1. Introduction

Very large-scale photonic integrated circuits (PICs) comprising hundreds and thousands of optoelectronic devices are being explored for complex chip architectures, such as those for optical switch fabrics, optical phased arrays, and coherent communications involving several degrees of multiplexing [1–5]. Silicon (Si) photonics, which uses the mature manufacturing infrastructure of CMOS electronics to fabricate photonic components, is a promising technology for high throughput production of very large-scale PICs due to the availability of large-area Si substrates and high-yield fabrication processes. Very large-scale PICs often have complex on-chip waveguide routing networks and many waveguide crossings (see, for example, [2] for a 32 × 32 optical switch). To make possible increasing integration densities, three-dimensional (3D) Si photonic platforms need to be considered. However, most Si photonic platforms today are two-dimensional (2D), with waveguides defined only in a single Si layer.

The limitations of 2D photonic platforms for very large-scale integration become evident when waveguide crossings are considered. With a single waveguide layer, waveguide crossings must be in-plane. The loss and crosstalk of a crossing is fundamentally limited by the refractive index contrast of the waveguides. For Si strip waveguides, crossings with losses less than 40 to 30 mdB (milli-decibels) have been demonstrated in the C- and O-bands, with crosstalk in the −35 dB to −40 dB range [6,7]. The loss can be reduced by lowering the index contrast of the Si waveguides. In a single layer, a lower effective index contrast can be achieved using subwavelength grating structures [8,9], but their production requires fine feature control which is difficult to maintain in a standard 248 nm or 193 nm deep ultraviolet (DUV) lithography process. A lower index contrast can also be accomplished using a rib waveguide geometry that uses the partially etched level of the Si, at the expense of a larger crossing size [10]. Using this method, crossing losses as low as 15 mdB per crossing have been predicted [11]. Using fully etched silicon nitride (SiN) waveguides, in-plane crossings with losses better than 70 mdB across the SCL bands have been demonstrated [12]. When hundreds of waveguide crossings are needed in a PIC, crossings can contribute to an insertion loss of the order of 10 dB as well as substantial crosstalk.

A 3D platform enables the realization of over- and under-pass types of waveguide crossings, in which a waveguide in a top (lower) layer can pass over (under) many waveguides in a lower (upper) layer. 3D platforms can be implemented by integrating SiN or amorphous Si waveguide layer(s) on a silicon-on-insulator (SOI) waveguide layer [13–19]. However, with two layers, a trade-off exists between the interlayer transition loss and the crossing loss/crosstalk. For efficient interlayer transitions with minimal losses, the two waveguide layers should be close to each other; but for the crossing to have low loss and low crosstalk, the two waveguide layers should be far apart. Most of the past demonstrations have reported crossing losses dominating the interlayer transition loss [13–15]. For example, in [15], the reported loss per transition and loss per crossing at 1550 nm were 10 mdB and 167 mdB, respectively. Thus far, the overall crossing loss in 2-layer SiN-on-SiN or SiN-on-Si photonic platforms has not been definitively lower than the loss of single-layer in-plane crossings; although, very recently, the crossing loss in a multilayer amorphous Si platform has been shown to be lower than that of single-layer in-plane crossings [18].

