Bayesian-Optimized Collapse-Mode CMUT with Trenched Membrane for High Output Pressure Authors: Yuanyu Yu, Xin Liu, Jiujiang Wang, Shuang Zhang, Sio Hang Pun Published: 2026 ABSTRACT Capacitive micromachined ultrasonic transducers (CMUTs) have been extensively investigated for applications in medical imaging and industrial non-destructive testing. However, their relatively low acoustic pressure output remains a major limitation to broader adoption. This paper proposes a CMUT structure that combines collapse-mode operation with a trenched membrane to enhance output performance. An analytical model based on von Kármán large-deflection plate theory is developed to estimate the optimal radial position of the trench, thereby defining the search space for subsequent Bayesian optimization. Single-parameter sequential Bayesian optimizations are first performed to identify the individual effects and optimal ranges of the trench’s radial position, depth, and width. Subsequently, a three-parameter global Bayesian optimization framework is employed for global parameter refinement. The three-parameter joint optimization reveals strong synergistic interactions among the design variables, achieving a higher output pressure of 72.34 kPa compared to 70.02 kPa from sequential approaches. Under identical operating conditions, the optimized trenched membrane CMUT exhibits a 100.15% increase in output acoustic pressure and a 28.77% improvement in the pressure-bandwidth product compared to a uniform membrane. Statistical analysis across multiple independent runs yielded a coefficient of variation (CV) of only 0.052% for the output pressure in the three-parameter global optimization results. This confirms that the proposed framework robustly optimizes CMUT designs for high output pressure, offering a promising technical approach to enhancing device performance. 1. Introduction CMUTs are devices that utilize electrostatic actuation to convert electrical energy into acoustic energy and vice versa. They are typically fabricated using silicon-based micro-electro-mechanical systems (MEMSs) techniques, representing a significant shift in ultrasonic transducer technology from traditional piezoelectric materials to semiconductor integration. Compared to piezoelectric ultrasonic transducers like lead zirconate titanate (PZT), CMUTs offer advantages such as high bandwidth and the ability to be integrated with CMOS circuits on a single chip. CMUTs have been explored in nearly all ultrasound applications [1]. Despite significant advancements in CMUT technology, there remain some challenges in its applications. One limitation is the relatively low output acoustic pressure of CMUTs in transmission mode, which restricts their use in medical ultrasound imaging and can compromise image quality. For example, Perez et al. conducted a cross-sectional study on handheld ultrasound devices for point-of-care ultrasound (POCUS), where 35 physicians compared ultrasound imaging of the right upper quadrant of the abdomen, the apical four-chamber view of the heart, as well as superficial views of the neck and lungs [2]. In their comparative study, it was found that the handheld ultrasound device Butterfly iQ+, which uses CMUT technology, scored lowest across imaging quality metrics, including detail resolution, contrast resolution, penetration, and clutter. It is mainly caused by the relative low output acoustic pressure of CMUTs. Therefore, increasing the output acoustic pressure of CMUTs is a research direction to enhance their performance. The output acoustic pressure of CMUTs is positively correlated with the membrane vibration amplitude. Theoretically, increasing the cavity height of the CMUT can enhance the output acoustic pressure; however, this approach may result in a decrease in receiving sensitivity. Currently, the methods to improve output acoustic pressure mainly include altering the operating mode, modifying the device structure, or combining both approaches. In both transmission and reception modes, CMUTs require an external DC bias voltage to operate at their working point. One way to enhance the output acoustic pressure of CMUTs is to change the applied DC bias voltage, allowing the device to enter unconventional operating modes. Research has shown that the collapse-snapback mode [3], collapse-mode [4,5,6], and deep-collapse mode [7] can all effectively enhance the output acoustic pressure of CMUTs. Among these operation modes, only the collapse-mode can improve both the transmission and reception performance of the CMUT while also increasing the bandwidth [8]. Collapse-mode refers to the operational state of a CMUT in which the applied DC bias voltage remains consistently above the pull-in voltage. In this mode, the central part of the membrane is in contact with the insulating layer of the substrate, while the annular region of the membrane that has not collapsed is responsible for transmitting and receiving ultrasound waves. In the white paper published by Health.E Lighthouse in 2021, several companies and institutions in Europe participated in benchmark testing of MEMS ultrasound transducers, including conventional CMUTs, collapse-mode CMUTs, and piezoelectric micromachined ultrasonic transducers (PMUTs). The results indicated that the collapse-mode CMUTs exhibited the highest peak-to-peak pressure and demonstrated the best