Directly Modulated Membrane Lasers on SiC Integrated With Spot-size Converters

Membrane lasers on a SiC substrate offer high-speed direct modulation due to the high thermal dissipation and large optical confinement provided by SiC's high-thermal conductivity and moderate refractive index (∼2.6). A remaining problem is the high coupling loss to a single-mode fiber (SMF) arising from the difficulty in producing a waveguide on SiC substrate whose effective refractive index matches that of the fiber. Here, to overcome this, the SiC substrate is partially removed and the dip is filled with SiO₂, and its surface is planarized to enable the fabrication of membrane lasers with spot-size converters (SSCs) using inversely tapered InP and SiO_x waveguides. The optical mode field is expanded through the inversely tapered InP region and transferred without coupling loss to an SiO_x core that can couple to a high-numerical-aperture SMF with high efficiency. Meanwhile, the laser must be integrated as close to the SiC substrate as possible to increase thermal dissipation. Therefore, the SiO₂ layer thickness under the laser must be less than a few tens of nanometers. A fabricated distributed reflector laser with a 120-μm-long active region exhibits a fiber-coupled output power of 7.3 dBm. The SSC enables us to obtain a fiber-coupling loss of 2.7 dB. Furthermore, we achieve an intrinsic 3-dB bandwidth of 57 GHz thanks to high-thermal-conductivity SiC and a cavity design that suppresses the damping effect. Consequently, we demonstrate 10-km fiber-amplifier-free 224-Gbit/s PAM-4 transmissions even without bandwidth enhancement by the photon-photon-resonance effect.

SECTION I. Introduction

In The face of rapidly growing Internet data traffic due to cloud and artificial intelligence services, it is increasingly important to realize 800 Gigabit and 1.6 Terabit Ethernet with deployments of high-speed and energy-efficient optical transmitters in data centers. For this purpose, optical interconnects supporting 200 Gbit/s per lane have been considered [1]. Directly modulated lasers (DMLs) are key components as compact, energy-efficient, and cost-effective optical transmitters. However, one drawback is that the 3-dB bandwidth f_3dB is at most ∼30 GHz. Thus, they are mainly deployed for short-reach optical interconnects based on 50-GBaud-class four-level pulse-amplitude modulation (PAM-4) [2], [3], [4], [5]. Although electroabsorption modulators and Mach–Zehnder modulators are capable of >100-GBaud operation [6], [7], [8], [9], they require external light sources or laser integration, which increases footprint, power consumption, and cost. Faster operation of DMLs is therefore highly desirable.

The f_3dB of DMLs is determined by the relaxation oscillation frequency f_r, the damping effect, and the product of resistance and capacitance (RC time constant), where f_r of ∼20 GHz is the main factor limiting f_3dB. The f_r can be increased by improving the differential gain, optical confinement factor Γ, and current density above the threshold [10]. Here, increasing Γ is crucial not only for improving f_r, but also for high-modulation efficiency and low power consumption. A membrane laser structure in which an active material is sandwiched by SiO₂ and air can enhance Γ [11]. Thus, we developed membrane lasers on 2-μm-thick SiO₂/Si and achieved an energy cost of 200 fJ/bit with non-return-to-zero (NRZ) 25.8-Gbit/s direct modulation [12], [13]. However, the 2-μm-thick, low-thermal-conductivity (1 W/m/K) SiO₂ resulted in insufficient heat dissipation. This caused a sharp temperature increase at the active region by current injection, leading to differential gain degradation at relatively low bias current density. As a result, the f_r and f_3dB saturated at the same level as in standard DMLs. To remove this limitation, we proposed and fabricated heterogeneously integrated membrane lasers on SiC substrate [14]. SiC, with a high thermal conductivity of 490 W/m/K [15] and a moderately low refractive index of ∼2.6 in the O-band [16], enabled us to reduce the degradation of differential gain at increased bias currents while keeping Γ high. A distributed reflector (DR) laser with a 50-μm-long active region on SiC exhibited f_r of 42 GHz [14]. Furthermore, together with a cavity design to reduce the damping effect by shortening photon lifetime, we obtained f_3dB of 60 GHz [14]. We also demonstrated the use of the optical feedback technique, which can further enhance the bandwidth [14], [17]. We achieved a f_3dB of ∼110 GHz with the photon-photon resonance (PPR) effect generated by optical feedback from the InP waveguide facet and demonstrated 256-Gbit/s PAM-4 at 25 °C [14] and uncooled 100-Gbit/s NRZ modulation up to 85 °C [17].

However, the fiber-coupled output power of the DR laser with the active length of 50 μm was as low as -1.4 dBm due to the large fiber-coupling loss of ∼6 dB [14]. Accordingly, we needed to use a fiber amplifier for the 2-km signal transmissions, which is not desirable from the standpoint of practical use and cost.

