De-multiplexing free on-chip low-loss multimode switch enabling reconfigurable inter-mode and inter-path routing

1 Introduction

Optical interconnects with broad bandwidth, low power dissipation and multiplexing compatibility have become a disruptive technology with a variety of applications such as intra-chip interconnects, short-reach communications in high-performance computers and datacenters and long-haul communication networks [1], [2]. Benefitting from diverse multiplexing techniques, great strides have been achieved in optical interconnects for single-mode optical systems. Among these techniques, the wavelength-division multiplexing (WDM) involving numerous wavelengths in a shared channel has been one of the most successful in the past decades [3], [4], [5], [6], [7], [8], [9]. To satisfy the exponentially increasing data demand, an emerging technique, mode-division multiplexing (MDM), is becoming more and more attractive as multiple spatial modes share a single-wavelength link to further improve the communication capacity [10], [11]. It is of great interest to fulfill high-bandwidth-density interconnects by combining these two techniques to construct a hybrid multiplexing system [12].

On the other hand, reconfigurable photonic integrated devices, which allow dynamic data rearrangement and increase flexibility as well as throughput, are the key components of optical networks. A compact and low-loss switching device is promising for data-hungry and power-consumption-sensitive applications such as optically interconnected processors and memories [13], [14]. Previously, many researches were focused on single-mode system, and various high-performance single-mode switch elements and fabrics with high-port count were achieved [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27]. Recently, switches for multimode signals were proposed and demonstrated [28], [29], [30], [31], [32], [33], [34]. While more ports and larger scale were developed, multiple dimensions were also introduced with reconfigurable and more powerful routing capability. In [28], Wu et al. realized a 1×3 switch processing three modes. Add-drop functionality was achieved in the 1×1 multimode switch for four-mode channels [29], [30]. The 1×2 multimode switches were demonstrated by combining WDM or polarization-division multiplexing with MDM [31], [32]. Compared with a single-input-port switch, a double-input-port case is more efficient since it can tackle two optical links. In [33], a 2×2 mode switch that routed four modes was demonstrated. The aforementioned switches are limited to path routing, and the signal is carried on the fixed mode. Recently, Yang et al. proposed a 2×2 switch. Reconfigurable global mode routing (not only path routing, but also mode routing among different spatial mode channels) was fulfilled to extend the freedom degree of reconfigurable routing [34]. The approach is de-multiplexing all the input multimode signals to fundamental modes and processing them individually by utilizing single-mode elements. Then the fundamental modes can be multiplexed back to the target output modes. Nevertheless, a large insertion loss (>10 dB), which is mainly attributed to the multiplexer (MUX)/demultiplexer (DEMUX) and complicated layout, was induced [34]. Such a demultiplexing-processing-multiplexing procedure results in a quite complicated configuration and puts forward higher requirements for the mode (DE)MUX performance. In contrast, previous works for one-stop multimode processing, which can be free from demultiplexing and multiplexing operations, have been realized and are promising for the 2×2 multimode switch [35], [36], [37].

Multiple 1×2 or 2×2 switching elements can be potentially cascaded with different network architectures to obtain higher-port-count switch fabrics and bandwidth-dense interconnects. The choice of switching elements determines the footprint, crosstalk, switching loss and the switch fabric architecture. In this paper, we present a compact and low loss 2×2 multimode switch that is easily scalable to high-port-count switch fabrics. It consists of a pair of 1×1 Mach-Zehnder interferometer (MZIs) multimode switches and a pair of micro-ring (MR) 2×2 single-mode switches. Reconfigurable inter-mode plus inter-path routing for fully flexible on-chip optical networks can be achieved. The input multiplexed signals can be processed simultaneously without a complicated demultiplexing-processing-multiplexing procedure. It is convenient to extend the design to support more mode channels by adopting a multi-arm MZI structure [38] and cascaded symmetric Y-junctions in a plane. The proposed switch is experimentally demonstrated with a low insertion loss of <1.2 dB and a low crosstalk of <−16.6 dB for all channels and switching configurations. We also demonstrate two representative types of 4×4 switch fabrics with cross-bar and ring-bus architectures by assembling multiple proposed switch elements. These two types of switch fabrics are optimized for different functionalities. The first realizes strictly non-blocking and almost path-independent insertion loss (PILOSS), while the latter one is non-blocking and can make full use of the optical paths. System characterization with 32 Gb/s high-speed non-return-to-zero on-off keying (NRZ-OOK) signals is carried out, reaching up to 256 Gb/s of aggregate throughout.