overall acoustic performance, along with comparable round-trip sensitivity [9]. Consequently, collapse-mode represents a promising operational mode for enhancing the output acoustic pressure of CMUTs. The output acoustic pressure of a CMUT is generated by the vibration of the membrane. Therefore, increasing the displacement during membrane vibration can lead to a higher output pressure. Since the edges of the membrane are fixed to the support, the amplitude of vibration is maximized at the center of the membrane, while the amplitude near the edges approaches zero. As a result, the average displacement of the membrane’s vibration is constrained by this structure. The output acoustic pressure can also be improved by modifying the structure of the CMUT to enable piston-like vibration of the membrane. One approach involves designing a non-uniform membrane that alters the membrane structure to achieve this piston-like motion. This design specifically includes features such as a center-mass and trench structures in the membrane. The center-mass method involves introducing a mass at the location of maximum amplitude of the membrane, which is at the center. This structure increases the local inertia of the membrane, facilitating the maintenance of coherent motion throughout the membrane. As a result, the piston-like mode becomes the dominant mode, leading to improvements in the output acoustic pressure [10,11]. Additionally, incorporating trench structures into the membrane can modify its stiffness, allowing the membrane to move more freely as a whole during applied forces in a piston-like manner. It enhances the volume displacement of the membrane and consequently increases the output acoustic pressure. Zhou et al. conducted a study using finite element analysis (FEA) to investigate the design of annular trenches on the membrane of conventional CMUTs. This approach can increase the membrane displacement, leading to more uniform deformation and ultimately enhancing the output acoustic pressure [12]. In their study, the output acoustic pressure of the CMUT with a single trench-structured membrane was improved by 14.4%. Jiang et al. investigated the reception performance of trench-structured membrane CMUTs, finding that a single trench structure improved the reception performance by 38.3% [13]. Their work indicates that trench-structured membranes can simultaneously enhance both the transmission and reception performance of CMUTs. In addition to non-uniform membrane techniques, enhancements can be achieved by modifying the electrode structure or cavity design to increase the acoustic output pressure. Guldiken et al. proposed a dual-top-electrode structure for CMUTs to enhance the output acoustic pressure. The principle of this design is to utilize the lever bending effect to increase the effective vertical displacement of the membrane and the overall volume displacement, thereby improving the output acoustic pressure [14,15]. Li et al. introduced a ring-shaped electrode structure that employs the electrostatic spring softening effect to alter the overall stiffness of the membrane. This modification causes the membrane to vibrate with piston-like deformation when subjected to force, thereby enhancing the output acoustic pressure [16]. Li et al. proposed a T-shaped cavity structure for CMUTs [17], while Jia et al. introduced “2-stage-Optimal” and “3-stage” cavity structures for CMUTs [18]. These methods all incorporate stepped cavity designs, which increase the electric field strength at the edges of the membrane. Additionally, they utilize the electrostatic stiffness softening effect to adjust the membrane’s stiffness, allowing the entire membrane to undergo piston-like deformation, thereby enhancing the output acoustic pressure. Additionally, Yuan et al. introduced a design method that combines an annular electrode with a membrane groove to enhance the output acoustic pressure of CMUTs. The annular electrode utilizes the electrostatic stiffness softening effect to partially adjust the stiffness of the membrane, while the groove effectively releases stress at the edges of the membrane, promoting piston-like deformation and increasing the average displacement [19]. Kim et al. designed a circular membrane and electrode structure with a separated suspension. By employing an indirectly clamped method, the circular membrane experiences piston-like deformation under electrostatic forces, thereby enhancing the output acoustic pressure [20]. Li et al. proposed a dual-layer CMUT design method that stacks a top circular CMUT cell on top of a bottom annular CMUT cell. This configuration reduces stiffness and allows the deflections of both layers to combine, making it easier for the membrane to deform under external forces and thus improving the output acoustic pressure [21]. Our research group previously proposed a method that combines collapse-mode operation with a center-mass design, referred to as an embossed pattern, to enhance the output acoustic pressure of CMUTs. Specifically, an annular mass structure is added at the center of vibration of the collapsed membrane to form the embossed pattern. Both FEA simulations and device experiments have demonstrated that this method can further increase the output acoustic pressure of collapse-mode CMUTs [22,23]. Previous research has demonstrated that introducing trench