Modern data centers leverage optical interconnects to provide high-speed communication connectivity among numerous servers [1]. To address the increasing high-bandwidth demands for optical interconnects, optical switch matrices become the essential components, enabling the dynamic routing across multiple optical channels [2]. As data traffic grows exponentially in the widespread applications of cloud computing and artificial intelligence, there is an increasing demand for large-scale optical switch matrices [3]. Thanks to the high integration density and compatibility with established complementary metal-oxide-semiconductor (CMOS) fabrication processes, silicon optical switch matrices have made significant advancements in recent years, and different structures have been proposed [4]. In silicon photonic switch matrices, Mach-Zehnder interferometers (MZIs) [5–7] or micro-resonators [8–10] are widely used as effective basic switching elements (SEs). Usually, the MZI-based switch matrices support a broadband optical operation, but a large footprint and high power consumption are suffered. In contrast, the switch matrices based on micro-resonators, such as micro-ring resonators (MRRs) and micro-disk resonators (MDRs), feature compact footprint, low power consumption, and inherent wavelength selectivity, which is well-suited for wavelength division multiplexing (WDM) applications.

Recent studies have extensively explored silicon photonic switch matrices utilizing MRRs [11–13]. Traditional implementations often employ a crossbar architecture, characterized by a two-dimensional mesh structure [14,15]. In each mesh cell, a single MRR configured in an add-drop arrangement typically serves as the SE at the waveguide crossing. Optical routing is achieved by toggling the MRR on or off. For example, when the MRR is off, forward routing occurs; when it is on, either upward or downward routing is enabled, as dictated by the design. Once the design is established, the routing configuration remains fixed. Cascading multiple mesh cells creates the switch matrix, allowing specific optical routing through the activation of the designated MRRs, thereby facilitating a non-blocking feature. However, the reliance on a single MRR per mesh cell limits routing availability to two directions, consequently constraining the input/output (I/O) port density and scalability of the switch matrix. To enhance routing flexibility and increase port density, the placement of two identical MRRs diagonally at each waveguide crossing has been proposed [16]. One MRR facilitates upward routing, while the other supports downward routing. This collaborative use of dual MRRs enables multiple routing options and enhances scalability. Nonetheless, a significant drawback of this configuration is its inherent internal blocking, which may lead to routing conflicts. To address this, a space-and-wavelength-selective 8×4×2λ crossbar switch matrix is implemented [17]. Each mesh cell integrates two second-order MRRs as SEs in place of a single MRR, enabling individual wavelength selectivity. By cascading such mesh cells, a rearrangeable non-blocking switch matrix with both space and wavelength selectivity is formed.

As switch density increases in scalable switch matrices, thermal crosstalk emerges as a significant challenge, particularly due to the extensive use of top-placed micro-heaters. Thermal leakage from adjacent micro-heaters causes unwanted resonance wavelength drift in the MRRs, which can severely impair the overall performance of the switch matrix [18]. To mitigate this wavelength drift, active calibration and stabilization techniques have been proposed [19–22]. However, these approaches are often associated with the need for additional electrical control circuits [23].

In this paper, we propose a scalable space-and-wavelength selective optical switch matrix based on ultra-compact, thermally tunable MDRs. To support two wavelength channels, dual MDRs are strategically placed at each waveguide crossing to enable flexible routing in multiple directions, ensuring rearrangeable non-blocking connectivity and enhancing routing capacity while increasing I/O port density. To mitigate thermal crosstalk between adjacent MDRs, specifically engineered routing waveguides are integrated into the matrix. A proof-of-concept silicon photonic 1×8×2λ switch chip is fabricated and evaluated. Using the fabricated chip, high-speed optical data transmission is experimentally demonstrated. The proposed optical switch matrix offers significant advantages, including scalability and reduced thermal crosstalk, underscoring its potential for future high-speed optical interconnection networks and communication systems.