To overcome this trade-off between the interlayer transition loss and crossing loss, we have proposed Si photonic platforms that use 3 waveguide layers [16]. With 3 layers, as in Fig. 1(a), the center layer can provide an intermediary transition between the bottommost and topmost layers, as shown in Figs. 1(b) and 1(c), such that a large total interlayer spacing can be achieved while maintaining low interlayer transition losses. In this article, we present a 3-layer Si photonic platform, consisting of two SiN waveguide layers and one Si waveguide layer realized on 8″ SOI substrates fabricated at the A*STAR Institute of Microelectronics. The lower Si layer allows for the incorporation of PN junctions for modulators and germanium photodiodes. While the upper SiN layers, in addition to enabling low-loss crossings, allow for high quality passive devices since, compared to Si, SiN waveguides have lower scattering losses, better variation tolerance due to the lower refractive index contrast, less temperature sensitivity, and better power handling [12, 20]. The platform was optimized for the C-band and measurements across the SCL bands are discussed. We had previously presented a similar platform optimized for the O-band in [16] integrated with high-speed modulators and detectors but with lossy 3-layer waveguide transitions. Here, we show overpass crossings with ultra-low losses between 0.28 mdB and 2.1 mdB and crosstalk of < −56 dB in the wavelength range of 1480 nm to 1620 nm, and a 3-layer transition loss < 150 mdB. The 3-layer transition loss was primarily limited by the Si to SiN transition performance at short wavelengths. The results show the feasibility of multilayer SiN on Si platforms to realize complex on-chip interconnect networks with extraordinarily low-loss and low-crosstalk crossings.

 

Fig. 1 (a) Schematic of the proposed three layer SiN-on-Si platform with the thicknesses labelled. SiN1 refers to the lower SiN layer, and SiN2 refers to the upper SiN layer. The buried oxide thickness is 2 µm. (b) Illustration of adiabatic interlayer transitions in the platform for transferring light between the three layers. (c) Illustration of low-loss waveguide crossings in the platform, in which SiN2 waveguides pass over Si rib waveguides with a contiguous partially etched Si slab. (d),(e): X-TEM images of (d) a fabricated SiN1 waveguide atop a Si waveguide and (e) a fabricated SiN2 waveguide atop a SiN1 waveguide in the platform.

Download Full Size | PDF

2. Design and fabrication of the multilayer platform

The cross-section schematic of the multilayer platform with nominal thicknesses is illustrated in Fig. 1(a). It consists of two 400 nm thick SiN waveguide layers and a 220 nm thick Si waveguide layer separated by silica (SiO₂) spacers. The separation between the lower and topmost SiN layers (referred to as SiN1 and SiN2 respectively) is 200 nm in height, and the separation between the Si and SiN1 layers is 250 nm. A partial etch step in the Si layer is used to define a 90 nm thick slab layer. The buried oxide (BOX) layer thickness is 2 µm. The platform was fabricated on 8″ SOI wafers, and full and partially-etched Si waveguides were formed using DUV lithography and reactive-ion etching (RIE). The first SiO₂ spacer layer was formed by SiO₂ deposition, a SiO₂ back etch, and chemical mechanical planarization (CMP). The SiN1 layer was deposited by low pressure chemical vapor deposition (LPCVD) to form nearly stoichiometric Si₃N₄, and SiN1 waveguides were formed using DUV lithography and RIE. The second SiO₂ spacer layer and SiN2 waveguide layer were formed using the same methods as the first spacer layer and SiN1 waveguides. Finally, the SiO₂ top cladding was deposited, deep trenches were etched to form edge couplers, and the wafer was diced. SiN deposited by LPCVD, rather than PECVD or ICP-CVD, was used since PECVD and ICP-CVD SiN typically have stronger absorption peaks near a wavelength of 1520 nm due to the hydrogen content of the SiN [12, 21].

Figures 1(d) and 1(e) show cross-section transmission electron micrographs (XTEMs) of a SiN1 waveguide atop a Si waveguide and a SiN2 waveguide atop a SiN1 waveguide, respectively, in the interlayer transitions of the fabricated platform. From the XTEMs, the measured Si, SiN1, SiN2 thicknesses were 217 nm, 385 nm, and 385 nm, respectively, and the measured Si-SiN1 and SiN1-SiN2 interlayer spacer thicknesses were 305 nm and 245 nm, respectively. These thicknesses are similar to the design in Fig. 1(a). Figure 1(d) shows that the SiN1 layer was not fully etched in some regions. The residual SiN1 layer may have led to higher interlayer transition losses than those simulated in Section 2.1. The wafer-scale SiN thickness non-uniformity is ~2%, and a previous report from A*STAR IME shows a CMP non-uniformity of ∼ 1% for a 1.1 µm thick SiO₂ layer [22].