structures on CMUT membranes can enhance output acoustic pressure. This study extends this concept to collapse-mode CMUTs by incorporating a trench into the vibrating membrane. To the best of the authors’ knowledge, the integration of collapse-mode operation with three-parameter trench-geometry optimization has not been reported previously. The primary contributions of this work are as follows: Firstly, the vibrating membrane of the collapse-mode CMUT is analyzed using the von Kármán large-deflection plate theory to analytically estimate the optimal radial position of the trench. Secondly, a three-parameter Bayesian optimization framework is implemented to globally optimize the radial position, depth, and width of the trench. Thirdly, the derived analytical criterion is utilized to constrain the search space for Bayesian optimization, thereby effectively coupling physical modeling with a data-driven method while enhancing computational efficiency. This paper is organized as follows. Section 2 describes the trenched membrane CMUT design, FEA models, and analysis methods. Section 3 presents the validation of the proposed approach. Section 4 reports the optimization process and results. Section 5 discusses the findings, and Section 6 concludes the paper. 2. Methodology 2.1. Conception of a Collapse-Mode CMUT with Trenched Membrane CMUTs are electrostatic transducers that operate by applying a DC bias voltage to deflect the membrane, followed by the superposition of an AC voltage or pulse to drive membrane vibration, thereby generating acoustic pressure. In conventional CMUTs, the applied DC bias voltage is maintained below the pull-in voltage. Once the DC bias voltage exceeds the pull-in voltage, the membrane collapses and comes into contact with the substrate insulating layer. The uncollapsed annular region of the membrane surrounding the collapsed area then vibrates to generate acoustic pressure. A trench structure is introduced into the annular vibrating membrane to improve acoustic pressure, as shown in Figure 1. Whether for a circular vibrating membrane in conventional mode or an annular vibrating membrane in collapse mode, the output acoustic pressure of CMUTs is inherently related to the amplitude of the membrane’s vibration. As mentioned above, the design of the trench structure reduces the local stiffness of the vibrating membrane, thereby increasing the volume displacement of the membrane and enhancing the output acoustic pressure. 2.2. Analytical Analysis of Optimal Trench Position To achieve the maximum increase in acoustic pressure, it is essential to investigate the optimal placement of the trench. In previous studies of conventional mode CMUTs, trenches are typically designed near the edge of the membrane [12,13,24]. However, since the vibrating membrane in collapse-mode CMUTs differs substantially from that in conventional-mode CMUTs, it is necessary to investigate the optimal trench position to enhance acoustic pressure, taking into account the characteristics of collapse-mode operation. 2.2.1. Analytical Model for Optimal Trench Position Figure 2 shows the cross-section view of a CMUT with trenched membrane in collapse-mode. The membrane is divided into two sections; the inner contact portion is constrained by the strong electrostatic force at the insulating layer of the substrate, as indicated at point A in Figure 2. The outer vibrating portion remains uncollapsed, which is fixed to the support, as shown at point B. In the transmitting mode, only the uncollapsed membrane vibrates. Therefore, the analysis focuses on the vibrating portion of the membrane. The dynamic behavior of the membrane is analyzed using von Kármán large-deflection plate theory, which accounts for geometric nonlinearity in the static deflection. For the small-signal AC dynamic response, the system is linearized around the statically deformed state. The amplitude of the AC voltage is small compared to the DC bias, and the residual stress in the membrane is neglected. This analysis helps estimate the optimal radial position of the trench and narrows the search space for Bayesian optimization. For the collapse-mode CMUT, the overall radius of the membrane is denoted as R. Under the influence of a DC bias voltage, the central region of the membrane is closely adhered to the substrate, forming a fixed contact area. The radius of the contact boundary is represented as r_c, which is uniquely determined by the DC bias voltage. The effective vibrating area of the membrane is limited to the annular non-contact region (r_c65 kPa, indicated in red) are densely clustered within a restricted three-parameter subspace characterized by a radial position of approximately 17.3–18.5 μm, a depth of approximately 0.35–0.41 μm, and a width of approximately 0.8–2.0 μm. The sharp decline in acoustic pressure outside this cluster suggests that high performance is only achievable through a precise combination of these variables. Such a distribution confirms the existence of strong synergistic interdependencies between the parameters, validating the requirement for simultaneous three-parameter global optimization rather than individual parameter tuning. Figure 25 presents the distribution of all sampling points via projections of the three-dimensional scatter plot onto the radial position–width, radial position–depth, and width–depth