Mechano-optically co-designed highly-scalable silicon photonic MEMS switches with quasi-buckling-free 2 × 2 horizontal adiabatic directional couplers Authors: Yinpeng Hu, Jiayue Zhu, Yunzhi Liu, Linyan Lyu, Ye Lu, Qian Ma, Huan Li, Liu Liu, Yaocheng Shi, Daoxin Dai Published: 2026 ABSTRACT AbstractMicro-electromechanical system (MEMS) photonic switches based on adiabatic directional couplers (ADCs) potentially offer the advantages including large fabrication tolerance and broad bandwidth, which is essential for large-scale reconfigurable photonic integrated circuits in optical networks and artificial intelligence computing infrastructures. However, the scalability of previous ADC-based MEMS switches is challenging due to the complicated fabrication of vertical ADC switches and buckling-induced performance deterioration of horizontal ADC switches. Here we propose a mechano-optically co-designed 2×2 MEMS switch based on a unique quasi-buckling-free horizontal ADC by introducing residual-strain management structures including slab anchors and extra S-bends. For the fabricated device, the waveguide buckling is alleviated significantly and high optical performance is achieved with low loss and low crosstalk over a broad bandwidth of 180 nm. It also features fast switching speed of ~2 μs, reliable durability with >7.2 billion switching cycles, and exceptional scalability with the realization of a 64 × 64 Benes switch array. The present 2 × 2 horizontal ADC switches are compatible with all mainstream array topologies and can be fabricated using simple standard silicon photonic foundry processes, which are not accessible for those 1×2 ADC switches reported previously. With these advantages, the present design provides a highly-scalable solution with great potential for MEMS/NEMS photonic devices used in versatile applications.Subject terms: Micro-optics, NEMS Introduction The rapid development of artificial intelligence, 5 G/6 G networks, and Internet of Things (IoT) has generated substantial demands for data transmission and processing, making photonic information transmission and processing an attractive solution due to its ultrahigh parallelism, large bandwidth, and low latency. Consequently, various large-scale photonic integrated circuits (PICs) have been demonstrated, including on-chip photonic networks1, programmable photonic chips2, photonic field-programmable gate arrays (FPGAs)3, optical neural networks (ONNs)4–6, as well as quantum information processors7,8, where highly-scalable photonic switches are essential components to enable flexible circuit reconfigurations. The scaling of conventional photonic switches9–14 utilizing Mach-Zehnder interferometers or micro-ring resonators is challenging due to the issues, including high power consumption for thermo-optic devices, high carrier absorption loss for electro-optic devices, and sophisticated calibration processes due to the analog nature of the phase shifters. In contrast, silicon photonic micro-electromechanical system15,16 (MEMS) switches17–26, which operate by modulating the mode coupling between waveguides in directional couplers (DCs) or adiabatic directional couplers (ADCs), offer a promising alternative. These switches potentially feature (1) low switching energy consumption of pJ-level and near-zero static power consumption27 for electrostatic actuation and (2) no carrier absorption, making them ideal for large-scale switch arrays. The pioneering works on MEMS switches utilizing DCs have been demonstrated with a large scale up to 50 × 5017. However, their practical applications are compromised due to the narrow bandwidth of DCs. Also, due to the analog nature of their operation principle, the voltage for electrostatic actuation must be accurately controlled for each switch cell, which requires cumbersome electronic circuits for control and drive. More recently, ADC-based MEMS switches21–25 emerged as a promising solution due to their advantages of low excess losses, large bandwidths, digital operation, and exceptional fabrication tolerance. Particularly, vertical ADCs (VADCs)21–23 and horizontal ADCs (HADCs)24,25 are employed successfully for MEMS switches. Unfortunately, both these previously reported VADC- and HADC-based elementary switches are limited to a 1 × 2 configuration, which is incompatible with widely-used topologies such as Benes, PI-Loss, Switch-and-Select, etc.,28,29 while each of these array topologies has its own uniqueness in scalability and functionality. In contrast, this paper focuses on the 2 × 2 elementary switch as the preference for flexible compatibility with various topologies available for diverse switching scales and applications. Besides, it is crucial to achieve high robustness and high performances with simple fabrication processes for elementary switches, especially when scaled to an N×N optical switch array with a large port count. For the VADC-based MEMS switches demonstrated previously21–23, impressive performances were achieved with low excess losses of <0.7 dB, high extinction ratios of >70 dB, large bandwidths of 300 nm and fast ON/OFF switching speed of 400/300 ns. Leveraging the elementary switch with such high performance, a 240 × 240 switch array with Cross-Bar topology23 and a 128 × 128 LiDAR system incorporating a focal plane switch array (FPSA)30 have been demonstrated. However, two silicon layers are required for the VADC design, which are not available in most silicon photonic foundry processes. Consequently, complicated back-end processes are required, which may limit the scalability due to high fabrication cost and compromised yield. Furthermore, the driving voltage of the elementary switch is as large as 60 V or beyond, which complicates the driving circuits due to the difficulties in implementing a multi-channel (100 s or even 10,000 s channels) voltage source with such a high voltage. In contrast, developing HADC-based MEMS switches is recognized as another attractive scheme because it only needs a single layer of silicon and is fully compatible with standard silicon photonic foundry processes, simplifying fabrication processes compared to VADC switches21–23. To date, 1 × 2 elementary switch25 and single-pole double-throw (SPDT) elementary switch24 have been proposed based on HADCs. For the pioneering HADC-based 1 × 2 elementary switch25, a three-comb-drive device design and 8 × 8 Cross-Bar switch array have been implemented. Unfortunately, the fabricated switch exhibits high excess losses of ~4 dB, high crosstalk of ~−10 dB over the bandwidth of 108 nm, and ON/OFF switching speed of 36.7/21.4 μs in experiment. Furthermore, the measured excess loss in the ON state varies up to ~1.7 dB, depending on the ADC gap, which indicates slightly nonadiabatic coupling in the fabricated device. For the HADC-based SPDT elementary switch24, a unique switch design and fast ON/OFF switching speed of 0.82/0.73 μs have been achieved. However, the excess-loss difference between the two output ports is as high as ~5 dB, and the bandwidth is only ~25 nm in experiment. In addition, each SPDT switch requires two, instead of one, electrical pads, which doubles the number of driving circuits and compromises the scalability to the array. These limitations highlight the need for significant performance improvements in HADC-based MEMS switches. It is important to note that suspended waveguide structures used in MEMS switches are susceptible to significant inadvertent structural deformation, such as beam buckling31,32, due to the release of the residual strain in the silicon-on-insulator (SOI) wafer used dominantly for PICs. The resultant waveguide deformation and/or misalignment often leads to significant and random switch performance deterioration, which greatly compromises the scalability. This issue is especially pronounced for digital photonic switches, for which fine-tuning to compensate the coupling change due to the waveguide deformation and/or misalignment is often undesired or infeasible. Therefore, systematic residue strain management is extremely critical for MEMS photonic switches by co-designing the photonic/mechanical structures, which have not been thoroughly investigated previously and will be the key focus here. In this paper, we propose and experimentally demonstrate a highly-scalable 2 × 2 silicon photonic MEMS switch based on a mechano-optically co-designed quasi-buckling-free HADC by introducing extra S-bends and partially-etched taper waveguides with slab anchors. All these structures of the HADC are designed optimally to achieve high photonic switching performances with low excess losses and high extinction ratios over a broad bandwidth. Meanwhile, these structures are also designed mechanically to effectively suppress the waveguide buckling caused by the release of the residual strain in the SOI wafer, which often causes significant misalignment between the two waveguides in conventional HADCs, leading to photonic performance deterioration or even failure. With such a novel structural design, the present HADC switch developed with a single