2 Switch design

Figure 1A illustrates the schematic of a high-port-count multimode switch fabric. Low loss waveguide crossings constitute the backbone of the optical network. To enable WDM grid compatibility, MRs are introduced to change the light propagation direction. Multiple WDM channels can be supported when the free-spectral range (FSR) is set as the wavelength spacing. Each input or output waveguide supports two modes. The numeric strings with red and green colors represent the signals carried on the fundamental transverse electric mode (TE₀) and first-order transverse electric mode (TE₁), respectively, as shown in Figure 1A. The input signals can be optionally converted to the desired output modes. Figure 1B shows the fundamental 2×2 multimode switch element, which is elaborately designed with a pair of MZI-based 1×1 multimode switches intersecting with each other and a pair of MR-based 2×2 single-mode switches. Two symmetric Y-junctions are employed to construct the MZI structure, and they act as power splitter and combiner for both TE₀ and TE₁ modes. For the splitter, the input TE₀ is equally split into two parts with zero phase difference, while TE₁ is split into two TE₀ modes with π phase difference. At the combiner, the two TE₀ modes are then combined back into TE₀ or TE₁ mode, depending on the phase difference [35]. A waveguide crossing is positioned close to a MR, forming a 2×2 single-mode switch. The MR heater H2 and MZI heaters H1 and H3 are utilized for reconfigurable control, performing two possible states: OFF and ON. In the OFF state, the resonance wavelengths of the two MRs deviate from the operating wavelength with half FSR to minimize the crosstalk, and thus no power couples into the MR and the signal is guided to the through port uninterruptedly. In the ON state, the light is coupled into the MRs as the resonance wavelengths are shifted to align with the operating wavelength, making a 90° direction change and outputting the light from the drop port. These two MRs always work with same switching states and resonance wavelengths. Note that the MR-based switch here is different from the traditional work in [24], where each MR serves as a 1×2 single-mode switch, and the resonance wavelengths of the MRs should be different to realize wavelength multiplexing or de-multiplexing. For the heaters H1 and H3, by controlling the phase difference between the two MZI arms, the data information carried on the two input modes can be exchanged or remain the same. In the OFF state, no phase difference is introduced; thus, the data are carried on the original modes. In the ON state, a π phase difference is introduced; thus, these two modes are exchanged with each other, as shown in Figure 1C. Therefore, the MZI-based 1×1 multimode switch functions equivalently as a structure consisting of a mode MUX, a DEMUX and a 2×2 single-mode switch, avoiding the performance degradation due to unnecessary (de)multiplexing.

Figure 1:

The proposed integrated multimode switch.

(A) Schematic of high-port-count multimode switch fabric. The numeric strings with red and green colors stand for the signals carried on different modes. The signal can be selectively guided to the desired output port as the desired mode. (B) A 2×2 multimode switch element consisting of two Mach-Zehnder interferometer (MZI) based 1×1 multimode switches orthogonally intersecting with each other and two micro-ring (MR) based 2×2 single-mode switches. The thermal heaters are used for reconfigurable control. (C) MZI-based 1×1 multimode switch. The TE₀ and TE₁ modes exchange with each other in the ON state. Schematic views of (D) the symmetric Y-junction and (E) waveguide crossing.