orthogonal planes. The distribution in the radial position–width plane in Figure 25a reveals that the high-performance region (>70 kPa) exhibits a distinct unimodal distribution along the radial position dimension, concentrated within a narrow range of 17.5–18.1 μm for radial position and 1.1–1.9 μm for width. When the radial position deviates below 17.3 μm or above 18.2 μm, the acoustic pressure drops sharply to below 60 kPa even if the width remains within its optimal range, indicating that radial position plays a dominant role in device performance. Its effective operating window is only approximately 0.6 μm, representing the most stringent precision requirement. Conversely, when the radial position is optimized, the acoustic pressure remains above 65 kPa for widths ranging from 1.0 to 2.0 μm. This suggests that the performance is relatively insensitive to width variations in this range, thereby offering a higher fabrication tolerance. The sampling point distribution in the radial position–depth plane of Figure 25b indicates that the high-performance region (>70 kPa) is similarly concentrated within the radial position range of 17.5–18.1 μm and a depth range of 0.39–0.41 μm, with an increasing density as the depth approaches 0.41 μm. Within the depth interval of 0.36–0.41 μm, the acoustic pressure can be maintained above 65 kPa as long as the radial position remains optimal, suggesting a high fabrication tolerance for the depth parameter. However, when the depth is less than 0.35 μm, the acoustic pressure generally drops below 63 kPa, even with an optimal radial position. This occurs because shallow trenches have a limited impact on membrane stiffness, resulting in a constrained increase in the output acoustic pressure. The resulting band-like distribution further confirms that radial position is the primary determinant of performance. The sampling point distribution in the depth–width plane of Figure 25c is the most dispersed among all projections. The high-performance region (>70 kPa) is scattered across a broad parameter space, specifically within a depth of 0.38–0.41 μm and a width of 0.9–2.0 μm, with the highest concentration observed in the 1.2–1.6 μm width range. However, when the depth is less than 0.35 μm, the acoustic pressure fails to exceed 63 kPa irrespective of the width, indicating a clear performance threshold where the acoustic pressure drops significantly. Consequently, the trench depth should ideally approach 0.41 μm to ensure high output acoustic pressure. The three-parameter global optimization results indicate that radial position governs the device performance due to its high sensitivity, with an effective performance window of approximately 0.6 μm. The abrupt drop in pressure upon deviation highlights the radial position as the most critical parameter for process stability. A general positive trend exists between pressure and depth, where values approaching 0.41 μm yield superior output, aligning with the sampling point distribution shown in Figure 17. Under optimized radial and depth conditions, the width exhibits significant robustness over a wide interval, offering the greatest fabrication tolerance. This three-parameter analysis reinforces the consistency of the trends identified in the preceding single-parameter sequential studies. The optimal sampling point identified from 12 independent three-parameter global Bayesian optimization runs is located at a radial position of 17.70 μm, a depth of 0.41 μm, and a width of 1.31 μm, corresponding to a peak acoustic pressure of 72.34 kPa. Figure 26 illustrates the convergence process of a representative three-parameter global Bayesian optimization run. During the first 18 iterations, a Sobol sequence was employed to generate initial sampling points, ensuring uniform coverage across the three-parameter space to establish the initial training dataset for the GP surrogate model. A significant performance leap occurred at the 19th iteration, marking the successful localization of the high-performance region. Iterations 19 to 33 represent the deep optimization phase, where the observed optimal values improved steadily, the search range narrowed, and the surrogate model predictions showed high consistency with the measured values. The global optimum within the considered design space was identified at the 34th iteration, and the observed value remained constant from iterations 35 to 80. The absence of better solutions over consecutive evaluations confirms that the optimization achieved stable global convergence. This figure demonstrates that Bayesian optimization can effectively locate high-pressure regions even when three parameters vary simultaneously with inter-parameter coupling. 