A curved-beam piezoelectric MEMS resonator featuring multiple temperature plateaus with enhanced stability Authors: Yancheng Lian, Yueyang Li, Fangzhou Chen, Zhuoyue Zheng, Jinuo Zhou, Junchen Liu, Fengzhe Xu, Jinbo Yang, Jianlin Chen, Chen Wang, Michael Kraft, Ka-Meng Lei, R P Martins, Pui-In Mak, Yuan Wang Published: 2026 ABSTRACT AbstractSilicon-based Micro-electromechanical Systems (MEMS) resonators face stability challenges due to their temperature susceptibility, which impedes the advancement towards high-precision applications. This work reports a multi-modal curved-beam piezoelectric resonator that enhances frequency stability through engineered irregular geometry and mechanical nonlinearity. Utilizing a standard Silicon-On-Insulator (SOI) - Aluminum Nitride (AlN) process, the proposed device engenders 17 vibration modes within 5 MHz, addressing the detection constraints and information extraction of non-inplane modes for capacitive counterparts and circumventing the needs of vacuum encapsulation. In addition, owing to the inimitable temperature coefficient of frequency (TCf) insensitive plateaus discovered in high-frequency modes (Mode 16 and Mode 18), a novel stability enhancement strategy has emerged. Through open-loop sweeps and phase-locked loop (PLL) tracking, it is suggested that these plateaus likely stem from modal coupling triggered by mechanical nonlinearities. This behavior, reminiscent of frequency veering, potentially enables energy redistribution between modes, leading to the resonance frequency deviating from the intrinsic thermal drift trajectory of silicon. Experimental results show that Mode 16 achieves a stability of 17.4 ppb at 46.390 s at a 62 °C plateau under ±1 °C fluctuations and reaches 2.0 ppb under stringent temperature regulation. Mode 18 reaches a peak stability of 37.9 ppb at 43 °C. This innovative approach presents a promising avenue for developing high-precision frequency references without the need for intricate TCf compensation, providing considerable potential for the evolution of next-generation silicon micro-resonators and enhancing their capabilities as stable timing sources in various applications. Subject terms: Electrical and electronic engineering, Engineering Introduction High-stability frequency sources serve as the fundamental backbone of the information society, playing an indispensable role in various applications, including clock generation, synchronous communication, and precision sensing. The rapid evolution of integrated circuits continuously pushes the boundaries of these timing units toward extreme miniaturization and enhanced stability. Traditionally, quartz crystal oscillators (QCOs) have long dominated the market by leveraging high quality factors (Q-factor) and excellent stability1–3. However, the intrinsic physical and material constraints of quartz crystals result in a bulky form factor and a fundamental lack of compatibility with standard CMOS manufacturing processes3,4. These limitations significantly hinder the potential of QCOs for a high-density monolithic integration system within advanced packaging architectures. To circumvent these critical bottlenecks, MEMS resonators have emerged as a transformative technology that aligns with the scaling laws of the semiconductor industry. Compared to quartz-based components, these micro-scale devices utilize standard silicon batch micromachining to achieve a substantially reduced footprint and lower power consumption. The primary advantage of MEMS technology lies in its ability to be fabricated directly alongside or even on top of driving and sensing circuitry5. This proximity effectively eliminates the parasitic effects associated with traditional wire bonding while simultaneously reducing system complexity and cost6,7. Beyond their integration merits, silicon MEMS resonators exhibit superior mechanical robustness against environmental vibrations and high-g acceleration shocks, which enhances long-term reliability for mobile applications8,9. Furthermore, the inherent scalability of the silicon platform allows for continuous performance optimization through the synergy of structural design and material engineering. Such multifaceted advantages enable MEMS resonators to provide high-precision timing across diverse sectors, including 5 G communications10,11, satellite navigation12,13, IoT sensor nodes14,15, and harsh-environment automotive electronics16,17. The operational utility of silicon MEMS resonators, despite their integration and robustness benefits, is significantly impeded by their susceptibility to ambient environmental changes (e.g., temperature fluctuations). In real-world applications, these resonators face the challenge of sustaining frequency stability amidst a range of external factors, including packaging-induced stress, fluctuations in ambient pressure, and changes in humidity levels18. Among these variables, temperature remains the most significant factor compromising frequency stability, posing a formidable challenge for high-precision applications19,20. In silicon-based micro-resonators, the resonance frequency is fundamentally governed by the elastic stiffness of the material and mechanical structure. As