Polarization-insensitive silicon nitride Mach-Zehnder lattice wavelength demultiplexers for CWDM in the O-band

1. Introduction

Coarse wavelength division multiplexing (CWDM) has emerged as a promising option to meet the growing bandwidth requirements in optical interconnects. Implementations of wavelength multiplexers/demultiplexers (MUXes/DEMUXes) most commonly include arrayed waveguide gratings (AWGs), Echelle gratings, and Mach-Zehnder interferometer lattice filters (MZI-LFs). The drawback of the grating devices, such as AWGs and Echelle gratings, is that it is generally difficult to incorporate wavelength tunability and phase error compensation. Phase error correction is particularly acute in silicon (Si) photonics, since the thermo-optic coefficient of Si is high and the high refractive index contrast leads to a sensitivity of the phase-shift to dimensional variations [1]. Although Si AWGs [2–4] and Echelle gratings [5–7], as well as designs that provide some tuning [8,9] have been demonstrated, generally speaking, MZI-LFs are preferred for wavelength transparency and flexibility [10,11]. Beyond the spectral characteristics, another challenge to Si MUXes/DEMUXes is that they are highly polarization sensitive and require polarization diversity at the receiver or in the link. The polarization splitter-rotators or polarization splitting grating couplers can increase optical insertion loss and signal crosstalk, reduce the operation bandwidth, and requires two ideally identical copies of the same photonic circuit.

Recently, silicon-nitride (SiN) has attracted significant interest as an alternative to Si for passive devices [12–15]. SiN is CMOS compatible and is amenable to low-cost, high volume production. Owing to its lower index contrast and a 5× smaller thermo-optic coefficient compared with Si, SiN-based passive devices exhibit lower losses, better dimensional tolerances and lower thermal sensitivity. Using multilayer SiN-on-Si platforms, SiN devices can be combined with Si for active functionality (i.e., modulators and detectors) [13,16–20]. Broadband and low-loss optical power transfer between the Si and SiN layers is achieved using adiabatic taper transitions [21]. A variety of MUXes/DEMUXes have been demonstrated in SiN platforms [22–24].

In this article, we present polarization-insensitive 4-channel MZI-LF MUXes/DEMUXes for coarse wavelength division multiplexing (CWDM) in the O-band. The design was implemented in the CEA-Leti SiN-on-Si platform [17,19], even though only the SiN layer was used. The design used single-mode waveguides with a square cross-section, as well as birefringence minimized directional couplers. Polarization-insensitive Si-based MZI-LFs have been theoretically proposed, but they suffer from poor dimensional tolerances [25,26]. The larger waveguide thickness of SiN compared with standard Si-based platforms makes our square-core design uniquely suited for implementation in SiN. The interchannel crosstalk was reduced using a cascaded design. For an interchannel spacing of 20 nm, the SiN MZI-LF DEMUX with the best performance exhibited channel insertion losses < 2.8 dB, crosstalk < −11.5 dB, and a passband shift between orthogonal polarizations < 1.5 nm. Across the 200 mm diameter wafer, the die-averaged insertion loss and maximum crosstalk were 3.1 dB and −10.6 dB respectively. A polarization independent MUX/DEMUX simplifies the photonic integrated circuits for CWDM applications.

2. Design

The schematic of the 4-channel DEMUX is shown in Fig. 1(a). It consists of two stages of 2 × 2 MZI-LF blocks connected in a binary tree configuration. Similar to previous MZI-LF demonstrations [10,24], the delay line lengths and splitting ratios of the couplers in the MZIs are chosen to achieve flat passbands. The free spectral ranges (FSRs) of the filter blocks are determined by the wavelength channel spacing, δλ, of 20 nm for CWDM. In Stage 1, a third order lattice filter with an FSR of 2δλ separates the even and odd channels. In Stage 2, the wavelength pairs in each half are separated using a second order lattice filter with an FSR twice that in the first stage.

 

Fig. 1 (a) Schematic of the 4-channel DEMUX. The delay lengths for the MZIs are indicated in green and the splitting ratios for the directional couplers in red. (b) Block-diagram schematic of the 4-channel DEMUX with stage doubling implemented.

Download Full Size | PPT Slide | PDF

The device used 600 nm × 600 nm single-mode waveguides with square cores for polarization insensitive operation. The material refractive index of the SiN core was taken to be 1.87. The effective and group indices, which were equal for both polarizations, were calculated with an eigenmode solver to be n_eff = 1.587 and n_g = 1.904 respectively at a wavelength of λ_o =1310 nm. The base MZI length difference ΔL, which sets the FSR of the device, is given by $\mathrm{\Delta }L=\frac{{\lambda }_o^2}{n_g\delta \lambda }$. For some MZIs, it is necessary to shift the spectra by fractions of a single FSR by slightly adjusting the delay line length difference. The additional delay length needed to shift the MZI spectrum by one full FSR is given by $L_{FSR}=\frac{{\lambda }_o}{n_{eff}}$. The channel spectra were designed to be centered at λ₁ = 1291 nm, λ₂ = 1311 nm, λ₃ = 1331 nm, and λ₄ = 1351 nm.

