Optical switches with multiple input/output ports perform optical signal routing in various optical networks^[1,2]. The silicon-on-insulator (SOI) platform shows great potential for integrated large-scale optical switches due to its complementary metal-oxide-semiconductor (CMOS) compatible fabrication and large refractive index contrast^[3–6]. Recently, some subwavelength structures have been introduced in integrated photonics to enhance the performance of the next generation of integrated photonic devices^[7–9]. Various silicon optical switches have been demonstrated based on Mach–Zehnder interferometers (MZIs)^[10–14], micro-ring resonators^[15,16], micro-electromechanical systems (MEMS)^[17–20], and so on. However, due to the large birefringence of the silicon waveguide with a typical height of 220 nm^[21,22], most silicon optical switches have shown large polarization-dependent loss (PDL)^[23,24], limiting their practical applications. For example, in optical fiber communication systems, the polarization state of optical signals varies randomly along the optical fibers^[25]. Consequently, the large PDL of silicon optical switches significantly deteriorates the performance of the entire optical network.

Therefore, it is highly demanded to design a large-scale polarization-insensitive silicon optical switch. Two methods have been widely adopted. One is to make all the optical elements insensitive to polarization by optimizing waveguide dimensions^[22,26–29]. However, the thickness of the waveguide must be larger than the commonly used 220 nm thickness^[27]. The other is to use the polarization diversity scheme^[24,30,31]. Input light is divided into two branches with orthogonal polarizations by a polarization beam splitter (PBS). Meanwhile, the polarization of one branch is rotated 90° by a polarization rotator (PR). These two branches pass two identical switch fabrics and finally are combined back into one path by a PR and a PBS^[32,33]. In the polarization diversity scheme, the optical elements are usually doubled, increasing the chip footprint and the control complexity. Recently, a non-duplicate polarization diversity silicon optical switch based on the path-independent insertion-loss (PILOSS) topology with extra waveguide crossings has been demonstrated^[31]. However, such a method is not commonly applied to most other switch topologies.

Previously, on-chip polarization controllers (PCs) have been proposed and demonstrated for polarization receivers^[34,35]. With active control of the phase shifters (PSs) in the PCs using a proper feedback algorithm, an arbitrary input polarization can be adjusted to the transverse electric (TE) mode. In this work, we experimentally demonstrate a polarization-insensitive $4\times 4$ silicon optical switch with on-chip PCs for automated input polarization adjustment. We develop a polarization control system to automatically adjust the polarization. Such polarization-insensitive optical switches are easy to scale up and can apply to any switching topology, greatly expanding the application scenarios of silicon optical switches.

The proposed $4\times 4$ polarization-insensitive optical switch is composed of four on-chip PCs followed by a $4\times 4$ dilated double-layer network (DLN) switch fabric based on MZI elements, as shown in Fig. 1(a). The input and output ports are marked as $I_m$ and $O_m$ ($m=1$, 2, 3, 4). Inverse tapers with a tip width of 120 nm are used for low fiber-to-chip coupling loss. The pitch is 50 µm. Light with an arbitrary polarization state is edge-coupled into the waveguide and automatically adjusted to the TE polarization by the on-chip PC. The $4\times 4$ MZI switch fabric is employed to establish routing paths for the TE polarized light.

The $4\times 4$ switch fabric includes eight $2\times 2$ MZIs and sixteen $1\times 2$ MZI switch elements. The switch element at the $i^{\mathrm{th}}$ column and $j^{\mathrm{th}}$ row is denoted as $E_{ij}$ ($i=1$, 2, 3, 4 and $j=1$, 2,…, 7, 8). Compared with our previous $4\times 4$ DLN optical switch chip^[11], we replace the $2\times 2$ MZIs in the second stage with sixteen $1\times 2$ MZIs to completely cancel the first-order crosstalk in the routing paths. Therefore, the switch fabric has lower crosstalk since no two optical paths share the same switch element. The 3 dB couplers in the MZI switches are based on $1\times 2$ and $2\times 2$ multimode interference (MMI) couplers. Unfortunately, due to the design error, there is a lateral shift of 20 nm at the joint of the MMI and the MZI arm for each $1\times 2$ MZI switch element. This increases the insertion loss and crosstalk. A thermal PS is integrated in one MZI arm of all switch elements for thermo-optic (TO) switching.

The detailed structure of the on-chip PC is illustrated in the inset of Fig. 1(a). It is constructed by a PR splitter (PRS)^[35], a $2\times 2$ MZI coupler, and a photodetector (PD) at the idle output port of the MZI. The other output port is connected to the $4\times 4$ switch fabric. The PRS is used to split the input light with TE and transverse magnetic (TM) polarizations into two TE branches. The insertion loss and the crosstalk of PRS are 0.35 dB (1.87 dB) and $-21\mathrm{dB}$ ($-13.4\mathrm{dB}$) for the TE (TM) light, respectively, according to our test device. Thermal PS1 is embedded in the upper branch to control the relative phase between the two branches. The two branches are then connected to an MZI coupler with thermal PS2 embedded in one MZI arm to control the optical power distribution in the two output ports. The on-chip PD is used to convert the optical wave to an electrical signal for feedback control.