the ambient temperature rises, silicon undergoes the thermal softening effect, where the energy of atomic vibrations interferes with the lattice stability. This results in a reduction of the effective stiffness of the resonant structure, leading to a concomitant downward shift in the resonance frequency4,21. Underpinned by this physical mechanism, silicon MEMS resonators typically exhibit a substantial negative TCf22,23. This intrinsic thermal sensitivity represents the primary obstacle to replacing prevailing QCOs in the frequency control market. To mitigate the inherent thermal drift and enhance the frequency stability of silicon MEMS resonators, researchers hitherto have explored two primary technical pathways: passive TCf regulation and active feedback control20. Passive compensation typically operates on the principle of cancellation of TCf-induced stiffness deviations, achieved by incorporating functional materials with a positive temperature coefficient into the resonant structure or utilizing structural design24,25. Hajjam. A et al.26 achieved temperature compensation through high-concentration phosphorus doping, resulting in a low temperature drift of 0.4 ppm/°C. Lan. K. W. et al.27 employed a vertically stepped temperature compensation scheme to achieve linear compensation, reducing the first-order TCf from −374.5 ppm/°C without compensation to −1.45 ppm/°C. As affirmed by technical reports and literature, this configuration effectively neutralizes the intrinsic thermal softening of the silicon material, thereby boosting the stability of silicon micro-resonators. Although this approach enables zero-power operation, the mismatch in the coefficient of thermal expansion between heterogeneous materials often incurs parasitic internal stress and interfacial damping. Such effects lead to significant energy dissipation, which inevitably degrades the Q-factor and is detrimental to the performance of micro-resonators20,28. Conversely, active compensation relies on environmental control as its central mechanism. Researchers employing this approach often incorporate micro-ovens to uphold a consistent thermal balance or utilize electronic tuning loops to readjust the frequency to its designated value29,30. Zadeh. S. A. G. et al.31 achieved simulated temperature stability of ± 0.74 ppm over the −45 °C to 85 °C range using a dual-resonator compensation technique with thermal feedback. Liu. C. S. et al32 fabricated a 77.7 MHz oven-controlled MEMS oscillator using the structural resistance of a resonant cavity as an embedded temperature sensor for temperature compensation, achieving a frequency stability of ± 0.3 ppm over a temperature range from −25 °C to 85 °C. Albeit offering superior precision and adaptability to environmental fluctuations, active methods are curtailed by elevated power consumption and increased system complexity. Consequently, achieving high-stability frequency references that reconcile high Q-factor with minimal power budgets continues to be a significant bottleneck in the development of silicon MEMS resonators. Building upon the afore-narrated background and encompassing the gist of TCf amelioration, this work proposes a piezoelectric MEMS resonator featuring a unique curved-beam structure. By leveraging standard micro-machining processes without auxiliary processing, this device offers the advantages of manufacturing simplicity and cost-effectiveness. In our previous research, a curved-beam capacitive resonator was developed with the objective of achieving multimodal output, successfully demonstrating six resonance modes that clustered within a modest frequency regime ( < 300 kHz)33. Nevertheless, the detection was confined to in-plane flexure modes due to the micro-machined capacitive architecture, resulting in the observed modal count falling short of theoretical expectations and, therefore, degrading the efficacy of the multimodal device and missing vital information of other modes. To address this limitation, we utilize a piezoelectric transduction mechanism to simultaneously detect in-plane, out-of-plane, bulk, and flexural vibration modes in ambient room air conditions. As a result, the proposed resonator showcases a diverse multimodal behavior, with the identification of 17 distinct modes within the 5 MHz range. As a preliminary investigation into this innovative subject and a proof-of-concept demonstration of functionality, the beat-frequency scheme was utilized across different vibration modes to enhance the device’s stability in a dynamic temperature environment, achieving a stability of 127.0 ppb across a dynamic temperature gradient from 25 °C to 75 °C34. Benefitting from the piezoelectric transduction to extract the information on out-of-plane/bulk vibration modes in the MHz range, another distinct attribute was unveiled, i.e., a temperature-insensitive plateau. Following the discovery of such an intriguing phenomenon, this study emphasizes an in-depth exploration of the exceptional frequency stability attained by these plateaus across multiple modes, despite ambient thermal fluctuations. These findings highlight the potential of advanced multi-mode MEMS resonators for multifunctional sensing and