2.1. Interlayer transitions

The multilayer photonic platform uses adiabatic linear taper transitions to efficiently transfer optical power between the waveguide layers. Figure 2(a) shows the designed transitions between the Si and SiN1 layers and the SiN1 and SiN2 layers. The inputs and outputs of the transitions are single-mode waveguides with widths of 500 nm in the Si layer and 900 nm in SiN layers. The Si waveguide tapers down to a blunt tip with a 140 nm width, and the SiN waveguides taper to blunt tips of 300 nm width. The taper lengths are 70 µm and 120 µm for the Si-SiN1 and SiN1-SiN2 transitions, respectively.

 

Fig. 2 (a) Top-down schematics of the Si-SiN1 and SiN1-SiN2 interlayer transitions. Mode intensity profiles of the fundamental TE mode at a 1550 nm wavelength are shown at multiple points along the transitions. (b) 3D-FDTD simulated insertion loss (IL) of the interlayer transitions for the TE polarization.

Download Full Size | PDF

Figure 2(b) shows the simulated interlayer transition losses for the transverse-electric (TE) polarization computed using the three-dimensional finite difference time domain method (3D-FDTD). The refractive indices were taken to be 3.48 and 1.96 for the Si and SiN layers, respectively, and the thicknesses in Fig. 1(a) were assumed. Over the wavelength range 1480 to 1600 nm, the Si-SiN1 transition loss is < 13 mdB and the SiN1-SiN2 transition loss is < 35 mdB. In both cases, the transition losses generally increase at shorter wavelengths as the optical confinement in the Si and SiN increases. Designs intended for shorter wavelengths in the O or S-bands may benefit from thinner waveguide and interlayer spacer layers [12]. Accounting for both the Si-SiN1 and SiN1-SiN2 transitions, the total loss incurred by transitioning from Si to SiN2 is < 48 mdB. Additional transition loss simulations were performed using the measured layer thicknesses from Figs. 1(d) and 1(e) at a wavelength of 1550 nm. The Si-SiN1 transition loss increased by < 3 mdB, and the SiN1-SiN2 transition loss decreased by < 6 mdB. In practice, the measured transition loss is expected to be larger due to waveguide propagation losses, which may contribute an additional loss to the Si-SiN1-SiN2 transition of 19–95 mdB for average propagation losses of 1–5 dB/cm.

2.2. Waveguide crossings

As shown in Fig. 3(a), which is a top-down view of Fig. 1(c), overpass crossings are formed when light confined in the SiN2 waveguide passes over Si waveguides. Underpass crossings are formed when light confined in a Si waveguide passes under SiN2 waveguides. As a result of the lower optical confinement in the SiN waveguide compared to the Si waveguide, the overpass crossing loss is generally higher than that of the underpass crossing, and thus, much of the design effort in the multilayer platform is in reducing the overpass crossing loss. In this work, we focus on 90° overpass and underpass crossings.

 

Fig. 3 (a) Top-down schematic view of overpass and underpass waveguide crossings using the Si and SiN2 layers. (b) Simulated overpass and underpass crossing insertion loss (IL) for the TE-polarization with and without the Si slab layer (i.e., using partially-etched Si rib waveguides versus fully-etched Si strip waveguides); < 1 mdB crossing losses are achieved for both overpass and underpass crossings when using partially-etched Si rib waveguides.