For clarity of presentation, we only choose the TE₀ mode injected from port I2 as an example to illustrate the reconfigurable inter-mode and inter-path routing. In this regard, the output results are only determined by the states of the heaters H2 and H3, which are independent of heater H1 that only impacts the light injected from port I1. The output results for all the permutations and combination states of H2 and H3 are shown in Figure 2. In the OFF states for both heaters H2 and H3, the input light propagates straight to port O2 with the output mode unchanged (Figure 2A). When the heater H3 is switched to the ON state, the input signal on TE₀ mode still travels to O2 while it transforms into the TE₁ mode at the output (Figure 2B). When the heaters H2 and H3 are in the ON and OFF states, respectively, the light turns to port O1 as the TE₀ mode (Figure 2C). For the ON states of both heaters H2 and H3, the light propagates to port O1 as the TE₁ mode (Figure 2D). In these scenarios, there are four different output results for the single input case. The MZI heater H3 is responsible for inter-mode routing, while the MR heater H2 determines the inter-path routing. The situation is similar when the TE₀ mode is injected from port I1. Therefore, the 2×2 multimode switch element has a complete set of eight routing states corresponding to three individual heaters, as summarized in Table 1. The signals 1 and 2 carried on the TE₀ and TE₁ modes are injected from I1, while signals 3 and 4 carried on the TE₀ and TE₁ mode are injected from I2. When the switch element is extended to M-modes, there will be a total of 2×(M!)2 routing states. For an N×N M-mode strictly non-blocking switch fabric, in which the paths from any input port to any output port can be established at the same time as the state of the switch elements on a given path is independent to all the others [18], [24], [39], [40], there are N!×(M!)N routing states in total.

Figure 2:

(A–D) Four different routing states for the TE₀ mode launched into port I2 for four switching configurations.

The orange stands for the ON state, while the gold stands for the OFF state.

The proposed switch is designed on the silicon-on-insulator platform with 220 nm thick top silicon layer. For the Y-junction shown in Figure 1D, the widths of the branch and stem parts are 0.4 and 0.8 μm, respectively. The branch is composed of two mirrored S-bends with radius and bend angle of 27.8 μm and 60°. Figure 1E shows the schematic view of the waveguide crossing composed of four multimode waveguides, which is divided into 12 segments with equal length. The particle swarm optimization method is utilized to optimize the geometry (w1-w13). The optimized parameters w1-w13 are 0.5, 0.6, 0.95, 1.32, 1.44, 1.46, 1.466, 1.52, 1.58, 1.62, 1.76, 2.15 and 0.5 μm, respectively. The waveguide crossing was reported in a previous paper with very low insertion loss of <0.03 dB and low crosstalk of <−37 dB [41]. The coupling gap and radii of the MR are chosen as 0.2 and 24 μm, so as to obtain a broad bandwidth and a high extinction ratio for high-speed transmission.

3 Results

To form the waveguide structure, 248 nm deep ultraviolet lithography and inductively coupled plasma etching were employed. Metal titanium was deposited on the top of the MRs and the arms of MZIs, serving as the heaters for thermal tuning. The SiO₂ cladding is utilized to cover the entire device, forming a buffer layer between the metal heater and waveguide. The microscope image of the fabricated device is shown in Figure 3. Extra mode MUXs based on adiabatic couplers are utilized to implement the multiplexing for input light coupled from an off-chip single-mode transmitter [42], as shown in Figure 3B. The region highlighted with a red dashed line shows the switch with a compact footprint as compact as 433 μm×433 μm. Figure 3C shows the zoom-in image of MR resonator.

Figure 3:

Fabricated multimode switch.

(A) Microscope image of the fabricated device consisting of the proposed multimode switch and mode (DE)MUXs. The region highlighted with the red dashed line shows the multimode switch. The areas highlighted in blue show the input/output multimode waveguides. Zoom-in images of (B) the mode MUX and (C) micro-ring resonator.

We measure the transmission spectra of the fabricated device over a wavelength range from 1520 to 1570 nm. For clarity, only spectra around 1550.29 and 1568.6 nm when the two modes are launched into port I2 are illustrated in Figure 4A and B. The legend indicates the states of the heaters H2 and H3; e.g. “F-O” refers to the transmission in the OFF state of H2 and ON state of H3. By deducting the loss caused by the grating couplers, mode MUX and DEMUX, the measured minimum and maximum insertion loss are 0.3 and 1.2 dB, respectively. The 3 dB optical bandwidth of the MR is 46.3 GHz, which agrees well with the designed value of 45 GHz. The mode-dependent insertion loss of <1 dB mainly results from the larger propagation loss of the Y-junction for TE₀ mode, since more power of TE₀ locates in the center of the waveguide, inducing a stronger reflection by the inner corner of the symmetric Y-junction. The path-dependent insertion loss is measured to be <0.5 dB, and the inter-path crosstalk and inter-mode crosstalk are <−16.6 dB and <−20 dB for all the cases. The major crosstalk comes from the inter-path one which can be further improved by adopting a MR with higher extinction ratio. The thermal instability of MRs can be overcome by implementing a thermal stabilization platform. For the MZI, the heating power and drive voltage are 18 mW and 7.1 V, when switching from OFF to the ON state. For the MR, the power of switching resonance wavelength to half FSR is 17.5 mW, and the voltage is 5.8 V. The thermal tuning structure is quite common and has been extensively applied in literatures, showing a typical switching time around tens of microseconds [43], [44].