high-precision frequency references, particularly for mid-frequency devices in IoT scenarios. The primary contributions of this research are presented in the following aspects where (1) an AlN-on-SOI piezoelectric MEMS resonator featuring 3rd-order Bézier geometries is demonstrated to achieve the multi-modal detection of 17 resonance modes within a 5 MHz bandwidth; (2) distinct temperature-insensitive plateaus are experimentally identified in high-frequency regimes where Mode 16 exhibits a frequency stability of up to 20 ppb under dynamic temperature fluctuations; and (3) comprehensive open-loop experimental evidence regarding secondary peak emergence and frequency veering is provided to elucidate the underlying mechanism of nonlinearity-mediated frequency stabilization within these plateau windows. Results Design of the device structure To address the constraints posed by the limited number of vibration modes in traditional MEMS resonators, this research introduces a piezoelectric MEMS resonator featuring a curved clamped-clamped beam structure. Unlike conventional resonators that utilize linear geometries with substantial structural symmetry, this innovative design deviates from the norm. While symmetry in conventional designs aids in modal orthogonality, it also hinders effective modal coupling between diverse modes, consequently confining the detectable vibration modes within a set frequency range. As such, given that the gist of the design criteria is to elevate modal density and foster modal clusters, driving the deliberate incorporation of a predefined curvature. The continuously varying curvilinear profile induces geometric nonlinearity, resulting in a redistribution of internal axial stress and promoting coupling between axial and bending deformations. Theoretically, this breaking of symmetry weakens modal orthogonality, establishing the structural conditions necessary for the coexistence of multiple modes. Finite element method (FEM) simulations using COMSOL Multiphysics demonstrate that the 3rd order Bézier profile supports a superior density of 17 resonance modes, whereas simpler straight or S-shaped designs yield only 12 to 15 modes within the frequency range of 5 MHz. This enhanced structural complexity enables intricate modal interactions and coupling strengths that remain fundamentally unattainable with standard linear, circular, or curved variations. In terms of practical implementation, the profile of the curved beam is designed based on the 3rd-order Bézier curve. The curvature parameters were iteratively optimized via FEM simulations using COMSOL Multiphysics. The resulting optimized initial curvature distribution function is defined as follows: To provide a concise dynamical analysis of the device, consider a clamped-clamped beam with a predefined initial profile ω0(x,t). Based on the Euler-Bernoulli beam theory and accounting for the geometric nonlinearity, the governing equation for transverse vibration is expressed as35: where E is the effective Young’s modulus of the beam, I is the moment of inertia, ρ is the average density, A is the cross-sectional area, ω and ω0 denote the transverse dynamic displacement and the predefined initial profile of the curved beam, respectively. Fpiezot denotes the time-dependent piezoelectric actuation force applied to the resonator. The architecture of the proposed device, the geometry dimensions, and the illustrative fabrication flow are depicted in Fig.1, while the details concerning the fabrication method and experimental platform are elaborated in Section 6. Materials and Methods. a Illustrative outline of the SOI piezoelectric (AlN) fabrication process. b Device dimensions and c Rear view of the device Basic Characterizations In this experiment, Lock-in Amplifier 1 is employed to assess the fundamental characteristics of the DUT. As illustrated in Fig. 2a, the baseline response encompasses the amplitude-frequency relationship along with the corresponding Q-factor. Additionally, we utilized COMSOL Multiphysics to perform FEM analysis for the visualization of modal shapes, as shown in Fig.2a. The acquired data indicate that the DUT exhibits 17 distinct resonant modes below 5 MHz, manifesting that the curved-beam piezoelectric resonator devised in this study offers a more diverse modal combination compared to traditional counterparts (e.g., straight clamped-clamped beam or double-ended tuning fork resonator). a Baseline characterizations of the proposed device operating in air show 17 distinct modes with FEM analysis within a 5 MHz range. b Q-factor measurement using the ringdown method for two selected working modes, i.e., Mode 5 and Mode 16 Pursuant to the garnered data, the quality factor demonstrates significant variation across different frequency regimes. The Q-factor of lower-frequency modes (Mode 1–7), whereby, fall in a modest range, roughly 150–2500. The mid-frequency modes (Mode 8–15) display a notable increase, with Q-factors reaching 2500–6000. As for the high-frequency modes (Mode 16 and Mode 17), the Q-factor continues to increase, approaching 7300. This trend aligns well with the reduced impact of thermoelastic and anchor-related losses at higher frequencies. Among these