Download Full Size | PDF

The overpass crossing loss is due to scattering from the Si waveguide, which perturbs the optical mode in the SiN2 waveguide. According to first-order perturbation theory (valid in the limit of weak scattering), the magnitude of the scattered field is proportional to the overlap of the unperturbed SiN2 waveguide mode in the region of the index perturbation. Therefore, to reduce the crossing loss, we reduce the volume of the index perturbation by passing over Si rib waveguides that share a contiguous slab as shown in Fig. 3(a) rather than Si strip waveguides. The partially-etched Si slab region is introduced underneath the SiN2 waveguide using adiabatic tapers, which have a simulated insertion loss of < 54 mdB per transition from 1480 to 1600 nm. To further reduce the crossing loss, we increase the modal confinement in the SiN2 waveguide by widening the waveguide width to 1.5 µm in the area of the crossing by means of a linear taper. Although the widened waveguide supports two TE-polarized modes, the adiabatic taper ensures only the fundamental mode is excited.

3D-FDTD simulations of the overpass and underpass crossing losses for the TE polarization over a wavelength range of 1480 to 1600 nm are shown in Fig. 3(b). The waveguide thicknesses and interlayer separations in Fig. 1(a) have been assumed. The figure compares crossing losses with and without the Si slab, i.e., for partially-etched Si rib waveguides versus fully-etched Si strip waveguides. Without the Si slab, the overpass crossing loss is < 18 mdB, and with the slab, the overpass loss is reduced to < 0.29 mdB, over an order of magnitude of improvement. The underpass loss is very low in both cases but also improved substantially by introducing the Si slab, which reduces underpass losses by increasing the portion of the modal power in the Si layer and reducing the modal overlap with the cladding. Without the Si slab, the underpass loss is < 0.26 mdB, and with the Si slab, the underpass loss is < 0.04 mdB and limited by the accuracy of the simulation. Overall, the overpass and underpass crossing designs in Fig. 3(a) enable sub-millidecibel losses, though in practice, propagation loss in the waveguide length between adjacent crossings may dominate the measured loss per crossing, e.g., crossings on a 5 µm pitch will have an additional 0.5 – 2.5 mdB of loss for waveguide propagation losses of 1 – 5 dB/cm. For completeness, in Fig. 3(b), from 1480 to 1600 nm, without the Si slab, the overpass/underpass loss varies between 9.00/−0.04 and 17.66/0.22 mdB, and with the Si slab, the overpass/underpass loss varies from 0.12/−0.01 to 0.24/0.02 mdB; negative losses are due to the limited simulation accuracy. Additional crossing loss simulations were performed using the measured layer thicknesses in Figs. 1(d) and 1(e) at a wavelength of 1550 nm for the design using the Si slab. The overpass loss decreased by about 0.1 mdB and the underpass loss remained < 0.04 mdB.

3. Measurements

In this section, measurement results are presented for the waveguide propagation loss, interlayer transition loss, and overpass crossing loss and crosstalk. The cutback method was used for the loss measurements. Although the transitions and crossings have been designed for the TE polarization, measurements for the transverse magnetic (TM) polarization have also been included for completeness. The devices were measured using a swept tunable laser source with an inline fiber polarization controller. Edge couplers and lensed fibers were used for on/off-chip coupling.

3.1. Propagation losses

TE propagation loss spectra for 900 nm wide, single-mode waveguides in the SiN1 and SiN2 layers are shown in Fig. 4. The inset shows the linear fit for the SiN1 waveguide cutback structures at a 1550 nm wavelength, where the loss was 1.09 dB/cm. The increased SiN1 and SiN2 propagation losses around a 1520 nm wavelength is a result of material absorption due to residual N-H bonds. This absorption may be significantly reduced by high-temperature thermal annealing [23]. At longer wavelengths > 1580 nm, the SiN layer losses were < 0.45 dB/cm, less than the Si rib waveguide losses. Over this wavelength range, 500 nm wide Si rib waveguides had a propagation loss between 0.8 – 1.5 dB/cm, in agreement with previous reports [24]. Due to mask space constraints, only three cutback structures were fabricated for the measurements in Fig. 4. Nonetheless, the linear fits were excellent from 1480 – 1565 nm (R ² > 0.99) and reasonable from 1565 – 1620 nm (R ² > 0.86) where the waveguide losses dropped below 0.7 dB/cm and the alignment error and Fabry Perot oscillations approached the resolution of the cutbacks. In addition, the waveguide losses in Fig. 4 agree with our previously reported LPCVD SiN waveguide losses [12], which used waveguides of the same nominal dimensions.