To solve this problem, integrating spot-size converters (SSCs) into membrane lasers on SiC is essential. For membrane lasers on 2-μm-thick SiO₂/Si, we have integrated an SSC consisting of an inversely tapered InP and SiO_x core waveguide, in which the optical mode field of the InP output waveguide is adiabatically coupled to the SiO_x core by an inversely tapered InP region [13]. The optical field of the SiO_x core is designed to have a mode-field diameter (MFD) of 3.5–4 μm, which is matched with that of a single-mode high-numerical-aperture fiber (HNAF), allowing us to obtain a high coupling efficiency. On the other hand, integrating SSCs on SiC is difficult because the refractive index of SiC is larger than that of the SiO_x core, and thus the light cannot be confined in the core. The refractive index of SiC is low enough to ensure a high Γ for the active region but too high for SiO_x-based SSCs.

In our recent preliminary work [18], we addressed this issue by removing a given region of the SiC substrate, embedding it with SiO₂, and fabricating SSCs on SiO₂. The SSC enabled a fiber coupling loss of 2.7 dB. To increase the output power, we enlarged the active length to 120 μm from that used in our previous work (50 μm) [14] while keeping the damping effect low. As a result, the fiber-coupled output power reached 7.3 dBm and the intrinsic f_3dB was kept high at 57 GHz at 25 °C. We demonstrated 2-km optical-amplifier-free 100-GBaud PAM-4 transmission even without using the PPR effect [18]. In this paper, we expand upon our recent results in [18] by elaborating on the detailed design and fabrication process of the SSC on SiC and demonstrating a faster transmission of 224-Gbit/s PAM-4 signals with extended single-mode fiber (SMF) transmission up to 10 km.

The rest of the paper is organized as follows. The design of the membrane laser on SiC integrated with the SSC is described in Section II. Section III explains the fabrication process. Section IV presents the experimental results, including the static characteristics and dynamic modulation characteristics. Section V concludes the paper.

SECTION II. Design: SSC-Integrated Membrane Laser on SiC

A schematic of the membrane DR laser on the SiC substrate integrated with the SSC is shown in Fig. 1(a). To form this structure, (i) a part of the SiC substrate has to be removed and the dip filled with SiO₂ to circumvent the refractive index mismatch with the SiO_x-based SSC, (ii) the thickness of the SiO₂ layer under the active region has to be thin enough to maintain high thermal dissipation, and (iii) the surface of the SiO₂ layer has to be flat enough to allow the direct bonding with the InP substrate. Fig. 1(b) shows a cross-sectional scanning electron microscope (SEM) image of the SiC/buried-SiO₂ boundary along the long side of the active region. A 2-μm-deep SiO₂ layer, acting as an undercladding layer, was placed under the InP output waveguide and SiO_x core, which allowed optical confinement in the SiO_x core. The design of SSCs on SiC is almost the same as that of 2-μm thick SiO₂/Si reported in [13], [19], since SSCs on a 2-μm thick SiO₂ layer are not affected by the substrate.

Fig. 1.

(a) Schematic bird's-eye view of the membrane DR laser on a SiC substrate. (b) Cross-sectional SEM image at the SiC and buried SiO₂ boundary along the direction for light propagation (long side of active region).

The key to making the operation speed of the membrane laser faster is to increase the heat dissipation by thinning the intermediate SiO₂ layer between the III-V section and the SiC substrate. In [14], the intermediate 40-nm-thick SiO₂ layer was thin enough to suppress the increase of active region temperature. However, since the SiO₂ layer thickness variation by chemical-mechanical polishing (CMP) process, which will be described later, could affect the thermal property and operation speed within the wafer, we calculate the impact on thermal resistance R_th and f_3dB based on the similar method in [14]. Fig. 2 shows the R_th of the membrane laser on SiC for unit active length (1 μm) as a function of the intermediate SiO₂ layer thickness d_SiO2 (defined in Fig. 1(a)), scaled by the previously reported calculation curve for the active length of 50 μm (see Fig. 2(c) in [14]). In addition, we calculated the small-signal responses (based on a similar method in [14], but further considering the temperature dependence of the gain of InGaAlAs-based MQWs, optimizing the parameters for lasers with 120-μm-long active region, and assuming the mirror loss of 50 cm⁻¹) and evaluated f_3dB as a function of d_SiO2, which is also plotted in Fig. 2. In fact, with our accuracy of CMP, the uniformity of d_SiO2 after CMP is ±10 nm within the wafer. If we set d_SiO2 as 30 ± 10 nm, we found from Fig. 2 that the uniformity causes the thermal resistance of the laser to range from 6.3 × 10³ to 8.4 × 10³ K/W but the f_3dB was kept high at 51–53 GHz. Hence, in this study, the target d_SiO2 was set to around 30 nm.

Fig. 2.

Calculated thermal resistance of the membrane laser on SiC for unit active length of 1 μm (left axis) and 3-dB bandwidth f_3dB (right axis) as a function of intermediate SiO₂ layer thickness d_SiO2. Note that the thermal resistance curve was scaled from the previously reported calculation result for the active length of 50 μm [14]).

To eliminate the need for a fiber amplifier, it is important to increase the output power itself, in addition to reducing the fiber-coupling loss by SSC integration. In our previous works, a laser with an active length of 50 μm was reported, but, in this work, the active length was increased to 120 μm to increase the output power. Therefore, we needed to redesign the grating coupling constant κ to suppress the strong damping effect [14]. In our previous work, we set κ = 600 cm⁻¹ for L = 50 μm (κL = 3), which successfully reduced the damping effect. In this study, we designed κ to be 250 cm⁻¹ for L = 120 μm (κL = 3).

