Ultra-High-Q Racetrack Resonators on Thick SOI Platform Through Hydrogen Annealing Smoothing

We implemented a hydrogen annealing based post-processing technique as a tool to improve the sidewall roughness of 3 μm thick silicon-on-insulator (SOI) waveguides and demonstrated ultra-high-Q factors on racetrack resonators leveraging on the propagation loss reduction achieved through the smoothing process. The designed racetracks are based on a combination of rib waveguides and strip-waveguide-based Euler bends. We measured intrinsic quality factors of 14 × 106 for a racetrack with a footprint of ∼5.5 mm2 and 10.7 × 106 for a smaller racetrack with footprint of ∼1.48 mm2. The estimated propagation loss for the rib waveguides was ∼2.7 dB/m, representing a ×3 reduction respect to the previously measured losses of 3 μm thick SOI rib waveguides treated with thermal oxidation smoothing. Overall, the post-processing technique allowed to significantly reduce the sidewall roughness without altering the geometry of the waveguides, unlike in sub-micron scale SOI platforms, making it an attractive solution for applications demanding ultra-low losses.


I. INTRODUCTION
S ILICON-ON-INSULATOR (SOI) technology has enabled a myriad of photonics applications leveraging on its CMOS compatibility and high integration density capability thanks to the high index contrast of its waveguides. This very characteristic plays an important role in the relatively high propagation losses of SOI waveguides, since it enhances the interaction of the optical field with the sidewall roughness introduced during fabrication processes such as lithography and etching [1]. This interaction is particularly strong in sub-micron scale waveguides see Fig. 1(a), producing losses of a few dB/cm [2], whereas for thicker Si layers, such as in the 3 μm thick SOI platform, the field has a higher confinement factor due to its larger core size (see Fig. 1), resulting in very low propagation losses [3].
Many emerging applications rely on the filtering and energy storing and enhancing properties of integrated resonators. They are used in lasers for linewidth suppression [4], in nonlinear photonics for enhancement of nonlinear interaction in frequency comb generation [5] and Brillouin lasers [6], in microwave photonics for signal processing [7], and, more recently, in quantum photonics, for communications [8] and metrology [9], among others. All of these applications require very narrow linewidths and low propagation losses, therefore a high quality (Q-) factor. The relation between the cavity losses, including waveguide propagation losses, and the intrinsic Q-factor of a resonator is inverse [10]. Efforts towards demonstrating high Q-factors on SOI technology have mainly revolved around reducing the field interaction with the sidewall roughness to in turn reduce the scattering losses. By increasing the width in sub-micron SOI waveguides, it is possible to reduce the scattering losses, however, by doing so, the waveguides become multimodal, so care must be taken to avoid higher order mode (HOM) excitation. One approach to avoid HOM excitation is to restrict the use of multimode (MM) waveguides to the straight sections of the racetrack and use single-mode (SM) waveguides for the bend sections, which also allows to achieve tight bending radii. Following this configuration, a compact racetrack used a combination of 2 μm wide MM and 500 nm SM ridge waveguides on 220 nm thick SOI to demonstrate a Q-factor of 1.1 × 10 6 and propagation loss down to 0.21 dB/cm [11]. Similarly, a Q-factor of 1.6 × 10 6 was measured for a large-area racetrack on 220 nm thick SOI using 3 μm wide MM rib waveguides and 500 nm wide SM strip waveguides in the coupling and bend sections. The propagation losses of the MM waveguides were as low as 8.5 dB/m, however the Q-factor was reduced by loss contributions from the adiabatic directional coupler, adiabatic bends and strip-to-rib conversion tapers [12], therefore avoiding this transition between different waveguide types could improve the performance. In resonators based only on MM waveguides the bends must be designed following a curve that prevents or minimizes HOMs excitation, since the Q-factor can be degraded in parts of the spectrum where the resonances of the fundamental mode and HOMs are close and the intermodal crosstalk is sufficiently high [13]. Racetracks incorporating so-called Euler bends can provide very low intermodal crosstalk. A racetrack based on 1.6 μm wide MM strip waveguides on 220 nm thick SOI achieved a 2.3 × 10 6 intrinsic Q-factor and estimated waveguide loss of 0.3 dB/cm by implementing Euler bends with an intermodal crosstalk bellow −30 dB between 1500 nm and 1600 nm [14]. A racetrack using a modified version of the Euler bend has also been proposed, showing a Q-factor of 10.2 × 10 6 . The estimated propagation losses for the used 3 μm wide by 220 nm thick strip SOI waveguides was down to 6.5 dB/m. For waveguides this wide, the field interaction with the sidewall roughness is greatly minimized, as it can be seen in Fig. 1(c), however the losses are still limited by the sidewall roughness, which could be improved by optimizing the fabrication methods [15]. A different configuration to avoid HOM excitation used MM