Figure 1(a) illustrates the schematic of the proposed scalable switch matrix implemented on a silicon photonic chip, featuring a two-dimensional mesh network with multiple input and output ports. The switch matrix utilizes ultra-compact thermally tunable MDRs configured in an add-drop arrangement as the SEs, and double-etched symmetric elliptical waveguide crossings at each waveguide intersection to facilitate horizontal and vertical routing. Figure 1(b) provides a zoom-in view of a single mesh cell, where dual MDRs are strategically positioned to enable routing in multiple directions. Specifically, when one MDR is on resonance, the transmitted optical signal is directed to its drop port, along the direction opposite to that of the input signal; when off resonance, the signal proceeds in the same direction as the input. By collaboratively activating or deactivating the two MDRs, the input optical signal can be routed to three directions, thereby enhancing routing capacity and increasing I/O port density compared to prior work [17]. Benefiting from the two MDRs, two wavelength channels are enabled. To address thermal crosstalk between the two MDRs in the mesh cell, engineered routing waveguides are introduced to increase the spacing between the dual MDRs without enlarging the mesh cell size. Additionally, two deep trenches are incorporated in the mesh cell to provide thermal isolation between neighboring cells. The single mesh cell has a side length of 510 μm. Figure 1(c) depicts the cross-sectional view of the MDR along the red dashed line in Fig. 1(b). The disk radius is designed to be 10 μm to achieve a large free spectral range (FSR). The bus rib waveguides are designed with a width of 0.6 μm to satisfy the phase-matching condition required for exciting the first-order whispering gallery mode (WGM) in the disk. Transition tapers are used to facilitate mode conversion between the bus rib waveguide and the 0.5-μm-wide wire routing waveguide. The coupling gaps between the disk and the bus rib waveguide are 200 nm. Different from the conventional MDR, an additional slab waveguide is incorporated to wrap the lateral sides of both the disk and the bus waveguide to strengthen the optical coupling between the disk and the bus waveguides. To enable wavelength tuning, two independent TiN micro-heaters are integrated atop the disks, allowing for collaborative control of the two MDRs and providing three-directional routing capabilities. With multiple cells incorporated into the switch matrix, the optical routing paths can be reconfigured by controlling the MDR switches, enabling rearrangeable non-blocking connectivity. The three-directional routing capability of each individual cell significantly increases the I/O port density of the switch matrix, which enhances its scalability.

As switch density increases in a scalable switch matrix, thermal crosstalk between the SEs becomes a critical challenge. Theoretically, an MDR can be toggled on or off by tuning its resonance wavelength via the thermo-optic effect.

Typically, the heat generated by a top-placed micro-heater induces the desired refractive index change in the waveguide to adjust the resonance wavelength. However, heat conduction can also cause unwanted temperature fluctuations, leading to refractive index changes in neighboring MDRs. Consequently, when the resonance wavelength of one MDR is tuned, neighboring MDRs experience resonance wavelength drift, a phenomenon known as thermal crosstalk. An effective strategy to minimize thermal crosstalk between MDRs is to introduce sufficient physical spacing between them. To illustrate this, a simulation is conducted. Figures 2(a) and 2(b) give a part of the mesh cell, highlighting the core area of the temperature distributions induced by the thermo-optic effect. In the proposed switch matrix, the routing waveguides are specifically engineered to provide a physical spacing of 250 μm between the two MDRs without enlarging the mesh cell size. In the simulation, in Fig. 2(a) when an electrical power of 25 mW is applied to the heater atop MDR1, the generated heat radiates outward, raising the center temperature from 300 K to 379 K. Notably, a significant temperature increase near the left side of MDR2 is observed. Figure 2(c) presents the simulated transmission spectra at the drop ports of the two MDRs, as shown in Fig. 2(a). The blue line represents the transmission spectra of both MDRs in their static state, where their spectra completely overlap due to their identical design. As the temperature increases, the green line shows the transmission spectrum of MDR1, with a desired wavelength shift of 4.95 nm, while the red line illustrates the transmission spectrum of MDR2, showing an undesired wavelength shift of 0.65 nm, indicating significant thermal crosstalk between the two MDRs. In contrast, when a physical spacing is introduced between the two MDRs, as shown in Fig. 2(b), the generated heat has a negligible impact on MDR2. Figure 2(d) presents the simulated transmission spectra at the drop ports of both MDRs depicted in Fig. 2(b). The blue line illustrates the transmission spectra of the two MDRs in their static state, where the spectra completely overlap owing to their identical design. As the temperature increases, the green line shows the transmission spectrum of MDR1, which exhibits a desired wavelength shift of 5.14 nm, while the red dashed line indicates that MDR2 experiences no wavelength shift. This indicates that thermal crosstalk can be effectively mitigated by carefully engineering the routing waveguide in the constrained physical space.