The control system to perform automated polarization adjustment consists of transimpedance amplifiers (TIAs), analog-to-digital converters (ADCs), a micro-control unit (MCU, STM32F4), digital-to-analog converters (DACs), and drivers. The optical signal from the idle output port of the MZI coupler is converted to a photocurrent by the on-chip PD, and then it is amplified by the TIA and sampled by the 12 bit ADC before feedback to the MCU. In the MCU, a polarization control algorithm based on a gradient descent method is implemented to achieve real-time feedback control.

Figure 1(b) shows the microscope image of the fabricated silicon optical switch chip. The footprint of the chip is $5\mathrm{mm}\times 3\mathrm{mm}$. We designed a silica-based planar lightwave circuit (PLC) with 2.5% refractive index contrast ($\mathrm{\Delta }n$) for fiber-to-chip coupling. The PLC waveguide has $3\mathrm{\mu m}\times 3\mathrm{\mu m}$ cross-sectional dimensions. The waveguide pitch in the PLC is gradually transformed from 50 µm to 127 µm to accommodate the pitch of a commercial fiber array. The electrical and optical packaging was done in the SJTU-Pinghu Institute of Intelligent Optoelectronics. Figure 1(c) displays the packaged chip.

Using the transfer matrix method, we can derive the normalized optical power at the feedback port of the on-chip PC as a function of phase changes ($\mathrm{\Delta }{\phi }_1$ and $\mathrm{\Delta }{\phi }_2$) of the two PSs:$$P_f={\left|\frac{1}{2}[e^{-i\mathrm{\Delta }{\phi }_1}(e^{-i\mathrm{\Delta }{\phi }_2}-1)e^{-i{\phi }_{\mathrm{TE}}}{E}_{\mathrm{TE}}-i(e^{-i\mathrm{\Delta }{\phi }_2}+1)e^{-i{\phi }_{\mathrm{TM}}}{E}_{\mathrm{TM}}]\right|}^2,$$where $E_{\mathrm{TE}}$ and $E_{\mathrm{TM}}$ represent the TE and TM electric field components of the input light before PRS (normalized to the total power, i.e., $|E_{\mathrm{TE}}{|}^2+|E_{\mathrm{TM}}{|}^2=1$^[36]), respectively, ${\phi }_{\mathrm{TE}}$ and ${\phi }_{\mathrm{TM}}$ are the phase changes induced in the PRS. Figure 2 shows the normalized optical power $P_f$ varied with $\mathrm{\Delta }{\phi }_1$ and $\mathrm{\Delta }{\phi }_2$ for different proportions of the input light. The phase difference between the two branches is set to be $\pi /2$ as a typical case without losing generality. It can be seen that there are two global minimum points for an arbitrary polarization. We can find the best phase shift values by reaching one of the global minimum points using a control algorithm so that the output power to the switch fabric is maximized.

We used a polarization control algorithm based on a gradient descent method to search for the global minimum^[34]. Figure 3 shows the algorithm flow chart. $U_{\phi 1}$ and $U_{\phi 2}$ are the applied DAC voltages to PS1 and PS2, respectively. $U_f$, $U_{f1}$, and $U_{f2}$ are recorded voltages from the ADC at the feedback port. $g_1$ and $g_2$ represent the gradient of $U_{f1}$ and $U_{f2}$ updated in each iteration, which determines if the adjustment is in the proper direction. $\alpha$ represents the learning rate, related to the step size of the applied voltages on PS1 and PS2^[34]. It is important for the algorithm to select proper parameters, such as the initial values of $g_1,g_2,U_{\phi 1},U_{\phi 2},\alpha$, and the threshold voltage $U_{\mathrm{f\_min}}$. Proper initial values can reduce the number of iterations and speed up the adjustment process. Likewise, $U_{\mathrm{f\_min}}$ is used to terminate the control loop when the monitored signal is less than the threshold.

To characterize the polarization adjustment of our chip, we first adjusted the on-chip PC in the $I_2-O_2$ path using the polarization control algorithm described in Fig. 3. A fiber-based commercial PC was inserted between a tunable laser and the optical switch chip to adjust the input polarization state. A digital phosphor oscilloscope (DPO, Tektronix 5054B) was connected to the TIA to capture the output signal at the feedback port. Figures 4(a) and 4(b) show the recorded voltages at the feedback port during the polarization tuning process for two random polarization states. The polarization recovery time is less than 4 ms. We also continuously perturbed the polarization state of the input light by manually adjusting the PC. It can be seen that without automated adjustment, the voltage at the feedback port changes accordingly, as shown in Fig. 4(c), which means that the optical power at the output port varies with the polarization of input light. Figure 4(d) shows the recorded voltages upon active adjustment. The signal at the feedback port is kept at the minimum under a continuously changed input polarization. Therefore, with the on-chip PC, our switch chip can accommodate any dynamically changed input polarization.