modes, Mode 5 (817.5 kHz, Q = 1800) and Mode 16 (4.2981 MHz, Q = 6700) exhibit comparatively stable performance and favorable signal-to-noise ratio (SNA), therefore are taken as representative examples and selected as operating modes. Furthermore, to ensure the accurate extraction of the working mode quality factor, the Q-factor of Mode 5 and Mode 16 is determined utilizing the ringdown method. The ringdown traces for Mode 5 and Mode 16 are depicted in Fig. 2b, both showcasing linear exponential decay. Mode 5 exhibits a relatively rapid decay, while Mode 16 maintains oscillation for a prolonged period, aligning precisely with their respective Q-factors. Temperature-Frequency Characterizations In this experiment, lock-in amplifier 1 is utilized to investigate the temperature-frequency properties of the DUT, focusing on Modes 5 and 16 for in-depth evaluation, particularly emphasizing the innovative temperature-frequency abnormal region discovered from this proposed subject. The resonant frequencies of both modes are monitored as the chamber temperature is incrementally raised from 0 °C to 80 °C, with the PLL PID module of the lock-in amplifier employed for frequency tracking. As such, the resultant frequency variations are shown in Fig. 3a. a Closed-loop measurement based on PLL resonant frequency tracking to characterize temperature-frequency relationship of the selected modes across 0 °C to 80 °C. b Open-loop frequency sweep measurement of Mode 5 across 0 °C to 80 °C Mode 5 exhibits a gradual decrease in resonance frequency as the temperature rises, in line with the expected Temperature Coefficient of frequency (TCf) behavior typical of silicon-based structures. This phenomenon is attributed to the decrease in the elastic modulus of silicon with increasing temperature, resulting in a decline in the structure’s effective stiffness and a subsequent negative temperature coefficient for the resonant frequency. As a result, Mode 5 serves as a reference for subsequent measurements. Mode 16 displays a similar overall trend, with an anomaly observed near 62 °C where its resonance frequency stabilizes instead of decreasing. A comparable temperature-insensitive “plateau” region is also identified in the higher-frequency Mode 17, suggesting that certain modes above 4 MHz may enter a temperature-insensitive regime. To delve deeper into this observation, lock-in amplifier 1 (operation bandwidth up to 5 MHz) is replaced with lock-in amplifier 2 (operation bandwidth up to 50 MHz) for further experiments to explore modes exceeding 5 MHz of the DUT. Following evaluation, Mode 18 (6.9128 MHz) is chosen as the representative mode, and its temperature-frequency response curve is presented in Fig. 3a. The temperature-frequency characteristics of Mode 18 generally mirror those of Mode 5, but with three distinct abnormal temperature regions. Subsequent sections will delve into a detailed analysis of this intriguing phenomenon. In addition, Fig. 3b presents the open-loop sweep results for Mode 5 while undergoing temperature elevation in the chamber from 0 °C to 80 °C, with incremental scans of 1 °C. These results from the open-loop sweep vividly illustrate the resonance frequency drift progression, establishing the control baseline for forthcoming experiments. Temperature-frequency abnormal region of Mode 16 This section presents a detailed investigation of the temperature–frequency response of Mode 16, with specific attention to the newly discovered abnormal region. The temperature characteristics of Mode 16 are illustrated in Fig. 4a. Throughout the temperature sweep, three notable abnormal regions emerge, with the segment around 62 °C displaying the most pronounced deviation. Consequently, this particular interval becomes the focal point for in-depth analysis in the subsequent investigations. a Temperature-frequency characteristics of Mode 16 with three abnormal temperature regions identified across 0 °C to 80 °C, wherein three observation points (P1–P3) are selected within the primary section for further investigations. b Open-loop frequency sweep measurement of Mode 16 across 0 °C to 80 °C. c Zoomed-view of the open-loop sweep for the abnormal region within the primary section The detailed scrutiny of the magnified view in Fig. 4a elucidates the progression of the resonant frequency within the primary abnormal region. Initially, the frequency declines with decreasing temperature, subsequently transitioning to a positive temperature coefficient, and eventually reverting to a negative coefficient as the temperature continues to rise. To delve into the rationale behind this atypical behavior, an open-loop frequency sweep of Mode 16 is conducted while heating the system from 0 to 80 °C, as shown in Fig. 4b. A comparison of the PLL-tracking frequency response illustrated in Fig. 4b with the results in Fig. 3b distinctly reveals the anomalous frequency shift around 62 °C. To quantitatively characterize the temperature insensitive region, the effective temperature bandwidth (ETB) is defined as the continuous temperature interval where the relative frequency deviation |∆f/f0| remains within a specific threshold. For the primary plateau of Mode 16, the ETB is