The proposed switch matrix utilizes dual ultra-compact MDRs to facilitate routing in multiple directions, significantly enhancing scalability in terms of I/O density and the number of SEs. By engineering the routing waveguides, sufficient physical spacing is introduced between the dual MDRs without altering the mesh cell size, thereby effectively reducing thermal crosstalk.

As a proof of concept, a silicon photonic 1×8×2λ switch chip is designed, as shown in Fig. 3(a). This design employs eight thermally tuned MDRs to facilitate the switching operation across eight channels. An edge coupler array with 10 ports serves as the I/O interface. When an optical signal is introduced at the input port labeled as I0, it can be directed to any of the output ports using the eight MDRs.

The proposed 1×8×2λ switch matrix chip is fabricated using the standard silicon-on-insulator (SOI) process provided by the CUMEC in China, featuring a 220-nm-thick top silicon layer and a 2-μm-thick buried oxide (BOX) layer. Figure 3(b) displays a microscope view of the fabricated 1×8 switch matrix chip, which has a footprint of 1.8mm×1.5mm. An edge coupler array, with a center-to-center spacing of 127 μm, is positioned at the bottom of the chip for I/O optical coupling. Nine electrode pads are located on top of the chip for electrical coupling, sharing a common ground pad. The insets illustrate zoom-in views of the MDR, highlighting the micro-heater situated atop the disk and the waveguide crossings for horizontal and vertical routing. Optical and electrical packaging is also conducted, as depicted in Fig. 3(c). The chip is initially die-bonded onto a metal heat sink, with both the chip and heat sink polished to achieve a smooth surface. This preparation enables the alignment and attachment of a 10-channel, 127-μm-pitch fiber array to the edge couplers using a UV-curable epoxy. Following optical packaging, the switch matrix chip is wire-bonded to a custom-designed printed circuit board (PCB) using nine gold wires for electrical fan-out. Finally, the packaged device is mounted onto an aluminum assembly. The inset provides a zoomed-in view of the photonic chip area, which includes the fabricated switch chip, the gold wires, and the fiber array.

The performance of the switch matrix chip is initially evaluated using an optical vector analyzer (OVA). Figure 4(a) presents the measured transmission spectra at all output ports in the static state. The red curve indicates the transmission spectrum at the output port O0, exhibiting a periodic notch response with an FSR of 10.5 nm and an extinction ratio (ER) of 32.1 dB. The other curves correspond to the drop responses of all the eight MDRs. As can be seen, the spectra exhibit clear FSRs and well-aligned resonance wavelengths. This clear spectrum reflects the excellent fabrication uniformity among the eight MDRs, which is attributed to the reduced random sidewall roughness due to the single waveguide sidewall of the disk structure, as well as the enhanced fabrication robustness provided by the additional slab waveguide incorporated in the MDR. The measured spectra at the output port O2 shows the drop response of a single MDR, featuring a 3 dB optical bandwidth of 164 GHz and an ER of 19.4 dB. The specialized MDR waveguide structure ensures its single mode operation. Using a top-placed micro-heater, each MDR can be thermally tuned. Figure 4(b) provides a zoom-in view of the wavelength tuning of the MDR at a resonance of 1540.7 nm. As the applied electrical power increases from 0 to 151 mW, the resonance wavelength is redshifted from 1540.7 nm to 1552.1 nm, achieving a shift of 11.4 nm, which corresponds to a tuning efficiency of 75.5 pm/mW. This wavelength tuning capability allows the MDR to function effectively as an optical switch element.

Switching speed is a critical metric for assessing the performance of the SE. An electrical square-wave signal with a frequency of 5 kHz is applied to the micro-heater atop the MDR, enabling the switch to toggle between on and off states. Consequently, the optical power at the drop port of the SE varies accordingly. An optical power meter captures the resulting signal, which is then analyzed using a real-time oscilloscope. Figure 5 displays the measured temporal profile of the generated electrical signal, with rise and fall times recorded at 14 μs and 16 μs, respectively. Thus, the switching time is determined to be 15 μs, which has the potential to be further accelerated through the utilization of the plasma dispersion effect.