Figure 5 shows the measured transmission spectra of the $I_2-O_2$ path after polarization adjustment. The spectra are normalized with a U-shaped reference waveguide. The on-chip insertion loss gradually decreases as the input light polarization rotates from the TE-like to the TM-like state. The PDL is ∼ 2 dB in the wavelength range of 1540 nm–1580 nm. The relatively large PDL is caused by the inverse taper at the input ports of the chip. We also notice that other than these two special polarization states, i.e., TE-like and TM-like states, the spectra show a periodic response with a free spectral range (FSR) of ∼6.6 nm. This is because there is an optical path difference between the two PRS branches in our design.

Figure 6 shows the measured transmission spectra of one switching state of the optical switch. The four routing paths are $I_1-O_1,I_2-O_2,I_3-O_4$, and $I_4-O_3$, as illustrated by the different colors in Fig. 1(a). Here, we adjusted the PC to maximize the output power while the on-chip PCs were turned off. The on-chip insertion loss of each path is 3.1 dB, 3.5 dB, 6.35 dB, and 6.4 dB at a 1550 nm wavelength, respectively. The worst crosstalk is −10 dB in this state, as shown in Fig. 6(c). The relatively larger insertion loss of paths $I_3-O_4$ and $I_4-O_3$ is due to the large insertion loss of switch elements $E_{27}$ and $E_{28}$, where a 20 nm lateral alignment error is incurred at the input of the elements in the mask layout design. All sixteen $1\times 2$ MZIs have lateral alignment errors in one of the MZI arms. This slightly increases the insertion loss, but significantly degrades the crosstalk. The insertion loss of path $I_4-O_3$ is broken down into ∼0.35 dB from the PRS, ∼1.2 dB from three $2\times 2$ MZIs, ∼3.4 dB from one $1\times 2\mathrm{MZI}(E_{28})$, ∼0.6 dB from the other $1\times 2$ MZI ($E_{37}$), and ∼0.85 dB from the crossings and connecting waveguides. Therefore, by correcting the design errors, the insertion loss can be improved to ∼3.2 dB. The crosstalk can also be reduced to ∼−50 dB, as no two optical paths share the same switch element. In Figs. 6(a) and 6(c), periodical ripples are observed in the measured spectra because the input light is not pure TE or TM polarization.

We performed a high-speed on–off keying (OOK) signal transmission experiment with various input polarization states for the $I_2-O_2$ path. A tunable laser source (TLS) generated a continuous wave (CW) light at the wavelength of 1550 nm. It was then modulated by a commercial modulator. A 32 Gbit/s pseudo-random binary sequence (PRBS) signal with a pattern length of $2^{31}-1$ generated from a pulse pattern generator (PPG, Keysight, N4951B) was used to drive the modulator. The modulated optical signal was amplified by an erbium-doped fiber amplifier (EDFA), and the polarization state of the input light was adjusted by a PC before it was edge-coupled to the switch chip. We used the polarization control system to actively adjust the polarization to the TE mode. The output signal of the switch chip was amplified by another EDFA and followed by an optical bandpass filter. Finally, it was detected by a 50 GHz photodiode (${\mathrm{u}}^2\mathrm{t}$, XPDV2120R) and received by a digital communication analyzer (DCA, Agilent DCA-X 86100D). We measured the eye diagrams for six different input polarization states, as shown in Fig. 7(a). No significant signal degradation was observed from the eye diagrams, indicating high signal integrity after passing the switch chip. These results indicate that the 4 × 4 optical switch chip can achieve polarization-insensitive optical signal routing.

To evaluate the optical signal transmission through our optical switch with different input polarization states, we also measured the error-free operation at a bit error rate (BER) of ${10}^{-9}$ for a 32 Gbit/s OOK signal. Figure 7(b) shows the relationship between the BER and the received optical power (ROP) with the input polarization gradually tuned from TE to TM ($\mathrm{POL}.1\to \mathrm{POL}.6$) mode together with the back-to-back (BTB) measurement as a reference. The ROP is changed by a VOA placed before the photodiode. The power penalty is about 1 dB for the worst polarization state at a BER of ${10}^{-9}$.

In conclusion, we have demonstrated a $4\times 4$ polarization-insensitive silicon optical switch based on a dilated DLN architecture. Four on-chip PCs are integrated before the switch fabric for real-time polarization adjustment. The polarization recovery time is less than 4 ms for any arbitrary input polarization state using the polarization control algorithm. The measured PDL is about 2 dB. The 32 Gbit/s OOK data transmission experiment shows that our optical switch can support high-speed data transmission with any input polarization. Our switch offers a new approach to overcome the polarization sensitivity issue to satisfy the practical applications in datacenter networks.