calculated as 2.0 °C within a ± 5 ppm range. To further clarify the behavior, Fig. 4c provides a locally enlarged view of the anomalous region. At the lower-temperature end of the interval, the amplitude–frequency response contains a single peak, whose resonant frequency gradually shifts toward lower values. As the temperature rises, a smaller secondary peak emerges on the right side of the original peak. Both the amplitude and the resonant frequency of this secondary peak evolve with temperature: the amplitude increases while the peak frequency continues to drift toward lower values. With further heating, the secondary peak grows and eventually supersedes the original peak, becoming the dominant resonance, while the original peak diminishes and fades away. At this stage, the response reverts to a single-peak profile, consistent with the frequency evolution tracked by the PLL. Following the appearance of the second peak in the frequency response curve, the behavior of both peaks aligns with a well-known energy coupling phenomenon, i.e., Frequency Veering, as denoted by the dotted line in Fig. 4c. This occurrence is likely ascribed to nonlinearities within the mechanical structure of the device, facilitating dynamic energy coupling and distribution between Mode 16 and a sideband of a high-frequency mode. The substantially reduced amplitude in the anomalous interval, as seen in Fig. 4b, further reinforces the presence of such mode-coupling effects. To examine the frequency stability within the anomalous region, three observation points are strategically positioned within the primary section at 60 °C (P1), 62 °C (P2), and 64 °C (P3). P1 and P3 are situated near the edges of the abnormal region, while P2 is positioned at its midpoint. The abnormal region spans approximately 2 °C, as illustrated in Fig. 4; hence, the dynamic-temperature assessments encompass a range of ±1 °C. For the dynamic experiment, a program was devised to regulate the temperature chamber. The protocol entails maintaining the target temperature for 300 seconds, followed by a gradual increase of 1 °C over 150 seconds, sustaining this level for another 150 s. Subsequently, a 1 °C descent occurs within 150 seconds, followed by a 150-second plateau. This pattern is then repeated with a second 1 °C decrease over 150 s and an additional 150-second hold, concluding with a 1 °C ascent in 150 s and a final 150-second stabilization phase, as depicted in Fig. 5a–c. Due to the thermal time constant of approximately 5 ms, this MEMS structure achieves thermal equilibrium almost instantaneously within a 150 s heating and stabilization window, thereby accurately tracking the oven temperature. Throughout this sequence, the PLL function of lock-in amplifier 2 continuously monitors the frequency of Mode 16, with the corresponding time-domain responses at P1–P3 illustrated in Fig. 5a–c. P1 and P3 exhibit the typical negative-TCf behavior under dynamic temperature variations, while P2 demonstrates much smaller fluctuations. Notably, during each temperature-holding interval, the frequency of P2 remains relatively unchanged, indicating significantly enhanced stability. a Temperature-frequency characteristics at P1 (60 °C) in ± 1 °C deviation range with the resonance frequency tracked by PLL. b Temperature-frequency characteristics at P2 (62 °C) in ± 1 °C range. c Temperature-frequency characteristics at P3 (64 °C) in ± 1 °C range. d Stability analysis using Allen Deviation method for P1, P2, and P3 within ± 1 °C deviation range. e PSD analysis of P1, P2, and P3 within ± 1 °C. It is noteworthy that the temperature is dynamically fluctuating for this experiment. f Stability analysis using Allen Deviation method for 62 °C and non-controlling. g PSD analysis of 62 °C and non-controlling To provide a direct comparison, the Allan Deviation of all three points is evaluated within the range of ±1 °C, as shown in Fig. 5d. P1 reaches 92.8 ppb at 1.495 s, P3 reaches 88.0 ppb at 1.540 s, whereas P2 achieves the best stability, which is 17.4 ppb at 46.390 s. Both the short-term and long-term characteristics distinctly illustrate the superior performance of P2 over P1 and P3. This finding underscores that Mode 16 exhibits significantly enhanced frequency stability within the temperature-insensitive region, particularly over extended averaging durations, under the context of dynamic temperature fluctuations. The power spectral densities (PSD) displayed in Fig. 5e corroborate that the noise floors at P1–P3 remain nearly identical. Additionally, precise temperature control experiments are conducted on P2 and compared with conditions lacking temperature control, as illustrated in Fig. 5f, g. At 62 °C with stringent temperature regulation, the DUT attains a stability level of 2.0 ppb within 2.510 seconds, showcasing an overall superior stability compared to the uncontrolled scenario. From the perspective of noise characteristics, the slight degradation in noise performance at 62 °C is likely attributed to thermomechanical noise at elevated temperatures. However, the overall stability improvement achieves in the temperature-insensitive region far outweighs this minor noise loss, demonstrating a favorable balance