 

Fig. 4 Measured TE propagation loss for SiN1 and SiN2 single-mode waveguides in the SCL-bands. Inset: Cutback measurement for a SiN1 single-mode waveguide for TE-polarized light at a wavelength of 1550 nm.

Download Full Size | PDF

3.2. Interlayer transitions

The Si-SiN1 and SiN1-SiN2 interlayer transition losses were measured using a series of cutback structures, which consisted of large, progressively increasing numbers of cascaded interlayer transitions. Cutback structures with up to 200 and 136 transitions were used for the Si-SiN1 and SiN1-SiN2 transitions, respectively. An optical micrograph of a Si-SiN1 transition and a SiN1-SiN2 transition in cascade is shown in Fig. 5(a). The edge couplers for all interlayer transition cutback structures were in the SiN1 layer. To prevent artifacts in the cutback measurements due to spectral ripple from periodic reflections at the blunt tip interfaces, the lengths of the straight waveguides between the interlayer transitions were randomized by drawing from a continuous uniform distribution between 5 µm and 8 µm. The linear fit of the cutback measurements to extrapolate the transition loss was excellent. As an example, Fig. 5(b) shows the fit for the SiN1-SiN2 transition at a wavelength of 1550 nm. Waveguide propagation losses have not been de-embedded from the interlayer transition loss measurements.

 

Fig. 5 (a) Optical micrograph of the three-layer transitions. (b) Cutback measurement for the SiN1-SiN2 interlayer transition for the TE polarization at a 1550 nm wavelength. (c),(d): Interlayer transition loss spectra for the (c) TE and (d) TM polarizations. (e),(f): TE interlayer transition loss spectra of multiple dies from the same wafer for (e) Si-SiN1 and (f) SiN1-SiN2 transitions. Die 1 refers to the die measured in (a)–(d), and unless otherwise stated, all other measurements in this paper refer to die 1.

Download Full Size | PDF

The TE transition loss spectra for both types of interlayer transitions are shown in Fig. 5(c). The SiN1-SiN2 TE transition loss was < 69 mdB over the measured 1480–1620 nm wavelength range and is highest at a wavelength of 1520 nm, where the SiN waveguide propagation loss was the highest. For the Si-SiN1 transition, the TE insertion loss was < 107 mdB across the measured wavelength range and was the highest at 1480 nm. In addition, Fig. 5(c) shows that the TE loss for the concatenated Si-SiN1 and SiN1-SiN2 three-layer transitions (i.e., Si-SiN2 transition) was < 150 mdB over the 1480–1620 nm wavelength range. In total, 4 dies from the wafer were measured, and the TE transition losses of all 4 dies are shown in Figs. 5(e) and 5(f). Only minor transition loss variations were observed between the dies, which may be partly due to measurement error. The higher than expected interlayer transition loss (compared to Fig. 2(b)), especially for the Si-SiN1 transition, may have been caused by the residual SiN1 layer, waveguide height variation, and the larger than designed interlayer separations. Despite the deviations of the fabricated waveguides from the ideal, the interlayer transitions exhibited low losses, showing the robustness of the adiabatic transition designs.

Figure 5(d) shows that the interlayer transitions also function for the TM polarization, albeit with higher losses. Over the 1480–1620 nm wavelength range, the TM transition losses were < 205 mdB, < 201 mdB, and < 350 mdB for the Si-SiN1, SiN1-SiN2, and Si-SiN1-SiN2 transitions, respectively.