layer of silicon with standard silicon photonic foundry processes is quasi-buckling-free and achieves unprecedented photonic performance for HADC switches (see Supplementary Section 1), including ultra-low excess losses, high extinction ratios, and broad bandwidth, as shown in Table 1. Fast ON/OFF switching speed of 2.2/1.9 μs and reliable durability with >7.2 billion switching-cycles are also demonstrated. Besides, the present MEMS switch features 2 × 2 design instead of the previous 1 × 2 design (see Supplementary Section 1), making it compatible with all widely used array topologies, including Benes, Cross-Bar, PI-Loss, and Switch-and-Select, etc. Furthermore, a 64×64 Benes switch array is realized as a proof of concept with high performances shown in Table 1. The switch array can be scaled to as large as 128 × 128 when the waveguide loss is lowered further by optimizing the fabrication processes, showing great potential for photonic interconnects, programmable PICs, photonic FPGAs, ONNs, quantum information processors, and FPSA LiDAR. Table 1Theoretical and experimental photonic performance of the devicesDevicesData typeStateExcess loss (dB)Crosstalk (dB)Bandwidth (nm)Elementary switchsimulatedOFF0–0.07–52.5 to –28.51400–1700ON0–0.06–47.7 to –21.1measuredOFF0.11–0.43–61.8 to –42.01420–1600 (Limited by the used grating coupler)ON0.05–0.99–24.9 to –10.364 × 64 Benes switch arraymeasuredAll-OFF4.1–9.7≤ –36.71530–1590 (Limited by the used grating coupler)Single-ON5.3–8.8≥ 38.5 (ON/OFF extinction ratio) Theoretical and experimental photonic performance of the devices Results Mechano-optical co-design of the quasi-buckling-free HADC MEMS switch Figure 1a shows the schematic configuration of the present quasi-buckling-free MEMS switch based on a gap-adjustable HADC, which includes a pair of centrosymmetric taper waveguides with a constant gap. Especially, the length of the linear taper waveguides used for ADCs is usually tens or even hundreds of micrometers to ensure adiabatic mode evolution over a broad bandwidth. However, such long suspended waveguides often buckle strongly due to the release of the residual strain in the used SOI wafer, which may cause significant misalignment between the two waveguides in the coupling region. Such misalignment disrupts the adiabatic mode evolution in the ON state, causing a portion of the input power to couple into another supermode that remains within the original waveguide at the output port, leading to higher excess loss and crosstalk. In severe cases, excessive misalignment will completely prohibit the coupling between the waveguides, resulting in switch failure. To address this issue, the taper parts used here for the HADC is designed to be partially etched with the slabs, which then function as the mechanical anchors. In particular, the slab anchors are tapered as well to avoid scattering losses due to the abrupt change of the waveguide cross-section, as shown in Fig. 1a. Additionally, extra S-bends are introduced to further reduce the waveguide buckling (see Fig. 1a). For the suspended waveguides of the HADC, one is connected to the stationary silicon core-layer by a slab anchor, while the other one is connected through another slab anchor to the MEMS actuator consisting of a mechanical stopper, a perforated shuttle beam, a structural aligner, folded springs, and electrostatic combs, which enables the movement in the y direction by the electrostatic actuation33. In the OFF state, the two waveguides are separated with a sufficiently large gap to be decoupled and thus the incident light eventually outputs at the bar port on the same side, as shown in Fig. 1b. In the ON state, the gap between the two waveguides is narrowed electrostatically and the input light is then coupled adiabatically to the adjacent waveguide and output to the cross port on the other side, as shown in Fig. 1c. Principle of the switch. a Schematic diagram of the MEMS structures. b, c Schematic diagram of the gap adjustable horizontal adiabatic coupler (HADC) in OFF and ON states, respectively To achieve a quasi-buckling-free HADC with precise residue strain management, it is essential to first accurately estimate the residual strain in the SOI wafer used. To this end, we have designed, fabricated, and characterized a group of doubly-clamped suspended beams along <110> and <100> crystallographic directions of silicon. The length and width of the beams are 50 μm and 4 μm, respectively. The optical microscope image of the doubly-clamped suspended beams is shown in Fig. 