between frequency accuracy and environmental resilience. Temperature-frequency abnormal region of Mode 18 This section delves deeper into examining the temperature-frequency characteristics of Mode 18, emphasizing the TCf anomalous region, with the general trend outlined in Fig. 6a. Three distinct abnormal regions are identified and labeled as Section 1, Section 2, and Section 3. The responses in Section 1 and Section 3 resemble the locally flat segments observed in the Primary Section of Mode 16. In contrast, the behavior observed in Section 2 is particularly intriguing, showcasing notably magnified positive-TCf characteristics, leading to a more pronounced thermally induced upward shift in frequency. Given that Mode 18 exhibits multiple “temperature-insensitive regions” with thermal characteristics distinct from Mode 16, we have provided individual definitions for its three primary sections to ensure a comprehensive evaluation. Specifically, for Mode 18, Section 1 yields the ETB of 4.5 °C at ±10 ppm threshold, Section 2 shows the ETB of 6.4 °C at ± 50 ppm threshold, and Section 3 results in the ETB of 2.2 °C at ± 10 ppm threshold. These three segments will be the primary focus of the ensuing discussion. a Temperature-frequency characteristics of Mode 18 with three abnormal temperature regions identified across 0 °C to 80 °C, wherein five observation points (P4–P9) are selected within three sections for further investigations. b Open-loop frequency sweep measurement of Mode 18 across 0 °C to 80 °C To clarify the origin of the multiple abnormal regions, the open-loop sweep is performed during the 0 °C–80 °C temperature ramp of Mode 18, and the result is shown in Fig. 6b. A comparison between Fig. 6b and Fig. 5b reveals a dispersion reminiscent of Frequency Veering extending throughout the response. This behavior typically signifies modal interaction involving energy coupling and distribution, potentially stemming from the sophisticated modal interaction entailing Mode 18 and multiple high-frequency sidebands of other modes. Further, concurrent coupling between Mode 18 and various sidebands could lead to a non-monotonic and thermally sensitive frequency shift. To evaluate the stability within each abnormal region, six observation points (P4–P9) are selected across Sections 1–3. P4 (26 °C) and P9 (64 °C) lie near the centers of Section 1 and Section 3, following the same selection strategy as P2, as shown in the zoom-in views of Fig. 6a. Within Section 2, P6 (41 °C) and P7 (43 °C) correspond to its two local extrema, whereas P5 (38 °C) and P8 (46 °C) are positioned near the edges and share the same frequencies as P7 and P6, respectively. Based on the temperature-frequency profile in Fig. 6a, the dynamic temperature range for P4 and P9 is set to ±1 °C using the same chamber program as in the previous section. For P5–P8, the range is narrowed to ±0.5 °C, with the temperature step refined from 1 °C to 0.5 °C. The resonant frequency of Mode 18 is tracked using the PLL function of lock-in amplifier 2, and the stability analysis employing Allan Deviation, as well as the noise performance assessed through Power Spectral Density (PSD) at six observation points, are consolidated in Fig. 7. The time-domain data for P4 and P9 within a dynamic temperature range of ±1 °C are displayed in Fig. 7a and Fig. 7j, respectively. Both P4 and P9 exhibit consistent frequencies within this temperature range, showcasing high stability. P4 demonstrates superior frequency stability compared to P9, attributed to P4 operating at a lower temperature than P9. Within the ±0.5 °C range, the time-domain data of P5–P8 are shown in Fig. 7d–g. P5 and P8 follow a negative-TCf trend, while P6 and P7 exhibit minimal variation, consistent with the local response shape of Section 2. Specifically, P5 reaches 168.6 ppb at 1.495 s, P8 reaches 152.4 ppb at 1.585 s, P6 reaches 44.4 ppb at 22.260 s, and P7 achieves the best stability, which is 37.9 ppb at 53.420 s. Both the short-term and long-term stability of P6 and P7 outperform those of P5 and P8, highlighting the stabilizing effect of temperature-insensitive regions. Additionally, Fig. 7i indicates that the noise levels of P5–P8 remain comparable. a Temperature-frequency characteristics at P4 (26 °C) in ± 1 °C deviation range with the resonance frequency tracked by PLL. b Stability analysis using Allen Deviation of P4 in section 1 within ± 1 °C. c PSD analysis of P4 in section 1 within ± 1 °C. d Temperature-frequency characteristics at P5 (38 °C) in ± 0.5 °C range. e Temperature-frequency characteristics at P6 (41 °C) in ± 0.5 °C range. f Temperature-frequency characteristics at P7 (43 °C) in ± 0.5 °C range. g Temperature-frequency characteristics at P8 (46 °C) in ± 0.5 °C range h Stability analysis using Allen Deviation of P5-P8 in section 2 within ± 0.5 °C range. i PSD analysis of P5-P8 in section 2 within ± 0.5 °C range. j Temperature-frequency characteristics at P9 (64 °C) in ± 1 °C range. k Stability analysis using Allen Deviation of P9 in section 3 within ± 1 °C range. l PSD analysis of P9 in section 3 within ± 1 °C range. It is noteworthy that the temperature is dynamically fluctuating for this experiment Discussion The