully resolve the vibration waveform. The measured input-referred shot-noise is approximately 2.0μg/Hz, while the calculated input-referred thermal-mechanical noise floor is approximately 8.6ng/Hz given Q≈1000,meff=19mg, and f0=1,315Hz, and following atm=8πkBTfomeffQ. It should be noted that the ultimate noise floor of the sensor is fundamentally limited by thermal mechanical noise. The input-referred shot-noise can be reduced by increasing the optical power. The optical shot noise on the photocurrent itself scales as Ppd and the sensitivity scales with Ppd, so that input-referred noise floor scales with 1/Ppd. The optical sensitivity also scales with N, the number of waveguide loops, so that input-referred shot-noise scales with 1/(NPpd). The number of waveguide loops used in this study is six. Owing to the low propagation loss of silicon nitride waveguides, a nearly proportional reduction in input-referred shot noise can be achieved by increasing the number of loops with minimal optical power penalty. For example, a device incorporating 60 waveguide loops and operating with optical power of 10 mW at the photodiodes is expected to achieve a shot noise floor equal to 14.4ng/Hz, comparable to the thermal mechanical noise limit. This work establishes the feasibility of an optomechanical sensor platform that leverages homodyne detection and phase-unwrapping to achieve superior noise performance and an expanded dynamic range. To the authors’ knowledge, this represents the first demonstration of an optomechanical accelerometer implemented in a PIC and using phase-unwrapping based on a dual MZI architecture and elliptic fitting. Conclusion In this study, an optomechanical sensor implemented on a PIC platform has been demonstrated. Post-processing steps, including DRIE, were applied to the PIC to form a flexural mechanical structure for inertial sensing. A model has been developed that used COMSOL to compute beam surface strain combined with Lumerical to compute the change in effective refractive index due to geometrical and stress changes. The validity was experimentally verified through vibration characterization using a shaker and Laser Doppler Vibrometry (LDV). The measured noise spectrum was compared with a shot noise model and shows good agreement at high frequencies, where environmental noise is no longer dominant. In addition, the proposed dual MZI architecture eliminates the need for biasing the sensor at a fixed operating point for linear operation. The sensor operates in an open-loop configuration, and an elliptical fitting algorithm is employed to unwrap the optical phase, thereby extending the achievable dynamic range. The experimental results demonstrate a clear and practical roadmap toward an optical MEMS accelerometer with a thermal-mechanical-limited input-referred noise floor of order 10ng/Hz and a significantly enhanced dynamic range. Supplementary information Simulation Method Noise Modelling Method Acknowledgements The authors gratefully acknowledge partial support from the U.S. Department of Energy through the Geothermal Prize and from the National Aeronautics and Space Administration under Contract No. 80NSSC25C0346. The authors also acknowledge Silicon Audio Seismic for its support of this work and for its engagement with the practical translation of the research. The authors further acknowledge the facilities, infrastructure, and technical environment of The University of Texas at Austin that enabled the design, fabrication, packaging, and characterization of the devices reported in this study. The authors also thank the supporting photonics foundry and fabrication partners for enabling device realization through access to silicon nitride photonic integration and associated fabrication processes. Author contributions Y.M. and Q.H. contributed equally to this work. Y.M. and N.A.H. conceived the idea. Y.M. designed the chip, fabricated, packaged, and characterized the device, and performed the data processing and analysis. Q.H. performed sensor modelling, data processing, figure preparation, manuscript writing, sensor packaging, and experimental measurements. E.V., Z.L. and W.W. contributed to manuscript preparation. All authors participated in scientific discussions and contributed to the interpretation of the results. All authors reviewed and approved the manuscript. Conflict of interest Co-author Neal A. Hall has an ownership interest in a business that develops and sells seismic sensors for geophysical applications. Footnotes These authors contributed equally: Yuqi Meng, Qiyun Hu. Supplementary information The online version contains supplementary material available at 10.1038/s41378-026-01410-1. References Associated Data Supplementary Materials Simulation Method Noise Modelling Method