2a. Laser confocal microscope (LCM) with 20 nm z-axis step size is used to measure the buckling of the suspended beams, as shown in Fig. 2b. Measurements show that the z-direction deflection is nearly the same for all beams, regardless of their in-plane orientation or the buckling direction (up or down). Therefore, we show the measured results along A–A’ in Fig. 2a to represent the buckling behavior, as the black circles show in Fig. 2c. Finally, the buckling with different residual strain is simulated with the finite element method (FEM) and the results with isotropic residual strain of 5 × 10–4 agree well with the measured results, as shown by the red curve in Fig. 2c. This confirms the accuracy of the residual strain estimation and provides a foundation for mechano-optical co-design of the quasi-buckling-free HADC structures with effective strain management. Residual strain analysis of the used SOI wafer. a Optical microscope image of the suspended beams. b Laser confocal microscope (LCM) image of the suspended beams. c Measured and simulated buckling of a suspended beam The three-dimensional (3D) finite-difference time-domain (FDTD) and finite element method (FEM) are used to mechano-optically co-design the photonic and mechanical structures of the HADC systematically. The schematic diagram of the HADC is shown in Fig. 3a, b and the parameters are chosen preliminarily as listed in Table 2 to achieve high-performance adiabatic coupling. Here, the total length of the coupling region is chosen as 60 µm, i.e., Lc = La + 2Lt = 60 µm, where Lt and La are respectively the lengths of the taper and the slab anchors. Initially, the lengths Lt and La are optimized without including the extra S-bends for achieving sufficiently low excess losses and avoid notable waveguide-buckling. Figure 3c gives the simulated excess losses of the HADC designed with Lt = 25, 20, 15, 10, and 5 µm, respectively. Definitely, in order to minimize the scattering loss due to the introduction of the slab anchors, the length Lt should be sufficiently long. For example, here we choose Lt ≥ 15 µm and obtain low scattering losses of 0–0.07 dB over the broad bandwidth of 1.4–1.7 µm. On the other hand, larger Lt results in a shorter anchor length La, which reduces the stability of the waveguides and might lead to stronger buckling both in the y and z directions. Note that the buckling of the stationary waveguide is stronger than the movable waveguide because the former is directly doubly-clamped by a pair of mode converters, while the latter has many other waveguide structures, including S-bends and meandering waveguides, to help release the residual strain. Consequently, only the buckling of the stationary waveguide is analyzed, as shown in Fig. 3d, e. The residual strain of the used SOI wafer is estimated to be 5 × 10–4 according to our measurement described above. Evidently, the waveguide buckling increases with decreased La. Therefore, the length La is chosen to be the maximum of 30 µm for weakening the buckling (dy ≤ 97 nm in the y direction, and dz ≤ 560 nm in the z direction) regarding that the condition of Lt ≥ 15 µm for achieving low scattering losses. Nevertheless, the waveguide buckling offsets dy and dz are still too large for the HADC to work. Design of the device. a Schematic diagram of the HADC. b Cross-section view through the A–A’ line in a. c Transmission spectra of the HADC in OFF state with different La. d, e Buckling of the stationary waveguide in the y and z directions with different La. f Buckling of the stationary waveguide in the y and z directions with extra S-bend. g, h Optimized alignment of the HADC in the ON state enabled by the minimized buckling. i, j Significant misalignment of the HADC in the ON state without the residual strain management structures, including the slab anchors and extra S-bends. The buckling is magnified by a factor of 2 for clarity. k, l Transmission spectra of the HADC in OFF and ON state respectively. m, n Light propagation of the HADC in OFF and ON state respectively. o Simulated displacement and maximum principal strain of the device in ON state. Residual strain of the SOI wafer has been considered. p Simulated electrostatic force arisen by different voltage and corresponding displacement of the movable mechanical structures Table 2Parameters of the designed HADCParametersValuesPreliminary parametersw1450 nmw2350 nmLs17.5 μmWs1 μmLc60 μmwt1 μmt170 nmt2150 nmgON25 nmOptimized parametersLa30 μmLt15 μm, (Lc – La)/2gOFF900 nm Parameters of the designed HADC