Figure 4:

Measured transmission spectra around the wavelengths of (A) 1550.29 nm and (B) 1568.6 nm for all of the permutations and combination states of the heaters H2 and H3 when the TE₀ mode and TE₁ mode are launched into port I2.

The letters in the legend “F-O” refer to the OFF state of heater H2 and the ON state of heater H3, respectively.

In order to achieve a real commercial success, high-level integration by assembling numerous photonic components is the development trend, as it can improve the link price, performance and power dissipation. Thus, a large-scale and high-density integration of the switch fabrics, combining multiple multimode switch elements and making MDM technique more attractive and valuable, is highly desired. Here, we fabricated two types of switch fabrics with cross-bar and ring-bus architectures based on our proposed elements to verify high-level integration, as shown in Figure 5A and B, respectively. They both consist of four switch elements with eight thermal heaters, and the simplified schematics are given in Figure 5C and D. The area highlighted in red refers to the switch element. The black arrows indicate the light propagation directions. These two types of switch fabrics are optimized for different functionalities. For the cross-bar switch fabric, light is switched (at most) twice through the MR. There is almost no cumulative switching loss, in spite of scaling up to higher port count. The switch fabric is not strictly non-blocking as each input port has different properties, e.g. the light launched into port I1 can be routed to any output port, while the light path between port I2 and port O1 cannot be established physically. There is a total of 384 routing states for a 4×4 strictly non-blocking multimode switch fabric. Excepting the physically unestablished optical paths such as “I2 to O1” and “I3 to O4”, there remain 224 routing states. If the ports I3, I4, O1 and O2 are idle and the rest are used, the architecture will be strictly non-blocking. The ring-bus switch fabric has a rotational symmetry structure. Thus, each input port has the same property, and it allows arbitrary access to any output ports. Such a reconfigurable optical switch fabric makes full use of the optical paths. The ring-bus architecture is rearrangeable and non-blocking. For instance, if the optical path “I1-O4” is established, the path “I2-O3” will be blocked. Therefore, it might be necessary to rearrange ongoing transmissions to satisfy a certain connection. There are a total of 192 non-blocking routing states.

Figure 5:

Fabricated multimode switch fabrics.

Microscope images of the two types of 4×4 switch fabric with (A) cross-bar and (B) ring-bus architecture based on the multimode switch elements. The wire bonding technique is used for electrical connection between the on-chip electrode pads and the printed circuit board pads. Simplified schematics of (C) cross-bar and (D) ring-bus switch fabrics. The areas highlighted in red refer to the switch element. The black arrows indicate the light propagation directions. Measured insertion loss and crosstalk at each output port for all the permutations and combination states of the heaters at 1550.29 nm when the light injected from port I1 of both the (E) cross-bar switch fabric and (F) ring-bus switch fabric.