On-off ER and channel crosstalk are another two critical metrics. To evaluate the on-off ER and channel crosstalk, the transmission spectra of eight channels are fully measured by applying electrical power to the specific MDR to switch the corresponding channel on or off. In the off-state measurement, no electrical power is applied to the chip. In the on-state measurement, electrical power is applied to a single MDR, tuning its resonance wavelength to the operation wavelength of 1547 nm. The selected operation wavelength of the on-sate MDR SEs is close to half of the FSR of the eight MDRs, which ensures minimal leakage power from the off-state MDRs and thereby maximizes the on-off ER. Figure 6(a) presents the on and off transmission responses of the channel from I0 to O2. As the MDR switch is tuned on and off, an ER as high as 20.0 dB is achieved. The on-off ER can be further enhanced by optimizing the MDR design to achieve near-critical coupling. Because of the limited on-off ER, residual optical power inevitably leaks into other channels, which is the channel crosstalk. Figure 6(b) shows the measured transmission spectra of all channels when the MDR in the channel from I0 to O2 is switched on. As can be seen, there is a channel crosstalk measured to be smaller than −22.7dB. Variations in channel crosstalk contribute to the slight differences in the off-responses of the other seven MDRs. Figure 6(c) summarizes the on-off ER and channel crosstalk of all eight channels. The red circle gives the measured optical power at each output port when the corresponding MDR is switched on; the green circle shows the measured optical power at each output port when the corresponding MDR is switched off, of which the difference is the on-off ER of each channel. The maximum on-off ER is measured to be 26.2 dB, and the minimum one is 20.0 dB. This variation arises from differences in the ER of each MDR, caused by fabrication imperfections. When the corresponding MDR in each channel is switched on, the output power of the other seven channels is also recorded, as shown by the gray circles. The channel crosstalk can be calculated in which the maximum is −16.7dB and the minimum is −22.7dB. Again, this difference is primarily attributed to the different ERs among the MDRs. Higher ERs generally provide better isolations between channels, thereby reducing crosstalk.

In addition, the insertion loss for each channel is also evaluated. The gray dashed line in Fig. 6(c) indicates an optical I/O coupling loss of about 15.0 dB from the pair of edge couplers. After accounting for coupling loss, the channel insertion loss is calculated to range from 2.0 to 5.8 dB, primarily due to MDR-related loss, waveguide propagation loss, waveguide crossing loss, and transition taper loss. Each waveguide crossing enables low-crosstalk horizontal and vertical optical propagation, whose insertion loss is as low as 0.1 dB and crosstalk is lower than −30dB. The transition tapers are employed to realize the mode transition between different types of waveguides, with each taper contributing a loss of approximately 0.2 dB. During the on-state measurements, the power consumption of each MDR ranges from 32.1 mW to 51.0 mW, with an average power consumption of 40.0 mW. The variation in power consumption is attributed to the differing lengths of the metal lines.

When a micro-heater is employed to tune the resonance wavelength of the MDR, thermal leakage from adjacent micro-heaters can induce unwanted resonance wavelength drift, known as thermal crosstalk. In large-scale optical switch matrices, thermal crosstalk presents a significant challenge, particularly in high-density integration of optical cavities. For comparison, a conventional mesh cell employing dual MDRs is evaluated. Figure 7(a) illustrates the resonance wavelength shifts of the two MDRs when an electrical power is applied solely to the micro-heater on top of MDR1, in which the inset shows the fabricated chip. As the electrical power is increased from 0 to 46.0 mW, the MDR1 exhibits a wavelength shift of up to 4.4 nm, while the MDR2 experiences a wavelength shift of 0.6 nm, attributed to thermal crosstalk. Figure 7(b) presents the measurement results for the proposed mesh cell, where the routing waveguide is specifically engineered to maintain a physical separation of 250 μm between the two MDRs, as depicted in the inset of the fabricated chip. As the electrical power is increased from 0 to 60 mW, the MDR1 exhibits a wavelength shift up to 4.4 nm, while the MDR2 shows a wavelength shift of 0.1 nm. Compared to the conventional mesh cell, when the MDR1 has a same wavelength shift, the MDR2 in the proposed mesh cell provides a largely reduced wavelength shift. This reduction verifies the effectiveness of the optimized routing design in mitigating thermal crosstalk.