experimental outcomes reveal an unconventional temperature-frequency response in the curved-beam piezoelectric MEMS resonator, diverging significantly from the usual monotonic frequency decline associated with the thermal softening of silicon. A critical discovery of this research is the observation of localized “temperature-insensitive plateaus” in the high-frequency regions, particularly in Mode 16 and Mode 18. The observed temperature insensitive plateaus are likely associated with a dynamic balance where the intrinsic thermal softening of the silicon material is countered by nonlinear strain stiffening effects governed by the Bézier curvature. In this proposed mechanism, the cubic displacement terms within the nonlinear governing equation may facilitate localized frequency corrections through modal coupling to maintain a relatively stationary frequency response against temperature fluctuations. Within these discrete temperature windows, the resonance frequency remains remarkably stable, effectively circumventing the intrinsic thermal drift limits of the structural material, while the tolerable temperature fluctuation range can exceed ±1 °C. This anomalous behavior likely originates from modal coupling mediated by the geometric nonlinearity of the curved clamped-clamped beam. In the proposed 3rd-order Bézier structure, non-uniform thermal expansion (attributed to the nonlinear curve structure) induces a continuous redistribution of internal stress fields, potentially altering the effective stiffness of different modes at different rates. This process can cause Mode 16 and an adjacent mode or a localized parasitic sideband to enter a state of near degeneracy within specific temperature windows. Experimental frequency sweeps reveal the emergence and evolution of the secondary resonance peak that interacts with the primary mode, manifesting characteristics analogous to frequency veering. However, based on the current experimental evidence, it remains challenging to definitively distinguish whether this phenomenon arises from classic frequency veering or more complex non-Hermitian dynamics, such as Exceptional Points (EPs) or Anti-Parity-Time (APT) symmetry breaking. Discriminating between these underlying coupling regimes requires further rigorous theoretical modeling and high-resolution experimental verification in future studies. Currently, this interaction facilitates a dynamic energy redistribution that allows the resonance frequency to deviate from its standard negative trajectory with respect to the TCf. The physical robustness of these identified plateaus is underscored by the consistency of the stability performance across different thermal environments. While Mode 16 and Mode 18 achieve ppb-level stability, the significance lies in the fact that the PSD noise floors remain uniform regardless of whether the device is operating inside or outside the plateau region. This observation confirms that the suppression of temperature sensitivity is an intrinsic deterministic mechanism rather than a stochastic reduction in fundamental noise. Although the temperature-insensitive region is an intrinsic feature of the design, its practical performance remains subject to manufacturing tolerances. To ensure high precision frequency stability, a one-time individual calibration is required for each device to pinpoint its specific plateau window. Furthermore, the ability to maintain such high stability under dynamic thermal fluctuations validates that the modal coupling is strong enough to buffer real-world environmental disturbances, providing a reliable physical basis for passive frequency compensation. Building upon these findings, the “self-compensation” observed in this nonlinear architecture provides a fertile ground for further investigation. Future efforts will first focus on the strategic optimization of the beam’s curvature and internal prestress to expand the “plateau window,” ensuring reliability over broader industrial temperature ranges or attaining a manipulatable plateau region. Concurrently, establishing a comprehensive analytical framework that integrates geometric nonlinearity with thermo-mechanical dynamics is essential to precisely engineer these TCf-insensitive windows across multiple vibration modes. Beyond structural tuning, exploring the influence of higher-order nonlinearities on coupling strength may further push the frequency stability toward the fundamental limits defined by thermomechanical noise. These advancements will be instrumental in evolving structural nonlinearities into autonomous, high-stability timing standards for next-generation low-power microsystems. This work demonstrates a multi-modal curved-beam piezoelectric MEMS resonator designed to achieve enhanced frequency stability. By employing a standard SOI-AlN process, the device leverages piezoelectric transduction to detect out-of-plane and bulk modes, which often pose detection challenges for conventional capacitive structures. The total of 17 distinct vibration modes, ranging from in-plane flexure modes to out-of-plane modes, are successfully captured within a 5 MHz frequency span. The significant finding is the existence of “temperature-insensitive