Therefore, extra S-bends are also introduced at the both ends of the stationary waveguide to further decrease the buckling. Theoretically, a larger S-bends enables lower bending loss and smaller waveguide buckling, but introduce larger footprints and higher propagation losses. Therefore, the extra S-bends are designed to achieve low excess loss and compact footprints (see Supplementary Section 2). As a result, the waveguide buckling is minimized to be <42 nm as shown in Fig. 3f. Thus, residual strain management structures have been implemented with the mechano-optically co-design, so that the misalignment between the stationary and movable waveguides is effectively minimized in the ON state, as shown in Fig. 3g, h. Here, a direct comparison is given by simulating a conventional MEMS switch without any mechano-optically co-designed structure consisting of the slab anchors and the extra S-bends, as shown in Fig. 3i, j, indicating that there occurs a significant buckling offset of 1.3 μm, which seriously prohibits the desired coupling between the waveguides of the HADC in the ON state. To ensure low crosstalk in the OFF state and acceptably low driving voltages for switching on, the initial width gOFF of the gap between the two waveguides in the OFF state is designed to be 900 nm. The simulated transmission spectra and light propagation at 1550 nm of the HADC in the OFF state are shown in Fig. 3k, m, which shows low excess losses of 0–0.07 dB and low crosstalk of –52.5 to –28.5 dB over the broad bandwidth of 1400–1700 nm. Higher crosstalk at longer wavelengths is attributed to the weaker optical mode confinement and thus stronger evanescent field coupling. In the ON state, the width gON of the gap between the two waveguides is designed to be 25 nm to ensure adiabatic mode evolution (see Supplementary Section 3) within a short coupling region over a broad bandwidth. The simulated transmission spectra and light propagation at 1550 nm in the ON state are shown in Fig. 3l, n, showing low excess losses of 0–0.06 dB and low crosstalk of –47.7 to –21.1 dB over a large bandwidth. Furthermore, the entire MEMS structure in the ON state is simulated numerically as shown in Fig. 3o, which indicates ~2.5 × 10–3 maximum principal strain at the structural aligners, well below the damage threshold of silicon. The actuation voltage of the device is simulated as shown in Fig. 3p. Here, the red line shows the quadratic relationship between the driving force generated by the electrostatic combs and the applied voltage, while the blue line shows the linear relationship between the driving force and the displacement. The elastic deformation of the mechanical structures follows Hooke’s law, with a spring constant of ~0.36 N/m. According to the electromechanical simulation results, the threshold voltage for switching on is ~24 V for the designed switch with an initial 900 nm gap. Furthermore, the energy consumption of the elementary switch is estimated to be ~0.64 pJ (see Supplementary Section 9). Experiment results The designed devices were fabricated by electron beam lithography (EBL) processes. The perspective scanning electron microscope (SEM) image of the elementary switch is shown in Fig. 4a, showing that the footprint of the device is approximately 100×100 μm2. The stationary waveguide, the movable waveguide and the S-bends are all suspended as shown in the stitched close-up view in Fig. 4b. The width of the taper waveguides and the gap between the waveguides are measured to be w1 = 440 nm, w2 = 340 nm, and gOFF = 905 nm, respectively, as shown in Fig. 4c. The travel distance of the movable waveguide determined by the mechanical stopper is measured to be 880 nm as shown in Fig. 4d, indicating a 25 nm-wide gap when the switch is ON. The gentle discrepancy between the designed and fabricated dimensions is attributed to the imperfect fabrication. The measured transmission spectra T12 with different driving voltage are given in Fig. 4e, showing that there is very low crosstalk in the OFF state (0 V). As the driving voltage increases, T12 increases as well because the gap between the two waveguides is narrowed by the MEMS actuator. The threshold voltage setting the switch to be ON was measured to be 22 V, which agrees well with the simulated result of 24 V. The gentle discrepancy is attributed to the lowered spring constant caused by fabrication errors for the mechanical structures. The switch remains in the ON state without any damage even operating with a driving voltage up to 65 V, indicating a