To investigate the tolerance of the interlayer transitions to mask misalignments, transition cutback structures incorporating lithographically defined ±60 nm offsets between the waveguide layers in the direction perpendicular to the optical propagation axis were measured. The Si-SiN1 and SiN1-SiN2 TE-polarization transition loss spectra including the offsets are shown in Figs. 6(a) and 6(b), respectively. At a 1550 nm wavelength, the excess loss due to waveguide misalignment was about 10 mdB for a Si-SiN1 transition and 14 mdB for a SiN1-SiN2 transition. Across the full 1480 – 1620 nm wavelength range, the TE loss was < 0.14 dB for both transition types with ±60 nm interlayer offsets. Similarly, the TM-polarization transition loss in Figs. 6(c) and 6(d) was < 0.21 dB for both types of transitions with ±60 nm offsets across the 1480 – 1620 nm wavelength range. Interlayer offsets parallel to the optical propagation axis are not explored here since these interlayer transitions are known to be relatively insensitive to even large 100 nm misalignments along this axis [12].

 

Fig. 6 (a),(b) TE and (c),(d) TM insertion loss spectra for (a),(c) Si-SiN1 and (b),(d) SiN1-SiN2 interlayer transitions for lithographically defined −60 nm, 0 nm, and +60 nm interlayer offsets. Offsets are defined relative to the SiN1 layer and are in the direction perpendicular to the optical propagation axis.

Download Full Size | PDF

3.3. Waveguide crossings

Figure 7(a) shows an optical micrograph of a portion of an overpass crossing cutback structure, which consisted of a serpentine SiN2 waveguide passing over a set of Si rib waveguides. The full cutback structure had 21 periods of the serpentine SiN2 waveguide. The pitch of the Si waveguide array was randomized by drawing from a continuous uniform distribution between 5 µm and 7 µm to avoid periodic reflections, and in the measurements discussed below, waveguide propagation losses in the pitch between adjacent crossings have not been removed. These cutback structures enable overpass crossing loss measurements as well as crosstalk measurements of the crossings. The four cutback structures have 840, 2100, 3360, and 4620 crossings. Due to mask space limitations, additional cutback structures necessary for underpass loss measurements were not fabricated. However, the underpass loss is not expected to exceed the measured overpass loss by more than 1 mdB since: 1) the simulated underpass loss in Fig. 3(b) is substantially less than the simulated overpass loss, 2) embedded waveguide propagation losses (Section 3.1) from the 5 – 7 µm crossing pitch will be larger for the underpass by at most 1 mdB, and only at longer wavelengths away from the SiN absorption peak.

 

Fig. 7 (a) Optical micrograph of a portion of an overpass crossing cutback structure. (b) TE overpass crossing loss measurements for multiple dies from the same wafer; the overpasses used a 1.5 µm SiN2 width. (c),(d): Die 1 overpass crossing losses for the (c) TE and (d) TM polarizations; overpass losses with 0.9 µm and 1.5 µm SiN2 widths are shown. (e),(f): TE cutback measurements for a 1.5 µm SiN2 width for (e) dies 2 and 4 at a wavelength of 1550 nm and (f) die 4 at a wavelength of 1582.76 nm.

Download Full Size | PDF

The measured TE overpass crossing losses in the SCL-bands for 4 dies from the same wafer are shown in Figs. 7(b) and 7(c). With a 90% confidence interval, for dies 1–4, the overpass losses with a 1.5 µm SiN2 width were < 2.6, < 2.1, < 2.0, and < 2.1 mdB, respectively, across the entire 1480 – 1620 nm measured wavelength range. In addition, the overpass losses of dies 2–4 were < 1 mdB from 1550 – 1620 nm with a 90% confidence interval; the die 1 overpass loss was < 1.5 mdB over this wavelength range due to higher error in the linear fits. The TM overpass losses in Fig. 7(d) are higher due to the reduced confinement of the TM0 mode in the SiN2 waveguide compared with the TE0 mode. The relatively large ripple in the TM losses may be due to the accumulation of reflections in the cutback structure, which are expected to be larger compared to the TE case. The TM overpass loss linear fits are generally less accurate than the TE case, and accounting for error, the TM overpass losses are < 14 mdB across the measured wavelength range with a 90% confidence interval. Figures 7(e) and 7(f) show example linear fits of the TE transmission of the set of crossing loss cutback structures (1.5 µm SiN2 width) from dies 2 and 4 at 1550 nm and 1582.76 nm wavelengths. The overpass loss of die 4 is minimized near 1582.76 nm, and the loss with a 90% confidence interval is < 0.28 mdB.