One limitation to the polarization-insensitive operation is the directional couplers. Although the couplers use square-core waveguides, the coupled waveguide geometry breaks the symmetry for the transverse electric (TE) and transverse magnetic (TM) modes, so the splitting ratio, K, is polarization dependent. The directional coupler geometries were designed using three-dimensional finite difference time domain (3D-FDTD) simulations. We selected the coupling gaps and the coupling lengths to minimize the objective function

We calculated the DEMUX response using a simplified transfer matrix model neglecting propagation losses. The computed channel spectra are shown in Fig. 2(a) with both polarizations superimposed. The flat-top passband shapes are nearly identical for TE and TM polarizations. For TE polarization, the design shows significant crosstalk in the λ = 1351 nm port where the coupler splitting ratios deviate maximally from their target values. To improve the crosstalk, we implement a stage doubling scheme in which each MZI-LF block is repeated as illustrated in Fig. 1(b). This stage doubled design reduces crosstalk at the expense of a larger footprint, higher loss, and a slightly reduced passband bandwidth. The calculated channel spectra with stage doubling is shown in Fig. 2(b). In this case, the crosstalk is maintained below -21 dB for all channels and polarizations.

 

Fig. 2 Computed transmission spectra of the 4-channel demultiplexer (a) without and (b) with stage doubling. The solid and dashed lines indicate TE and TM polarization respectively.

Download Full Size | PPT Slide | PDF

3. Measurements

The devices were fabricated on a 200 mm silicon-on-insulator wafer in a SiN-on-Si platform at CEA-LETI [17,19]. The SiN was deposited using plasma-enhanced chemical vapor deposition (PECVD) and patterned using 248 nm deep ultraviolet photolithography. Figure 3(a) shows an optical micrograph of the fabricated 4-channel demultiplexer with stage doubling. The optical input and output ports are labelled by their wavelengths. The other ports of the MZI-LF routed to the chip edge were used for debugging purposes. We used inverse taper SiN edge couplers for on/off chip coupling to lensed fibers. Without the edge couplers, the device had a footprint of 2.5 mm × 0.9 mm.

 

Fig. 3 (a) Optical micrograph of the fabricated 4-channel demultiplexer with stage doubling. (b) Schematic of the measurement setup. TLS is the tunable laser source. DUT is the device under test.

Download Full Size | PPT Slide | PDF

Figure 3(b) shows the measurement setup. An O-band tunable laser source (TLS) provided optical input and an optical component analyzer (Keysight N7788B) was used to set and monitor the polarization state. The integrated polarization controller is comprised of a series of programmable waveplates. We found that applying voltages to three of the five available waveplates provided full control of the polarization state while maintaining a uniform polarization over the measurement bandwidth. The 3 dB fiber power splitter following the polarization controller routed half the signal to the polarimeter for polarization state monitoring, while the other half was coupled to the Device Under Test (DUT). Finally, the DUT transmission was measured using an optical power meter. The insertion loss of the polarization controller was around 3 dB. The edge coupler losses as determined from a straight waveguide calibration structure was about 4.5 dB. The SiN square waveguide propagation loss, which was similar for both polarizations, was measured to be < 0.9 dB/cm at a 1310 nm wavelength and < 1.9 dB/cm across the 1270-1370 nm measurement bandwidth.

We measured the device transmission with the polarization controller set to scramble the polarization state at a rate much faster than the integration time of the power meter. This had the functional effect of sending a balanced TE/TM polarization mixture into the DUT, and the spectral measurements show the polarization averaged transmission. Figure 4(a) shows the device transmission spectra after the losses from the experimental setup and edge couplers have been de-embedded. We find that the demultiplexing function is preserved for a polarization-averaged input. The insertion loss ranged from 0.9 dB to 2.8 dB among the channels. Within a 10 nm bandwidth about the passband center wavelengths, the passband ripple was < 0.8 dB and the worst-case channel crosstalk was < −11.5 dB across all channels. The passband centers were red-shifted by about 6 nm compared to the design values, which could be the result of deviations in the waveguide dimensions or the SiN material index.

 

Fig. 4 (a) Measured channel spectra with polarization scrambled (polarization-averaged) input. (b) Measured transmission of a single port with scrambled (Avg) input and two orthogonal polarization states (Max,Ortho).