maximum driving voltage rating beyond 65 V. Such a large tolerance warrants robust drive schemes for large-scale switch arrays. For precise measurement of the excess loss of the switch, ten elementary switches in cascade were measured (see Supplementary Section 5). Figure 4f gives the averaged transmission spectra of the switch in OFF state, showing an excess loss of 0.11–0.43 dB and crosstalk of –61.8 to –42 dB over the large bandwidth of 1420–1600 nm. Note that the measured bandwidth is limited by the grating couplers. The excess loss in the OFF state is not as low as the simulated one due to the relatively high propagation loss caused by rough-sidewall scattering, excess mode conversion losses, and extra scattering losses of the slab-anchors due to slight EBL alignment offset (see Supplementary Section 4, 5). In the ON state, the measured transmission spectra are given in Fig. 4g, which shows that the excess loss is 0.05–0.99 dB and the crosstalk is –24.9 to –10.3 dB over the same large bandwidth. It can be seen that the experimental results are not as excellent as the simulated ones. By using a laser confocal microscope (LCM), the maximum buckling in the z direction of the stationary waveguide in the OFF state from 10 devices is measured to be ~100 nm consistently (see the red curve and shaded area in Fig. 4h), which is much larger than the simulated results due to the waveguide narrowing and the slab-anchor over-etching. Note that the simulated results for the buckling of the stationary waveguide with the parameters of w1 = 440 nm, w2 = 340 nm and t1 = 40 nm agree well with the experimental results, as shown by the blue curve in Fig. 4h, while the designed parameters are w1 = 450 nm, w2 = 350 nm and t1 = 70 nm. Such fabrication variations, together with the larger-than-expected buckling and significant propagation loss, contribute to the ON-state performance degradation of the switch (see Supplementary Section 6). Also note that the consistent buckling profiles from all the measured devices indicate excellent uniformity of the fabricated devices. The temporal experimental results show that the present switch works well with fast ON/OFF switching time of 2.2/1.9 μs, as shown in Fig. 4i. To demonstrate the durability, the switch was operated over 7.2 billion on-off cycles with a square-wave voltage at the frequency of 50 kHz for more than 40 hours. Negligible performance degradations were observed from the measured transmission at 1550 nm shown in Fig. 4j and the measured transmission spectra T12 given in Supplementary Section 7. Experimental results of the elementary switch. a Perspective SEM image of the elementary switch. b Stitched close-up view of the HADC. c, d Close-up view of the taper waveguides and the mechanical stopper. e Transmission T12 with different actuation voltages. The threshold voltage is measured to be 22 V and the maximum driving voltage rating is beyond 65 V. f, g Measured transmission spectra of the switch in OFF and ON state respectively. h The red shaded area shows the measured buckling in the z direction of the stationary waveguide in the OFF state from 10 devices. The red curve shows the averaged buckling. The blue curve shows the simulated results with the fabrication errors including the narrowing of waveguides and over-etch of the slab anchors. i Temporal response of the switch, which demonstrates ON/OFF switching times of 2.2 μs and 1.9 μs, respectively. j Durability measurement of the switch. The switch is driven by a square voltage with 50 kHz frequency for over 40 hours and the total operating cycles exceeds 7.2 billion, after which the switch shows no performance degradation The present 2 × 2 HADC switch is then scaled to a 64 × 64 array designed with the Benes topology, which consists of 352 switch cells and has a footprint of 8.4 × 5.2 mm2, as shown in Fig. 5a. Here varied-width multimode interference (VWMMI) ridge waveguide crossings are employed to achieve low loss and low crosstalk33, while grating couplers are used for fiber-chip coupling with two 66-channel fiber arrays at the input/output ports. The switch array in the initial all-OFF state (i.e., all 352 switch cells are OFF, labelled as ‘A-0’) was characterized by coupling light into the i-th (i = 1–64) input port in sequence, and the measured transmission Ti,jA−0 from the i-th input port to the j-th output port are shown in Fig. 5b, where i, j = 1–64. The maximum transmission Ti,jA−0 with i ≠ j is 36.7 dB lower than the transmission Ti,jA−0 with i = j over the bandwidth of 1530–1590 