concentric racetracks combined with a broadband directional coupler based on a pulley configuration demonstrating a Q-factor of 1.4 × 10 6 over a wavelength range from 1240 to 1680 nm. The two supermodes of the concentric waveguides had slightly detuned resonance wavelengths, which helps counteract the Q-factor deterioration due to coupling to HOMs at the fundamental mode resonances [16]. Another approach to achieve high Q-factors is to smoothen the sidewall roughness through different processing techniques. A combination of surface smoothing through oxidation and resist reflow to improve the etch mask pattern allowed to demonstrate an internal Q-factor of 22 × 10 6 using a ring resonator based on shallow-etched rib waveguides. Besides the improvements obtained through processing, the geometry of the waveguides reduces the field interaction with the sidewalls, translating into losses of 2.7 dB/m, however to limit the contribution from bending losses a large radius of 2.45 mm was needed [17]. In 3 μm thick SOI high intrinsic Q-factors have been demonstrated through the combination of large cross-section waveguides and thermal oxidation (TOX) smoothing post-processing [18]. A Qfactor of 8 × 10 6 was demonstrated for a ring resonator based on SM rib waveguides, with a radius of 2.6 mm (21 mm 2 footprint). A more compact racetrack configuration was achieved by combining the same type of rib waveguides for the straight sections and MM strip waveguides for the bend sections, resulting in an intrinsic Q of 4.3 × 10 6 . In both configurations the Q-factor was limited by the propagation losses of ∼0.1 dB/cm. The TOX process has some drawbacks, such as linewidth reduction, introduction of surface stress [19], and rounding at the inner corners of the waveguides, which can be detrimental to the performance of some devices. Additionally, it is desirable to further reduce the scattering losses of silicon waveguides beyond what has already been demonstrated in this platform.
In this work we implement a hydrogen annealing based post-processing technique to smoothen the sidewall roughness of 3 μm thick SOI waveguides and reduce the propagation losses. This process was originally proposed in the MEMS field [20] and it was later used to smoothen sub-micron scale SOI waveguides [21], enabling losses down to 0.1 dB/cm in the C-band [22]. However, for waveguides at this size scale preserving the shape and dimensions can be challenging [19]. We have previously demonstrated propagation losses of ∼4 dB/m in strip waveguides on thick SOI through this process [23]. Here we exploit this result to demonstrate ultra-high Q resonators. We designed two racetracks with footprints of ∼1.48 mm 2 and ∼5.5 mm 2 [24]. The smaller racetrack had an intrinsic Q-factor of 10.7 × 10 6 and the larger one 14.3 × 10 6 , corresponding to a loss for the rib waveguides of ∼2.7 dB/m. These results are an improvement respect to the ones obtained through TOX in [18], while requiring a much smaller footprint. Additionally, the devices were fabricated in a multi-project wafer (MPW) run, meaning that the process is completely compatible with our current open-access fabrication process, hence it can be integrated into the process flow, enabling new applications that could benefit from the demonstrated ultra-low losses.

II. DESIGN
A schematic of the proposed add-drop racetrack configuration is shown in Fig. 2(a). The straight sections consist of rib waveguides with a width of 3 μm and etch depth of 1.2 μm. These dimensions guarantee a single-mode operation [25] over an ultra-broad wavelength range (1.2-4 μm), along with ultra-low propagation loss [3] thanks to the field being so well confined inside the core see Fig. 1(f)-(g). Rib waveguides have high bending losses due to radiative coupling of HOMs into the slab, therefore requiring a large bending radius for the resonator losses to be dominated by the propagation losses [18]. To circumvent this, we used strip waveguides for the bends. The single mode operation of 3 μm thick Si strip waveguides would require sub-micron widths, resulting in a high aspect ratio which makes waveguides prone to cracking due to stress, and also leads to higher propagation losses, due to the field being less confined inside the core. Increasing the width to a micron scale results in multi-mode operation, therefore an arc bend would require a large radius to avoid HOM excitation, unless a configuration such as the so-called "matched bends" is used. However, this approach does not allow to achieve a very small radius and provides a limited wavelength range of operation [26]. To guarantee that the power of the fundamental mode coupled at the input of the bend is in turn coupled to the fundamental mode of the waveguide at its output, while using a tight bending radius, the bends of the proposed racetrack follow a so-called Euler curve. This type of bend has been successfully demonstrated on VTT's thick SOI platform with very low losses in a wide wavelength range and supporting both TE and TM polarizations [27]. For the coupling sections, we used a directional coupler based on rib waveguides (see Fig. 2(a)) since on thick SOI the strong field confinement prevents efficient enough coupling between strip waveguides (see Fig. 1(d)-(e)).