4.3.2. Repeated Runs and Stability of Three-Parameter Global Optimization Statistical analysis was performed on 12 optimal sampling points from repeated independent three-parameter global Bayesian optimization runs, as summarized in Table 6. The mean radial position is 17.71979 μm with a CV of 0.139%, exhibiting a fluctuation range of only approximately 0.077 μm. The CV for depth is 0.084%, the lowest among all parameters, indicating that depth possesses the highest convergence stability. Although the width shows a higher CV of 2.591%, the corresponding fluctuation in acoustic pressure is only about 105 Pa, which validates the previously discussed existence of a flat response interval near the optimal width. The mean acoustic pressure across the 12 runs is 72,289.77 Pa with a standard deviation of only 37.33 Pa (CV = 0.052%). These results demonstrate that the optimization method consistently converges to nearly identical optimal solutions regardless of stochastic initial conditions, thereby exhibiting superior engineering reliability. The clustering characteristics of high-pressure points in Figure 24 further validate the efficacy of the proposed method. The convergence trajectories for 12 independent three-parameter global Bayesian optimization runs are depicted in Figure 27. By employing a Sobol sequence for the 18 initial deterministic samples, the optimization paths coincide initially before diverging during the adaptive exploration phase. The iterations required for steady-state convergence span from 34 to 77, with a mean of 64 evaluations. Following convergence, the curves plateau, signifying that no marginal performance gains are achieved. The predefined budget of 80 evaluations successfully captures the convergence inflection points for all trials, ensuring that even late-converging cases reliably identify the global optimum within the considered design space without premature termination. The pressure CV of 0.0516336% in Table 6 underscores the sufficiency of this 80-iteration limit for the three-parameter trench coupling optimization. This budget strikes an optimal balance between ensuring complete convergence and minimizing FEA computational overhead. The ability of multiple independent trials to consistently pinpoint the same global performance regime demonstrates the robust search stability and reliability of the proposed Bayesian optimization framework. 4.3.3. Optimized Trenched Membrane Design Across the 960 total sampling points generated from the 12 independent rounds of three-parameter global Bayesian optimization, the maximum output acoustic pressure was identified as 72.34 kPa. The corresponding optimal configuration was found at a radial position of 17.70 μm, a depth of 0.41 μm, and a width of 1.31 μm. This peak value, consistently approached by all independent trials, represents the global optimum within the considered design space. Figure 28 shows the comparison of output acoustic pressure between the optimized trenched membrane CMUT and the uniform membrane CMUT under the same conditions. The trenched CMUT achieved a peak acoustic pressure of 72.34 kPa at 4.25 MHz, with a 3-dB bandwidth of 0.92 MHz. In comparison, the uniform membrane CMUT obtained a peak acoustic pressure of 36.14 kPa at 5.23 MHz, with a 3-dB bandwidth of 1.43 MHz. The output acoustic pressure of the optimized trenched CMUT was improved by 100.15%. Here, the pressure-bandwidth product is introduced as an indicator to evaluate the output acoustic pressure and bandwidth performance of the CMUTs [36]. The pressure-bandwidth products of the trenched membrane CMUT and the uniform membrane CMUT were 66.56 kPa MHz and 51.69 kPa MHz, respectively, resulting in an improvement of 28.77%. It can also be observed that, compared to the uniform membrane CMUT, the central frequency decreases when adding the trench structure to the vibrating membrane. The primary reason is that the trench structure reduces the equivalent stiffness of the membrane. Figure 29 shows the radial distribution of membrane deflection for the trenched and uniform membrane CMUT under 130 V DC bias and excitation of 1 V AC voltage. The trenched membrane exhibits a significantly larger peak deflection (20.71 nm) at the center, which is approximately 3.98 times that of the uniform membrane (5.20 nm). Moreover, the deflection profile of the trenched membrane is more concentrated near the center, resembling a piston-like vibration, while the uniform membrane shows a broader, more diffuse deflection distribution. This improvement occurs because the trench structure makes the membrane less stiff in the vibrating area. This reduction in stiffness helps to loosen mechanical limits and allows for a more uniform and effective vibration. To quantitatively evaluate the effective vibration output, the volume displacement (proportional to volume velocity at a given frequency) is calculated as the area-weighted integral of membrane displacement:(35)V=∫0Rw(r)·2πrdr The volume displacement was 14.35 μm3 for the trenched membrane CMUT and 4.51 μm3 for the uniform membrane CMUT, yielding a ratio of 3.18. The resonant frequency of the trenched membrane was 4.25 MHz, which was lower than the 5.23 MHz of the uniform membrane. Nevertheless, the substantial increase in volume displacement fully compensates for the frequency reduction, resulting in a larger volume velocity and thus a higher output acoustic pressure. 