nm, exhibiting high extinction ratios. The excess losses of all the transmission Ti,iA−0 at the center wavelength of 1550 nm are given by the red circles in Fig. 5c, showing a variation ranging from 4.06 dB to 9.73 dB. Notably, among the 64 signal paths, path Ti,jA−0 (i = 40) exhibits a significantly higher loss, as indicated by the grey cross in Fig. 5c. This is attributed to fabrication defects occasionally introduced during the EBL process. To prevent confusion between the defective signal path and non-signal paths, the transmission spectrum of this particular path has been excluded from Fig. 5b. Here, since the chip is not packaged yet, each 2×2 switch cell can electrically be addressed by a pair of probes manually in the experiment, in which case we go further characterize the 32 specific states with only one cell switched on (labelled as ‘S-1’, i.e., all switch cells are still OFF except the selected one,). In the experiment, the 32 switch cells at the center stage (stage 5) of the Benes array were switched on one by one. As a result, the signal incident from the i-th input port is alternatively routed to the j-th output port, where j = i + 32 when i = 1–32 or j = i – 32 when i = 33–64), and the corresponding measured transmission spectra Ti,jS−1 are shown in Fig. 5d. As it can be seen, the extinction ratios (given as Ti,jS−1−Ti,jA−0) are higher than 38.5 dB over the bandwidth of 1530–1590 nm, while the excess losses of these 64 transmissions Ti,jS−1 are 5.31–8.75 dB at 1550 nm (see the red rings in Fig. 5e). The defective path i = 40 is marked in Fig. 5e and excluded in Fig. 5d for the same reason mentioned above. On the other hand, according to the measured excess/propagation losses of ridge waveguide crossings, 90° arc-bends, Euler S-bends and 600 nm-wide ridge waveguides (see Supplementary Section 4), the total excess losses of the signal paths in the all-OFF and single-ON states are respectively estimated as 3.92–8.93 dB and 5.65–8.15 dB (see Supplementary Section 8), which agrees well with the directly measured results, as shown by the blue rings in Fig. 5c and e, indicating that all the elements in the array work well with high uniformity, while the slight discrepancy is attributed to some slight fabrication defects, minor variations of grating couplers, inter-channel variation of the fiber arrays as well as some random particulate contamination on top of the chip. Experimental results of the switch array. a Optical microscope image of the switch array. b Transmission spectra Ti,jA−0 (i, j = 1–64) of the switch array in all-OFF state. c Measured and estimated excess loss of the 64 signal paths in all-OFF state at 1550 nm. d Transmission spectra Ti,jS−1 of the switch array in single-ON state. e Measured and estimated excess loss of the 64 signal paths in single-ON state at 1550 nm Discussions Residual strain of the SOI wafer For silicon photonic MEMS device design, residual strain management is crucial for doubly-clamped structures, and thus it is demanded to co-design optical and mechanical structures optimally. The residual strain of the SOI wafer is estimated to be 5 × 10–4 by measuring the buckling of suspended doubly-clamped beams and fitting the measured data with the FEM simulation result. Particularly, we introduce extra waveguide S-bends and slab anchors to reduce the buckling of the HADC waveguides to be less than 40 nm in theory, thus alleviating the negative impact on the performance of the switch. Such a mechano-optical co-design approach is beneficial to the development of most photonic MEMS devices and is applicable to wafers with a wide range of residual strain (see Supplementary Section 10). For wafers with residual strain larger than 1 × 10–3, employing a larger S-bend can help achieve a similarly low buckling. In experiment, the buckling of the fabricated HADC was alleviated to be ~100 nm, while the corresponding excess loss and crosstalk of the switch in the ON state are 0.05–0.99 dB and –24.9 to –10.3 dB, respectively, which is not as excellent as the design. The reason is mainly due to the fabrication variations, including the ~10 nm waveguide narrowing and the ~30 nm slab-anchor over-etching, which both directly degrade the optical performance and makes the HADC more sensitive to the release of residual strain. The performance of the switch can be improved by adopting a standard commercial silicon photonic foundry process, which offers more precise control over etching depth, and by appropriately compensating the waveguide width variation