To determine the optimal bend radii for the 180°Euler (U-) bend, we simulated the transmission for different radii following the procedure detailed in [27] using FIMMPROP. We designed a so-called large racetrack (shown in Fig. 2(b)) with a minimum radius R min = 194.87 μm (effective radius R eff = 268.9 μm) for a transmission of 99.98% in the fundamental mode, and a so-called small racetrack (see Fig. 2(c)) with R min = 97.3 μm (R eff = 134.3 μm) for a transmission of 99.91% in the fundamental mode. For the transition between the straight and bend sections we used rib-to-strip converters (see Fig. 2(a)) to couple the fundamental mode of the SM rib waveguide to the fundamental mode of the MM strip waveguide. The converters are characterized by a length of 200 μm and an insertion loss of ∼0.014 dB per converter [14]. Fig. 3 shows the estimation of the intrinsic Q versus the total length of straight rib waveguide sections of the racetrack, considering the previously mentioned bend radii and converter length. To estimate the intrinsic or loss Q, we need to know the propagation loss of the rib waveguides after hydrogen annealing smoothing. The losses measured for rib waveguides processed with TOX smoothing are 0.1 dB/cm [3], and we have also demonstrated a ×3 reduction of the propagation loss in strip waveguides through hydrogen annealing [23],  therefore we assumed a loss for the ribs of ∼0.03 dB/cm. From these estimations, we chose a total rib waveguide length L rib = 11.6 mm for the large racetrack to obtain an estimated 13.77 × 10 6 intrinsic Q and free spectral range FSR = 5.3 GHz. For the smaller racetrack we chose L rib = 6.6 mm for an intrinsic Q of 8.87 × 10 6 and FSR = 9.1 GHz.
Regarding the external coupling to the cavities, the target was to design weakly coupled racetracks, so that the Q of the resonator was dominated by the propagation losses of the waveguides. However, since the exact propagation losses of the rib waveguides after the H 2 annealing process were unknown, racetracks with identical cavities but different coupling coefficients (different gap between waveguides in the directional coupler) were cascaded, sharing the same input bus (see diagrams in Figs. 8(a) and 9(a)).

III. FABRICATION
The devices were fabricated in an MPW run at VTT. The waveguides were defined on 3 μm thick silicon wafers using stepper lithography (i-line), followed by inductively coupled plasma -reactive ion etching (ICP-RIE) with a Bosch process [28]. This type of etching leads to scallop formation on the sidewall, that adds to the roughness already introduced during lithography (see Fig. 4(a)). Next, the wafer was annealed in  an epi-reactor within 100% pure hydrogen atmosphere, with a pressure of 70 Torr at 900°C for 15 min. These annealing parameters allowed to significantly smoothen the sidewall roughness, as seen in Fig. 4(b) for waveguide-like test structures processed on silicon monitor wafers for optimization of the annealing process. The dimensions and geometry of the waveguides were also preserved, as shown in Fig. 5(a) for rib waveguides (targeted width = 3 μm) and 5.b for strip waveguides (targeted width = 1.875 μm). These results show an improvement over the sidewall smoothing by thermal oxidation, which in contrast affects the linewidth and results in rounding at the inner corners of the rib waveguides, as evidenced in Fig. 6. The final fabrication step was deposition of a 0.5 μm thick TEOS cladding layer on top of the waveguides through LPCVD.

IV. EXPERIMENTAL RESULTS
The measurement setup is shown in Fig. 7. The optical source was the tunable laser TSL-550 by Santec, characterized by a 200 kHz linewidth. For the detection we used a power meter module MPM-211, also by Santec. The Si waveguides support both TE and TM polarizations, therefore careful coupling of the desired single polarization into the chip is required. For this, we used a polarization control stage consisting of a fiber polarization controller, which rotates the polarized light with arbitrary angle at the input into 45 degrees at the output, followed by a switch that routes the input light into one of two outputs. Each of these outputs is connected to a fiber optic polarizing beam splitter, so according to the switch state either TE or TM polarization is obtained at the output. The whole stage has a polarization extinction ratio of ∼20 dB. To couple the light into the chip through edge coupling we used tapered fibers to approximately match the mode size of the Si rib waveguides at the facets of the chip. A polarization maintaining fiber was used at the input, whereas a standard single-mode fiber was used at the output.