Download Full Size | PPT Slide | PDF

To determine the polarization dependence, we measured the transmission for two orthogonally polarized inputs. First, the polarization state was adjusted to maximize transmission for a representative wavelength within the channel passband. We optimized the polarization states at the wavelengths of 1300, 1320, 1340, and 1360 nm, which were within the channel passbands for all the devices that we measured. The Stokes parameters for the corresponding polarization state at the polarimeter port were recorded. Next, we adjusted the polarization controller to produce an orthogonal polarization at the polarimeter port. This also produced the orthogonal polarization at the DUT input, since the fibers of the experimental setup were not perturbed between measurements. We found that the orthogonal polarization state was nearly identical to the polarization state which minimized transmission at the wavelength of optimization. Figure 4(b) shows the measured transmission for the λ = 1300 nm port when the input polarization state was scrambled (‘Avg’), maximized at 1300 nm (‘Max’), and set to the orthogonal state (‘Ortho’). The passband shift between ‘Max’ and ‘Ortho’ polarizations was < 1.5 nm for all 4 channels, and the polarization dependant loss (PDL) was < 0.9 dB. Although we have chosen to quote our results based on directly measured powers, we also characterized the device using a Mueller matrix extraction method with the polarimeter port connected to the device output [27]. With this method, we found a TE/TM passband shift < 1.6 nm and a PDL < 1.3 dB.

To illustrate the robustness of the DEMUX, we measured devices from 10 dies across the wafer. With the exception of a single channel on one die, all of the tested devices functioned for both input polarizations. The overlapping histograms in Fig. 5 summarize the performance of each channel for the nine working devices. The DEMUX performance was mostly limited by the λ₂ port centered near 1320 nm, which had the highest insertion loss, passband ripple, and crosstalk. The cross-die averages and standard deviations (in dB) of the worst-performing channel are −3.1 ± 0.5 dB for the insertion loss, 1.4 ± 0.4 dB for the PDL, 0.8 ± 0.1 dB for the passband ripple, and −10.6 ± 0.9 dB for the maximum passband crosstalk. The passband centre wavelengths were fairly consistent between dies; for all four channels, the standard deviation of the center wavelengths across the wafer was between 0.9 nm and 1 nm.

 

Fig. 5 Overlapping channel histograms of (a) insertion loss (b) polarization-dependent loss (c) passband ripple and (d) crosstalk for DEMUXes on 9 dies from across the wafer.

Download Full Size | PPT Slide | PDF

4. Discussion

Compared to the designed values, the fabricated DEMUXes exhibited a red-shift in the passband spectra and significantly higher channel crosstalk. The deviation was most evident at shorter wavelengths, where the overall device crosstalk was limited by the crosstalk of the λ = 1300 nm port into the λ = 1320 nm passband. In addition, the adjacent order passband centered at λ = 1280 nm was severely degraded compared to simulation. To determine the possible causes of these effects, we used a first derivative approximation to examine the simulated channel response with respect to deviations in the SiN waveguide width Δw, SiN waveguide height Δh, and the directional coupler gaps Δg. Changes to the geometrical parameters of the device affect the spectral response by introducing phase errors in the MZI-LFs and altering the splitting ratios of the directional couplers. We did not explicitly include the waveguide sidewall angle as a geometrical parameter because we found that, for small sidewall angles, the waveguide cross-section was well approximated by a rectangular waveguide of width equal to the average of the widths at the top and the bottom of the waveguide.

Figure 6(a) shows the comparison between the measured data and the simulated response for a polarization averaged input when [Δw,Δh,Δg] = [10 nm, 35 nm, 45 nm], which was the combination that minimized the deviation between measured and simulated normalized spectra. With these parameters, our model produces a close qualitative match with the measured spectra. These dimensional offsets are also consistent with cross sectional scanning electron micrographs of calibration structures that showed the SiN thickness was approximately 630 nm. The computed insertion loss for the worst channel is −1.85 dB, which is lower than the measurements because our model assumes lossless waveguides and couplers. The computed maximum crosstalk is −13.2 dB, close to the measured values of −10.6 ± 0.9 dB across the wafer. The principal cause of the increased crosstalk was the lower than expected cross-coupling coefficients at short wavelengths, leading to a reduced extinction ratio in the bottom port of Stage 1 and a spectral broadening of the λ = 1300 nm port. Figure 6(b) shows the simulated response for TE and TM polarized input. We find that the dimensional offsets induced an average passband shift between the polarizations of 1.5 nm, which is consistent with measured values.

 

Fig. 6 (a) Measured (solid lines) and simulated (dashed lines) spectra for a polarization averaged input. (b) Simulated spectra for TE (dotted lines) and TM (dashed lines) input polarizations.

Download Full Size | PPT Slide | PDF

5. Conclusion

In summary, we have presented a SiN polarization insensitive cascaded MZI wavelength DEMUX in the O-band. With a double-stage design, the channel insertion loss was < 2.8 dB and the crosstalk was < −11.5 dB for the best device. For nine dies measured across the wafer, the die-averaged insertion loss and maximum crosstalk were 3.1 dB and −10.6 dB respectively. The high crosstalk was likely due to the dimensional variations, which may be improved in the future. This work is synergistic with the development of SiN-on-Si multilayer polarization-independant grating couplers [28]. This proof-of-concept demonstration shows that the SiN MZI-LF DEMUX is a promising alternative to polarization diversity for CWDM.