5. Discussion 5.1. Advantages of Bayesian Optimization in CMUTs Design Theoretical and FEA simulations indicate that introducing a trench on the membrane of a collapse-mode CMUT can significantly enhance the output acoustic pressure. The trench structure involves three spatial parameters of radial position, depth, and width. However, there is no explicit analytical relationship exists between these three parameters and the resulting acoustic pressure. If traditional grid search methods were used, optimizing the combinations of the three parameters could require up to thousands of FEA simulations, resulting in significant consumption of computational resources and a high likelihood of getting trapped in local optima. Bayesian optimization is a global optimization method designed for black-box functions, particularly suited to evaluating costly objective functions that lack analytical expressions. Its core advantage lies in its ability to approximate the global optimum with very few evaluations, typically in the range of tens to a few hundred [37]. This method constructs a probabilistic surrogate model of the objective function and intelligently balances exploration and exploitation through an acquisition function. It significantly outperforms traditional grid search or random search in terms of sample efficiency, uncertainty quantification, and handling complex constraints. In this study, to further enhance the efficiency of Bayesian optimization in trenched membrane design, the mechanical behavior and acoustic radiation characteristics of the annular vibrating membrane in the collapse mode were first analyzed using von Kármán large-deflection plate theory. This analysis identified the positions on the membrane that contribute most significantly to acoustic radiation, allowing for an estimation of the optimal radial position for the trench. Based on this theoretical estimation, the search space for the trench’s radial position was narrowed from [10 μm, 25 μm] to [16 μm, 19 μm], effectively reducing the search space to 20% of its original range. Consequently, for the same total number of iterations, the sampling density in the high-potential region is increased, leading to significantly faster convergence. For the single-parameter sequential optimizations, the radial position, depth, and width required only 8, 7, and 8 simulation iterations, respectively, to reach convergence, all of which are less than half of the preset total iteration number of 24. Based on these findings, a three-dimensional global optimization was further conducted to jointly optimize the trench radial position, depth, and width. The three-parameter optimization, consisting of 80 evaluations, achieved a higher output pressure than the sequential single-parameter approach (72.34 kPa vs. 70.02 kPa), confirming that the three parameters are strongly coupled and their synergistic interactions cannot be captured by sequential single-parameter approaches. This “theory-guided Bayesian optimization” strategy provides a simple and effective starting point for trench design in collapse-mode CMUTs and can be extended to the optimization design of other CMUT structural parameters. 5.2. Comparison of Different CMUT Structures for Pressure Enhancement This section compares the optimized collapse-mode trenched CMUT with other structural modification strategies reported in the literature for enhancing output pressure, such as embossed membranes, annular electrodes, stepped cavities, and center-mass structures. A comprehensive comparison of these structures is presented in Table 7. The trenched CMUT exhibits a high acoustic pressure output due to its operation in a collapsed mode. Both the trenched membrane and embossed membrane CMUTs operate in collapsed mode. Apart from the specific structural differences between the trench and embossed patterns, all other parameters in the FE models, including dimensions, materials, DC bias, and excitation AC signals, are identical to facilitate a fair comparison. The results demonstrate that the trenched membrane CMUT achieves a 6.67% increase in output acoustic pressure compared to the embossed membrane structure. 5.3. Frequency Shift and Bandwidth Variation As shown in the previous section, the design of the trench structure results in a significant increase in output acoustic pressure. However, this enhancement is accompanied by a decrease in center frequency, which shifts from 5.23 MHz to 4.25 MHz, along with a narrowing of the bandwidth. The downward shift in center frequency is primarily due to the localized reduction in bending stiffness induced by the trench. According to plate theory, the flexural rigidity D is proportional to the cube of the local thickness, and the trench design effectively reduces the overall mechanical stiffness of the vibrating membrane. Although the trench simultaneously removes some mass from the structure, the stiffness reduction dominates the dynamic response during collapse-mode operation, resulting in a decrease in the center frequency. In conventional abdominal ultrasound diagnostics, frequencies of 5 MHz and below are typically sufficient to meet most clinical requirements [38]. For organs such as the liver, pancreas, and kidneys, standard convex array probes generally operate within the 2–5 MHz range. This frequency band provides good penetration and acceptable resolution for typical adults. The optimized frequency of 4.25 MHz remains well within this range, ensuring full compatibility with existing imaging systems. More importantly, the increase in acoustic pressure, which directly enhances the signal-to-noise ratio (SNR) and penetration depth, is achieved at a clinically relevant frequency. While the center frequency and fractional bandwidth of the trenched membrane CMUT slightly reduce, leading to a minor decrease in axial and lateral resolution, the output acoustic pressure increases by over 100%, resulting in a signal gain of +6 dB. Since the acoustic