To access the data transmission performance with a single wavelength channel, the fabricated switch matrix chip is incorporated in an optical transmission system, as shown in Fig. 8(a). A continuous-wave (CW) optical signal at a wavelength of 1547 nm is generated by a laser diode (LD) and subsequently sent to a Mach–Zehnder modulator (MZM), where 20 Gbps non-return-to-zero (NRZ) on-off keying (OOK) data is modulated onto the optical carrier. The data is generated by an arbitrary waveform generator (AWG) using a 210−1pseudo-random bit sequence (PRBS10). The modulated optical signal, with a power of −3.5dBm, is coupled into the fabricated switch chip. At the output of the chip, the optical signal is amplified by an erbium-doped fiber amplifier (EDFA) to provide a power as large as 0 dBm. Using a high-speed photodetector (PD), the modulated optical signal is down-converted to an electrical signal. The recovered electrical NRZ signal is then captured using a real-time oscilloscope (OSC). The eye diagrams are directly recorded from the OSC, without applying additional digital signal processing (DSP) or filtering. For comparison, a back-to-back (B2B) operation is also performed, in which a variable optical attenuator (VOA) is introduced to maintain the same insertion loss as the chip. During the measurement, electrical power is applied only to the MDR in the target channel, while the MDRs in non-target channels remained unpowered. Figure 8(b) shows the measured eye diagrams for each channel of the switch chip as well as for the B2B operation. The ERs of the eye diagrams are also calculated, whose values are 18.4 dB, 14.4 dB, 15.4 dB, 14.8 dB, 17.5 dB, 14.9 dB, 14.0 dB, 15.1 dB, and 18.9 dB. The differences result from the different insertion losses of each channel. Clear and open eyes can be seen, which verifies the effectiveness of the system for OOK signal transmission using the fabricated switch matrix. Noticeably, the time jitters can be observed in the measured eye diagrams, which are primarily attributed to the high insertion loss of the chip and the large relative intensity noise (RIN) from the laser source. The 20 Gbps data rate is constrained by the bandwidth limitations of MZM and PD used in the test setup. The fabricated switch matrix utilizes dual ultra-compact MDRs to facilitate flexible routings and a large transmission bandwidth.

To evaluate the data transmission performance with dual wavelength channels, the fabricated chip is integrated into an optical transmission setup. In this setup, two CW optical carriers at 1550.1 nm (λ1) and at 1548.5 nm (λ2) are generated by two LDs, with a wavelength spacing of 1.6 nm, consistent with dense WDM standards. Each optical carrier is sent to individual MZMs, where a 20 Gbps NRZ OOK data is modulated onto the optical signal at λ1 and a 15 Gbps NRZ OOK data is modulated onto the optical signal at λ2. The data signals are generated by an AWG using 210−1 PRBS10. The output powers of the modulated signals from both modulators are approximately 0 dBm. The two modulated signals are then combined using a fiber coupler (FC) and injected into the 1×8 switch chip. By controlling the MDRs, the optical signal at λ1 is routed to O1 and the optical signal at λ2 is routed to O4. At the output of the chip, the optical signals at each output port are amplified by EDFA to a power of 0 dBm. Then, each modulated signal is detected by a high-speed PD and converted into an electrical signal, which is then captured by a real-time OSC.

Figure 9(a) gives the measured transmission spectra of the two channels using an OVA. The routing channel from I0 to O1 exhibits a channel crosstalk of −20.0dB, while the channel from I0 to O4 has a crosstalk of −9.2dB. Figure 9(b) shows the optical spectra of the combined modulated optical signals before injecting into the chip measured by an optical spectrum analyzer (OSA). After amplification by an EDFA, the measured spectra at output ports O1 and O4 are presented in Figs. 9(c) and 9(d), showing a power suppression of 12.3 dB at output port O1 and 13.1 dB at output port O4. To assess the effects of wavelength channel crosstalk, eye diagrams are evaluated in three different conditions: B2B, single-wavelength-channel routing, and dual-wavelength-channel routing. The corresponding results are presented in Fig. 9(e). For the 20 Gbps data at 1550.1 nm, the ERs are 19.8 dB (B2B), 19.6 dB (single-wavelength-channel routing), and 16.8 dB (dual-wavelength-channel routing). For 15 Gbps data at 1548.5 nm, the ERs are 17.3 dB (B2B), 17.0 dB (single-wavelength-channel routing), and 16.4 dB (dual-wavelength-channel routing). As can be seen, despite minor degradation in the dual-channel condition, the clear and open eye patterns confirm the viability of dual wavelength data transmission using the fabricated switch.