plateaus” (TCf abnormal region) in high-frequency modes (Mode 16 and Mode 18). Comprehensive open-loop sweeps and PLL tracking characterize the underlying modal coupling. Within specific temperature intervals, the secondary peak emerges and eventually dominates the primary response—a behavior analogous to the frequency veering effect. This occurrence is likely ascribed to nonlinearities within the mechanical structure of the device, facilitating dynamic energy coupling and distribution between the operating mode and a sideband of a high-frequency mode. Experimental data show that Mode 16 achieves the frequency stability of 17.4 ppb (averaging time: 46.390 s) at a 62 °C plateau under ±1 °C thermal fluctuations. Similarly, Mode 18 exhibits multiple stability plateaus, reaching the best stability of 37.9 ppb (averaging time: 53.420 s) at 43 °C. In summary, the proposed resonator, characterized by its simplified fabrication, eases the implementation of active compensation and resilience to environmental factors, emerging as a competent contender for ppb-level frequency references in advanced timing applications. Materials and Methods Fabrication The proposed device is manufactured on an n-type SOI wafer using an AlN piezoelectric process, with the microfabrication sequence outlined in Fig. 1a. The completed structure comprises five functional layers: the pad metal layer, AlN piezoelectric film, isolation oxide, device silicon, and silicon substrate. The fabrication process commences by thermally oxidizing the 10 μm thick device silicon layer to create a 0.2 μm SiO₂ isolation layer, ensuring electrical insulation for grounding the silicon. Subsequently, the oxide layer is patterned using photolithography and reactive-ion etching (RIE). A 0.5 μm c-axis oriented AlN film is then deposited through reactive sputtering, ensuring consistent thickness across the wafer. Contact windows in the AlN layer are subsequently defined through a wet-etching step. Metallization follows, involving electron-beam evaporation of a 1.02 μm metal layer, which is then patterned through lift-off to form the electrode structures. Device features are delineated by deep reactive-ion etching (DRIE) of the device silicon layer with a lithographically patterned photoresist mask. The process is finalized with backside DRIE to eliminate the silicon substrate beneath the beams, creating a 400 μm cavity that releases the suspended structures while preserving their mechanical integrity. Experimental Platform The device under test (DUT) in this study is a curved clamped–clamped beam resonator with a boundary length of 800 μm and a boundary width of 70 μm. The optical micrograph of the fabricated device is shown in Fig. 8b, and the illustrative three-dimensional model is depicted in Fig. 8c. a Demonstration of the experimental platform. b Actual image of the as-fabricated MEMS device. c Illustrative model of the proposed device with labelled layers The experimental platform is illustrated in Fig. 8a. Owing to the transduction mechanisms of the piezoelectric effect and the devised structure, the DUT can operate directly in air without the need for vacuum encapsulation, being affixed to a Printed Circuit Board (PCB) interface board. During the experiment, the PCB is situated within a temperature chamber, with input and output signals routed through a sealed feedthrough on the chamber wall to the lock-in amplifier. Moreover, an external trigger from a function generator synchronizes data acquisition with the lock-in amplifier. To gain a clearer understanding of the platform setup, the relevant instrumentations utilized in this study are detailed in Table 1. Table 1Information on the instrumentsEquipmentDetailsLock-In Amplifier 1Zurich Instrument HFLI 5 MHzLock-In Amplifier 2Zurich Instrument HF2LI 50 MHzTemperature ChamberEspec SH-261Frequency GeneratorKeysight 33500B Information on the instruments The frequency response of the DUT is measured using lock-in amplifier 1, where a 500 mV AC excitation is applied under temperature variations from 0 °C to 80 °C. The excitation signal is provided from the lock-in amplifier, while the temperature change is controlled by the temperature chamber. Acknowledgements This work was supported by the Science and Technology Development Fund, Macau SAR (FDCT), 0051/2024/AGJ, the University of Macau, MYRG/GRG2024/00025/IME, the University of Macau, SRG2023-00051-IME, the Science and Technology Development Fund, Macau SAR (FDCT), 004/2023/SKL, and the Funding Scheme for Scientific Research Platforms National Platforms, 0001/2025/NRP. Author contributions Y.L. was responsible for system implementation, measurement, validation, writing the original draft and editing drafts; Y.L., F.C., Z.Z., J.Z., J.L., and J.Y. were responsible for measurement and experimental platform setup; J.C. and K.L. were responsible for device fabrication and editing draft; Y.W. was responsible for supervision, reviewing, editing draft and funding acquisition; C.W. and M.K. were responsible for advising, reviewing and editing drafts; R.M. and P.M. were responsible for funding acquisition and reviewing drafts. Conflict of interest The authors declare no competing interests. References