attenuation coefficient of human tissue is proportional to frequency, and considering that abdominal imaging involves deep tissue assessment, the intensity of echo signals tends to diminish due to attenuation. Consequently, the combination of doubled output acoustic pressure and the resulting reduction in attenuation at the lower frequency typically leads to an enhancement in overall image SNR. This ultimately means that details in deeper regions of the image become clearer, thereby improving far-field imaging performance. In the design of trenched membrane CMUT devices, it is essential to select appropriate dimensions and operating points based on the specific application of the ultrasound probe, while also optimizing the trench structure. Furthermore, for devices operating in collapse mode, the center frequency is primarily determined by the width of the vibrating membrane. Consequently, the center frequency of the collapse-mode CMUT can be tuned by adjusting the DC bias voltage. Although an optimal trench position exists, Figure 16 demonstrates that for a fixed trench configuration, the output acoustic pressure remains enhanced relative to the uniform membrane CMUT across a DC bias range of 110 V to 150 V. Specifically, within the range of 120 V to 140 V, the pressure enhancement exceeds 57%. 5.4. Proposed Fabrication Process 5.4.1. Fabrication Process A feasible fabrication process for the trenched membrane CMUT is presented in Figure 30. Similar to our previously developed embossed pattern CMUT fabrication process [39], this is a customized six-mask sacrificial release technique. First, on an N-type heavily doped silicon wafer, SiO2 is grown using dry oxidation, followed by the deposition of Si3N4 as an insulating layer using the low-pressure chemical vapor deposition (LPCVD) method, with polysilicon serving as the sacrificial layer, as shown in Figure 30a. Polysilicon is then etched with Mask #1 and reactive ion etching (RIE) to create the shape of the circular CMUT cell cavity and the etching channels, as illustrated in Figure 30b. Next, low-stress Si3N4 is deposited using PECVD or LPCVD as the structural layer and membrane. With Mask #2 and RIE, etching holes are formed in the release channels surrounding the CMUT cell, creating pathways leading to the polysilicon, as depicted in Figure 30c. The wafer is then immersed in potassium hydroxide (KOH) solution, and wet etching is performed to completely release the polysilicon sacrificial layer from the release holes, forming the cavity structure, as shown in Figure 30d. Following this, PECVD is used to fill the etching holes by depositing Si3N4. Subsequently, Mask #3 and RIE are employed to thin the membrane, restoring its thickness to the initial value, which ultimately seals the cavity, as illustrated in Figure 30e. Using Mask #4 and RIE, an annular trench is etched onto each CMUT’s circular membrane, forming the trench structure shown in Figure 30f. After that, Mask #5 and RIE are used to completely etch away the Si3N4 and SiO2 in the bottom electrode pad region, creating pathways to the heavily doped silicon substrate, as shown in Figure 30g. Finally, a metal layer is deposited using e-beam evaporation, and Mask #6 is applied with dry or wet etching method to form the top electrode pad, bottom electrode pad, and connections, as shown in Figure 30h. 5.4.2. Fabrication Considerations According to the global three-parameter optimization analysis, the key to enhancing the output pressure of trenched collapsed-mode CMUTs lies in the synergistic optimization of the trench’s radial position, depth, and width, rather than the independent contribution of a single geometric parameter. Consequently, the impact of fabrication tolerances and material properties on the performance enhancement must be evaluated during the device manufacturing process. The enhancement of acoustic pressure is highly sensitive to the radial position of the trench, which features a narrow optimal window. This imposes a stringent requirement on alignment precision during fabrication. In contrast, the trench width exhibits high process redundancy, as high output pressure can be maintained across a relatively broad range. Furthermore, the magnitude of pressure enhancement is positively correlated with the trench depth. However, it indicates that when the trench depth exceeds 0.43 μm in a 0.60 μm thick membrane—corresponding to 72% of the thickness and a residual thickness of less than 0.17 μm—the membrane vibration enters a nonlinear regime. It results in an abrupt jump followed by a decay in output pressure, which may compromise device stability. Therefore, the etching depth must be strictly controlled during fabrication to avoid approaching this critical threshold. Membranes are fabricated using either PECVD or LPCVD processes, where the membrane thickness directly governs the center frequency, output acoustic pressure, and collapse voltage. To mitigate the impact of deposition non-uniformity, it is recommended to select dies from the central region of the wafer, where thickness uniformity is better. Dielectric charging represents a critical reliability bottleneck for collapsed-mode CMUTs during long-term operation. Under identical conditions, LPCVD silicon nitride exhibits significantly superior resistance to dielectric charging compared to PECVD nitride due to its lower