The transmission spectra for TM and TE polarization at the through and drop ports are shown in Fig. 8, for the large racetrack, and Fig. 9, for the small racetrack. The measured FSRs were 5.6 GHz and 9.6 GHz respectively. We can observe the presence of off-resonance notches that are stronger than others in the through spectra of both racetracks (pointed out with an encircled 2 and 2/3 in Figs. 8 and 9 respectively), which are not aligned with the drop port resonance peaks. These actually correspond to the resonance of cascaded resonators that share the same through port as the devices under test as explained in section II. We can also see in Fig. 8(b) the presence of off-resonance peaks for the TE polarization. These likely belong to a HOM since the measured rib waveguides were slightly over-etched to a depth of ∼1.4 μm instead of the target depth of 1.2 μm, making the rib waveguide slightly multimodal, capable of supporting a higher order TE mode. However, the extinction ratio between this and the fundamental mode resonances is >20 dB and we did not notice significant induced asymmetry of the fundamental mode resonances. For the large racetracks there were two cascaded racetracks, as shown in Fig. 8(a), whereas for the smaller racetracks there were three of them as shown in Fig. 9(a). Here we show the results for the racetracks with weaker coupling coefficients (larger coupler gap).
The linewidths and Q-factors of the resonance peaks were extracted following temporal coupled-mode theory. According to this, the transmission spectrum of an add-drop resonator is given by the coupled-mode equation solution [29]: where |s drop | 2 is the power at the drop port, |s in | 2 is the power at the input port, γ ext is the external coupling rate, ω 0 is the angular resonance frequency, and γ T is the total linewidth given by: where γ i is the intrinsic linewidth. The loaded Q-factor Q T = ω 0 /γ T is then estimated by performing Lorentzian fit of the resonance peaks at the drop port based on (1), which also allowed to extract the external Q-factor Q ext = ω 0 /γ ext , and therefore the intrinsic or internal Q-factor Q i = ω 0 /γ i from (2). The calculated mean intrinsic Q of the large racetrack between 1560.5 nm and 1561.5 nm for TM was 12.2 × 10 6 , with a maximum value of 14.3 × 10 6 at 1560.61 nm, corresponding to a linewidth of 0.12 pm (14.84 MHz), as shown in Fig. 10(a), whereas for TE polarization the mean Q i was 9.2 × 10 6 , with a maximum of 12.7 × 10 6 at 1561.13 nm, corresponding to a linewidth of 0.14 pm (17.77 MHz), as shown in Fig. 10(b). For the small racetrack the mean intrinsic Q between 1560 nm and 1561 nm for TM polarization was 9.2 × 10 6 , with a maximum value of 10.7 × 10 6 at 1560.58 nm, corresponding to a linewidth of 0.18 pm (21.80 MHz), as shown in Fig. 11(a), and for TE the mean between 1559 nm and 1560 was 7.1 × 10 6 , with a maximum of 8.2 × 10 6 at 1559.26 nm, as shown in Fig. 11(b). The slightly higher Q for TM compared to TE for both racetracks is likely due to the stronger confinement of the optical field for TM, since this leads to less interaction with the sidewall roughness, and therefore, lower scattering losses. The estimated propagation loss extracted from the measured maximum intrinsic Q was down to ∼2.7 dB/m. These results are in good agreement with the estimations from Section II.

V. CONCLUSION
We have experimentally demonstrated ultra-high-Q factors on racetrack resonators fabricated on a thick SOI platform by reducing the propagation losses of the waveguides through an MPW-compatible hydrogen annealing based sidewall roughness smoothing post-processing technique. For a racetrack with an area of ∼5.5 mm 2 we measured an intrinsic Q of 14.3 × 10 6 , whereas for a smaller one with a ∼1.48 mm 2 area we measured 10.7 × 10 6 . The estimated propagation losses extracted from the measured Q-factor was ∼2.7 dB/m. These ultra-low losses will potentially enable the development of emerging PIC applications, such as photonic quantum processors, and various photonic sensors.