Despite using cutback structures with up to 4620 crossings, over portions of the spectrum, the ultra-low crossing loss led to the fiber coupling loss misalignment error and Fabry-Perot oscillations from the edge couplers being non-negligible compared to the loss difference between subsequent cutback structures. For dies 1–4, the range of wavelengths over which linear fits with R² > 0.9 were measured was 1503–1537 nm, 1480–1570 nm, 1480–1582 nm, and 1480–1590 nm, respectively. Outside these wavelength ranges, the linear fits are less accurate and the confidence intervals corresponding to the traces in Fig. 7(b) are larger. Importantly, for the highest crossing loss wavelength point, R² = 0.964 for die 1 and R² > 0.998 for dies 2–4, which confirms the accuracy of the measured 2.0–2.6 mdB upper bounds on the loss per overpass across the SCL-bands.

3.4. Crosstalk extraction

The overpass and underpass crosstalks were measured using the shortest overpass cutback structure, in which a Si rib waveguide passed under the SiN2 serpentine waveguide 42 times. The overpass(underpass) measurement involved injecting light into the SiN2(Si) waveguide and measuring the crosstalk in the Si(SiN2) waveguide. The crosstalk was determined by de-embedding the edge coupler losses from the cross-port transmission measurement. Since the crosstalk signal was low, the measurement was sensitive to fiber-to-chip misalignment and to scattered light in the substrate and superstrate.

Here, we describe the extraction procedure for the overpass crosstalk; the underpass crosstalk extraction proceeds in an analogous manner. First, the Si edge coupler response, η_Si, was obtained by measuring the transmission, $T_1={\eta }_{Si}^2$, of a single Si rib waveguide. The input lensed fiber was then moved to the SiN2 waveguide input port to obtain the raw cross-port transmission spectrum, T_SiN2→Si, which is approximately given by

Next, we measured the SiN2 waveguide transmission with the output lensed fiber moved to the SiN2 waveguide output port and the input fiber position fixed. This gives the transmission, $T_2={\eta }_{SiN2}^{\prime }{\eta }_{SiN2}$, where η_SiN2 is the optimally aligned SiN2 edge coupler response. Finally, we remeasured the spectrum after adjusting the position of the input fiber to maximize the transmission to obtain $T_3={\eta }_{SiN2}^2$. Since ${\eta }_{SiN2}^{\prime }=T_2\sqrt{T_3}$, Eq. 1 can be used to calculate κ_SiN2→Si. In the measurements, η_SiN2 could be higher than ${\eta }_{SiN2}^{\prime }$ by as much as 6 dB.

Equation 1 makes the simplifying assumptions that coherence effects are negligible and each of the 42 crossings makes an equal contribution to the measured cross-port signal. In actuality, the crossings do not make an equal contribution due to waveguide propagation loss. In Eq. 1, the waveguide propagation loss is lumped with η_Si and ${\eta }_{SiN2}^{\prime }$. We compared the crosstalk calculated from Eq. 1 with a more accurate approach in which the accumulated loss for the pathway through each individual crossing is computed and summed together. We find that Eq. 1 results in an overestimation of at most 0.1 dB in the crosstalk values. In addition, due to the crossings relevant to the crosstalk measurement being separated by long waveguide lengths ≥100(210) µm for underpasses(overpasses), coherence effects (i.e., the constructive and destructive interference between the crosstalk contributions) are only expected to influence the fine spectral features of the crosstalk and not significantly affect the maximum measured crosstalk across the SCL bands.