hydrogen content and charge trap density, making it more suitable for continuous operation. While residual stress in both types of silicon nitride can be maintained within reasonable limits by tuning deposition parameters, geometric discontinuities at the trench roots and sharp corners induce significant stress concentration. These localized stresses may far exceed bulk levels, potentially compromising mechanical integrity over the long term. In contrast, LPCVD silicon nitride offers a distinct advantage in mechanical reliability for trenched structures, owing to its superior compactness and fracture toughness. In summary, the device design and process implementation of trenched collapsed-mode CMUTs require a strategic balance between performance enhancement and process feasibility, taking into full account the combined effects of etching tolerance, thickness uniformity, dielectric charging, and trench-induced stress concentration. 5.5. Limitations and Outlook Although this study has achieved an optimal trenched membrane collapse-mode CMUT structure with high output acoustic pressure using Bayesian optimization and FEA simulation, there are still certain limitations. First, the theoretical model relies on simplified assumptions. The von Kármán large-deflection plate theory guides the prior search space for Bayesian optimization, which accounts for geometric nonlinearity under large static deflection. However, when the CMUT operates under high AC excitation or in a deep-collapse mode, the deformation of the membrane may further exceed the applicability of the current theoretical model. In such cases, discrepancies may arise between the predicted maximum deflection position and the actual value. Second, the surrogate model in Bayesian optimization depends on the accuracy of the FEA simulation. In this study, the fitting of the GP is based entirely on data derived from FEA simulations, which utilize idealized material parameters, ideal clamped boundary conditions, and linear small-deformation assumptions. In reality, CMUT devices often incorporate non-ideal factors during fabrication, such as residual stress gradients and variations in membrane thickness, which are challenging to model accurately in simulations. Consequently, there may be a discrepancy between the optimal trench parameters obtained from simulations and actual devices. Third, the proposed trenched CMUT design has not yet been experimentally validated. While the FEA model employed in this study is developed from the same framework as our previously validated embossed CMUT model [22,23], the specific trenched structure requires direct experimental verification. The absence of fabricated device measurements means that the predicted performance enhancement remains to be confirmed through physical testing. Future research will focus on fabricating trenched membrane CMUT devices based on the process proposed in this study. Experimental tests will be conducted to evaluate their vibration characteristics, center frequencies, and acoustic pressure outputs, thereby validating the effectiveness of the FEA model. In addition, further multi-objective optimization studies will be conducted to comprehensively balance key factors such as acoustic performance, structural reliability, and process tolerances. 6. Conclusions This study presents a systematic approach to enhance the output acoustic pressure of collapse-mode CMUTs by optimizing trench geometry. An analytical model based on von Kármán large-deflection plate theory was established to estimate the optimal trench radial position, which was then used to define the search space for Bayesian optimization. A three-dimensional global Bayesian optimization framework was subsequently employed to globally optimize the trench radial position, depth, and width. The analytical prediction of the optimal trench position is validated by the results of both single-parameter sequential and three-parameter global Bayesian optimizations, which yielded a high degree of consistency with a deviation of less than 1.15% (17.50 μm vs. 17.70 μm). This confirms that the theoretical model effectively captures the radial trend of the volume-velocity integrand and provides a reliable starting point for optimization. The three-parameter joint optimization reveals strong synergistic interactions among the radial position, depth, and width, achieving a higher output pressure than sequential single-parameter approaches (72.34 kPa vs. 70.02 kPa). The consistency between the analytical prediction and the numerical optimization, together with the superior performance of the three-dimensional joint optimization, demonstrates that integrating physical modeling with data-driven optimization offers an efficient and reliable pathway for CMUT structural design. The optimized trenched membrane CMUT demonstrated a 100.15% increase in output acoustic pressure compared to the uniform membrane CMUT, along with a 28.77% enhancement in the pressure-bandwidth product. Future work will focus on the fabrication and experimental characterization of the optimized trenched CMUT, including the evaluation of its vibration characteristics, center frequency, and output acoustic pressure. Further multi-objective optimization studies will also be conducted to comprehensively balance acoustic performance, structural reliability, and process tolerances.