Regarding scalability of the proposed switch, wavelength channel, propagation loss, and the size of the mesh cell are the key considerations. In our proposed N×2N×2λ switch, the number of supported wavelength channels depends on the number of MDRs in each mesh cell. However, due to a periodical optical response of the MDR, the FSR finally limits the maximum number of the wavelength channel allowed within a wavelength range. To accommodate more MDRs supporting more wavelength channels, a larger FSR is preferred, which is exactly the key advantage of the MDR. Thanks to the single sidewall configuration, the MDRs can have a small radius and a large FSR. For example, we have demonstrated an MDR with 3.7 μm radius, enabling 31.1 nm FSR [24]. With small-radius MDRs, a high wavelength scalability is enabled.

As the switch matrix scales up, on-chip optical propagation loss becomes a limiting factor for the number of mesh cells, with MDR insertion loss being the primary contributor. Based on the measured on-chip path loss of 2.0 to 5.8 dB in the 1×8 prototype, each MDR is estimated to contribute ∼2dB at the drop port and ∼0.5dB at the through port. For an N×2N×2λ switch, the minimum propagation loss happens in the shortest path where only one MDR is involved. The value is 2 dB. The maximum propagation loss happens in the longest path where 4N–1 MDRs are involved. The value is (1+2N)dB, including one drop loss and 4N–2 through loss. For instance, assuming our proposed switch has a size of 8×16×2λ, the maximum propagation loss would incur up to ∼17dB. The MDR insertion loss can be further reduced by optimizing the MDR structure or adopting low-loss silicon nitride platforms [25].

The physical space of the chip constrains the number of the mesh cells. In the fabricated 1×8×2λ switch, each mesh cell occupies a footprint of 510μm×510μm. The routing waveguide is specifically engineered to provide a 250 μm spacing between the MDRs for thermal crosstalk alleviation. To further minimize the mesh cell size, MDRs can be designed with a significantly smaller radius, such as 3.7 μm, resulting in reduced power consumption and minimized thermal diffusion. This reduction in thermal effects facilitates a decrease in the physical spacing between resonators, thereby enabling a more compact mesh cell architecture. These ultra-compact MDRs present strong potential for enhancing both scalability and switch density. Furthermore, the spacing can be reduced through improved thermal management, such as optimizing the placement of air trenches and integrating high-thermal-conductivity heat sinks to suppress lateral heat diffusion [26]. These design strategies collectively help balance the trade-off between footprint and thermal crosstalk, presenting a promising approach for scaling optical matrices within the physical chip constraints.

Although an MDR-based two-dimensional network configuration has been implemented in a photonic field-programmable disk array (FPDA) signal processor [27], the motivations are distinct. The FPDA signal processor focused on general-purpose photonic microwave signal processing, whereas the switch matrix targets high-bandwidth digital optical interconnects. To achieve routing of high-speed digital signals, specific design modifications are necessary, including the integration of engineered routing waveguides to mitigate thermal crosstalk in densely integrated arrays, as well as optimizing MDRs for broad bandwidth operation of 164 GHz to support high-speed data routing.

In conclusion, we proposed a scalable optical switch matrix that enables both space and wavelength selectivity by employing ultra-compact thermally tunable MDRs. To support two wavelength channels, dual MDRs were strategically positioned at each waveguide crossing to facilitate routing in multiple directions, enhancing I/O port density. To mitigate thermal crosstalk between adjacent MDRs, specially engineered routing waveguides were integrated into the matrix. A prototype silicon photonic 1×8×2λ switch chip was fabricated and successfully demonstrated optical data transmission. The proposed switch matrix offers a promising solution for scaling optical matrices with low thermal crosstalk in future optical interconnect networks.