The measured overpass and underpass (1.5 µm SiN2 width) crosstalk spectra for TE and TM-polarized inputs are shown in Fig. 8. For the TE polarization, the overpass/underpass maximum crosstalk is −56/−61 dB, and for the TM polarization, the overpass/underpass maximum crosstalk is −46.5/−52 dB. Overall, for a TE-polarized input, the crosstalk is < −56 dB across the measurement wavelength range, while the worst case crosstalk is −46.5 dB for a TM-polarized input. This is, to our knowledge, the lowest crosstalk ever reported for an on-chip optical crossing. To test our coherence assumption, the overpass TE crosstalk measurements were repeated for two longer cutback structures with larger waveguide lengths between the relevant overpasses. Coherence effects are expected to be sensitive to the length between crossings. However, the maximum extracted crosstalk per crossing in the additional measurements was similar and slightly less than that of Fig. 8(a), which supports the assumption.

 

Fig. 8 Extracted overpass and underpass crosstalk per crossing for the (a) TE and (b) TM polarizations.

Download Full Size | PDF

4. Discussion

Table 1 compares the measured crossing and interlayer transition losses in this work with other multilayer and single-layer demonstrations. All quoted values are for the TE polarization. In the comparison, if the measured values were only taken at a single wavelength rather than a range of wavelengths, the wavelength for the data is specified. Our work reported here demonstrates ultra-low and broadband crossing loss and crosstalk across the SCL-bands. This is made possible by the large interlayer separation between the topmost and bottommost waveguide layers in the platform.

Table 1. Performance comparison of waveguide crossings (TE polarization)

View Table

We illustrate the relative performance of our platform architecture by considering, as an example, a waveguide crossing array with 100 crossings and interlayer transitions at the input and output. The total excess loss for such an array is < 0.63 dB across the SCL bands using the measured Si-SiN1-SiN2 transition loss, measured overpass loss, and computed loss of the Si slab tapers in Fig. 3(a). In comparison, the architectures in Table 1 would incur 16.72 dB (at 1.55 µm) [15], 66.4 dB (at 1.55 µm) [14], 100 dB (at 1.55 µm) [13], 7 dB (between 1.52–1.58 µm) [6], 1.9 dB (at 1.55 µm) [9], and < 0.31 dB (at 1.54 µm) [18] of loss in equivalent implementations. The most competitive single layer design, [9], required fine (≈ 50 nm) features and electron beam lithography for fabrication, which limits the practicality of this design for high-volume, high-yield production. The three layer a-Si platform in [18] is promising, but a direct comparison with our work is difficult because the crossing loss was reported at a single wavelength and the waveguide propagation losses have been subtracted from the reported crossing loss, while we specify the maximum crossing loss across a 140 nm bandwidth in the SCL bands including waveguide propagation loss. Low-loss crossings between adjacent layers are also reported in [18], but with higher loss than the three layer crossings. Overall, the multilayer platform shown here is best used with a very large number (hundreds or thousands) of crossings, such that the crossing loss dominates the loss penalty of the interlayer transitions.

5. Conclusion

In summary, we have demonstrated a multilayer SiN on Si photonic platform for low-loss interlayer transitions and ultra-low-loss, low-crosstalk waveguide crossings. The multiple waveguide layers enable flexibility in optical routing and overcome the trade-off between the interlayer transition losses and crossing loss. In the presence of hundreds or thousands of crossings, the loss penalty incurred by the interlayer transitions becomes lower than the crossing loss. Multilayer over/underpass types of crossings are best utilized in very large-scale integrated photonic circuits requiring complex on-chip optical interconnectivity.