on the layout. Crosstalk in the switch array Our presented HADC switch exhibits non-negligible crosstalk only in the ON state, while maintaining low crosstalk in the OFF state. In a switch array, when two signal paths intersect at an ON (OFF) state switch, the inter-path crosstalk is comparable to the crosstalk of this ON (OFF) state switch. This means that in a Benes topology, where the ON-state intersecting scenario may occur, some paths may experience non-negligible crosstalk. Fortunately, in most commonly used strictly non-blocking topologies such as Crossbar and Dilated Banyan, any two signal paths either intersect at an OFF-state switch or do not intersect at any switch at all. Consequently, switch array in these topologies can achieve low crosstalk when our proposed HADC switch is applied. Scalability In this paper, the switches were fabricated with manual EBL processes in order to make rapid experimental verification. Unfortunately, the propagation loss of the fabricated silicon photonic waveguides is 8.3 dB/cm due to the rough sidewalls’ scattering. When using standard silicon photonic foundry processes, the propagation loss of 600-nm-wide ridge waveguides can be as low as 1.0 dB/cm or less, which is nine times lower than the waveguide fabricated with EBL processes in the lab. This will lead to a significant reduction in the excess loss of our HADC switch (see Supplementary Section 5). Meanwhile, as demonstrated in our previous work, the excess loss of the waveguide crossings can potentially be reduced to <10 mdB when fabricated with silicon photonic foundry processes34. Accordingly, the total excess loss could be reduced greatly to 4.5 dB on average for the 64×64 array. When scaled to be 128×128, the array has an average excess loss of 6.9 dB, which is still acceptably low, showing the great potential for high scalability. As a summary, the present HADC switch provides a highly viable solution for realizing large-scale silicon photonic switches available for a broad spectrum of practical applications such as photonic interconnect, programmable and reconfigurable PICs, photonic FPGAs, ONNs/PNNs, quantum information processors, and FPSA LiDAR. Supplementary information The revised Supplementary Information of the article Acknowledgements The authors thank the ZJU Micro-Nano Fabrication Center and the Westlake Center for Micro/Nano Fabrication and Instrumentation for the facility support. Author contributions Y. H., Q. M., L. Lyu, H. L., and D. D. designed the structures. Y. H., J. Z., Y. Lu, and Y. Liu fabricated the devices. Y. H. and J. Z. characterized the devices. Y. H., J. Z., H. L., and D. D. contributed to the data analyses. Y. H., J. Z., H. L., and D. D. wrote the manuscript. All authors discussed the results and contributed to the manuscript. H. L. and D. D. conceived the project. Funding This work is funded by National Key Research and Development Program of China (2024YFB2908302), National Science Fund for Distinguished Young Scholars (61725503), National Natural Science Foundation of China (U23B2047, 62321166651, 92150302, 62341508, 62375240), Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China (No. JYB2025XDXM106), Leading Innovative and Entrepreneur Team Introduction Program of Zhejiang (2021R01001), Zhejiang Provincial Major Research and Development Program (2021C01199), Natural Science Foundation of Zhejiang Province (LZ22F050006), Fundamental Research Funds for the Central Universities (226202400171), and Startup Foundation for Hundred-Talent Program of Zhejiang University. Data availability The data and materials that support the plots within this paper are available from the corresponding authors upon request. Competing interests H.L., Y.H., and D.D. have filed a patent application (application number: CN115826143A) on the silicon photonic MEMS switches based on adiabatic directional couplers presented in this paper, which is also in progress of PCT patent application (application number: PCT/CN2023/118315). All other authors declare they have no competing interests. Footnotes These authors contributed equally: Yinpeng Hu, Jiayue Zhu. Contributor Information Huan Li, Email: lihuan20@zju.edu.cn. Daoxin Dai, Email: dxdai@zju.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41378-026-01298-x. References Associated Data Supplementary Materials The revised Supplementary Information of the article Data Availability Statement The data and materials that support the plots within this paper are available from the corresponding authors upon request.