These two switch fabrics are fully characterized for all the channels and switching configurations. Both TE₀ and TE₁ modes at 1550.29 nm are injected from port I1 of the cross-bar switch fabric (Figure 5E) and ring-bus switch fabric (Figure 5F). The insertion losses are extracted by subtracting the loss of fiber-chip coupling, mode multiplexing and demultiplexing. For the cross-bar fabric, the insertion loss is 1.5–3.0 dB, and the path-dependent loss is <0.9 dB. Compared to the fundamental switch element, the cross-bar switch fabric has no significant increase in insertion loss. Approximate PILOSS functionality can be realized. For the ring-bus fabric, the insertion loss is 1.3–3.8 dB, and the path-dependent loss is <1.9 dB. The larger path-independent loss is attributed to the cumulative switching loss in the light path. The measured crosstalk for the two fabrics is lower than −15 dB. Note that the I1-O4 path, which includes eight MRs (four MRs in OFF state, the other in ON state) with 3.8 dB insertion loss, is the longest and most complicated path. Within the cumulative switching loss, a non-negligible part comes from the MRs’ resonance wavelength misalignment caused by thermal crosstalk. It can be efficiently cut down by adopting a larger gap or by using a deeply etched isolation trench between these MRs [17]. Compared with the reported 2×2 four-mode switch with loss of 10–14.9 dB [32], our 4×4 dual-mode switch fabrics achieve a much smaller insertion loss (3.8 dB) with similar scale, which are readily applied in high-port-count optical networks.

A modulated signal at 32 Gb/s was used to further test these two switch fabrics, and the experimental setup is shown in Figure 6A. Continuous wave light at 1550.29 nm generated by a tunable laser was launched into the Mach-Zehnder modulator driven by a NRZ-OOK signal (2³¹-1 pseudo-random binary sequences) from the bit pattern generator. The polarization controller was employed to optimize the polarization state of input signal to maximize the coupling efficiency between fiber and chip. After coupling out of the chip, the processed signal is firstly amplified by the erbium-doped fiber amplifier and then filtered by a band pass filter. The variable optical attenuator is utilized for controlling the power launched into a communication signal analyzer and an error analyzer for eye diagrams and bit error rate (BER) measurements, respectively. For convenience of presentation, here we only show the eye diagrams from port I1 to each output port for the cross-bar switch fabric and ring-bus switch fabric, as shown in Figure 6B. Clean and open eyes were observed for all the paths, indicating a good routing performance of these two switch fabrics. Finally, the BER measurements of the shortest and longest optical paths, i.e. “I1-O1” and “I1-O2” for the cross-bar switch fabric and “I1-O1” and “I1-O4” for the ring-bus switch fabric, are performed, as plotted in Figure 6C and D. For the cross-bar fabric, it can be seen that the maximal power penalty for different optical paths is less than 1 dB, which can be attributed to more MRs involved in the path. For the ring-bus fabric, similar results can be observed. Limited by the experimental facility, the demonstration with simultaneous dual-mode inputting cannot be performed. However, in our opinion, this will not affect the characterization of the proposed scheme. The same characterization method had been utilized by many previous multimode switches [31], [34].

Figure 6:

High-speed signal characterization for the two switch fabrics.

(A) High-speed signal testing setup including a tunable laser, a polarization controller (PC), a Mach-Zehnder modulator (MZM), a bit pattern generator (BPG), an Er-doped fiber amplifier (EDFA), a tunable bandpass filter, a variable optical attenuator (VOA), a communications signal analyzer (CSA) and an error analyzer (EA). (B) Measured 32 Gb/s eye diagrams for the B2B case, and each output port at 1550.29 nm when the light injected from port I1 of the cross-bar switch and ring-bus switch fabrics. Bit error rates of (C) optical paths “I1-O1” and “I1-O2” for the cross-bar switch fabric and (D) “I1-O1” and “I1-O4” for the ring-bus switch fabric.

4 Conclusion

In summary, we have proposed general solutions of 2×2 multimode switch, enabling reconfigurable inter-mode and inter-path routing. The switch has a simplified and low-loss design as it can process multiple modes at the same time while avoiding a demultiplexing-processing-multiplexing procedure. Minimum and maximum insertion losses as low as 0.3 and 1.2 dB and low crosstalk of <−16.6 dB for all the channels and switching configurations are experimentally obtained. Two types of switch fabric with cross-bar and ring-bus architectures based on the proposed switch element are fabricated, verifying high feasibility of high-density and large-scale integration. Measured results show a low insertion loss of <3.8 dB and a low crosstalk of <−15 dB. For high-speed communication, 32 Gb/s NRZ-OOK signal transmission is carried out, reaching up to 256 Gb/s of aggregate throughout. The proposed multimode switch can further promote the advanced and flexible MDM optical networks.