We designed a DR laser consisting of a 120-μm-long active distributed feedback (DFB) section and a 120-μm-long distributed Bragg reflector (DBR) mirror. Single-mode lasing is satisfied by having the DBR mirror select one of the DFB lasing modes. Compared to a λ/4-shifted grating at the DFB section, the uniform grating relaxes localization of the optical power distribution, which is advantageous for high-speed operation [20]. Given the above cavity design, the stopband width of DFB and DBR was designed to be ∼5 nm.

SECTION III. Fabrication Process

The fabrication process of the SSC-integrated membrane laser on SiC is shown in Fig. 3(a)–​(l); the cross sections of the active region and SSC along the direction perpendicular to the light propagation are illustrated. First, we prepared a 2-inch SiC substrate. After 2-μm dry etching of a given area where the SSC will be fabricated [Fig. 3(a)], we deposited SiO₂ to fill the dip [Fig. 3(b)]. We then polished the surface by CMP to thin and flatten the SiO₂ layer [Fig. 3(c)]. To achieve high thermal dissipation at the active region, we carefully carried out the CMP process so that the rest of the SiO₂ layer thickness on the unremoved SiC area to be around 30 nm. Note that the SiO₂ layer thickness was estimated by spectroscopic reflectometry. Meanwhile, 9-period InGaAlAs MQWs and an InGaAs etch stop layer were grown on a 2-inch InP (001) substrate using metal-organic vapor phase epitaxy (MOVPE). After a 5-nm SiO₂ layer had been deposited on the InP substrate, the InP and SiC substrates were directly bonded using a O₂-plasma assisted-bonding technique [Fig. 3(d)]. Note that we had the SiO₂ thickness be as thin as possible and defined as 5 nm, where we defined the same thickness in our previous work aimed at improving the yield [17]. As shown in Fig. 2, the 5-nm increase in thickness does not largely affect the f_3dB degradation. In fact, the total intermediate SiO₂ layer thickness d_SiO2 was 30 nm, which was confirmed by a cross-sectional transmission electron microscope image. The InP substrate was removed by mechanical polishing and wet etching using the etch stop layer [Fig. 3(e)]. An optical microscope image of the InP/SiC template taken after this process is shown in Fig. 4. There are no significant voids, which means that the surface polished by CMP was smooth enough and the bonding process was successfully performed even though the SiC substrate included the region buried with SiO₂. Next, we deposited and etched a SiO₂ mask, followed by a mesa-stripe formation by dry and wet etching [Fig. 3(f)]. The mesa was buried in an intrinsic InP layer by epitaxial regrowth using MOVPE. Then, Si-ion implantation and Zn thermal diffusion were performed to form the lateral p-i-n junction [Fig. 3(g)]. The thickness of the InP slab was ∼350 nm. Next, after surface gratings had been formed on the top InP region to define the DFB and DBR sections, the InP membrane layer was selectively removed [Fig. 3(h)]. The InP waveguide at the tapered region is illustrated in Fig. 3(h). Contact metals were deposited on the n-InP and p-InP [Fig. 3(i)]. We then deposited the SiO_x layer [Fig. 3(j)]. By partially removing the SiO_x layer by reactive ion etching, the SiO_x core was defined [Fig. 3(k)]. Finally, a 5-μm SiO₂ overcladding layer was formed on the SSC, followed by opening the contact holes for electrodes [Fig. 3(l)].

Fig. 3.

Fabrication process for the membrane laser on a SiC substrate integrated with the SSC. Cross-sectional views are shown in the direction perpendicular to the light propagation for the active region (right) and SSC section (left). (a) Etching of SiC, (b) deposition of SiO₂, (c) CMP, (d) wafer bonding, (e) removal of InP substrate, (f) dry and wet etching, (g) n- and p-doping, (h) grating formation and selective removal of InP, (i) deposition of contact metal, (j) deposition of SiO_x, (k) formation of SiO_x core, and (l) opening of contact holes.

Fig. 4.

Optical microscope image of the 2-inch InP/SiC template.

SECTION IV. Results and Discussion

The measured output light versus bias current (L-I curves) at 25 °C are shown in Fig. 5, with one curve (red) for detecting output power using a photodetector (PD) placed in front of the output facet and the other (blue) for detecting output power using a single-mode HNAF butt-coupled with the SiO_x waveguide facet. We performed all the measurements at a stage-controlled temperature of 25 °C. The threshold current was 2.7 mA. The undulation at ∼30 mA in the L-I curve measured using the PD is due to the reflection at the SiO_x facet. On the other hand, when the HNAF was butt-coupled to the SiO_x facet, the reflection at the edge was suppressed, and thus the undulation was not so significant; the slightly remaining undulation could originate from very little reflection at the output facet. The maximum output power detected using the PD was 10.3 mW (10.1 dBm) owing to the enlarged active length of 120 μm. The maximum fiber-coupled output power was 5.4 mW (7.3 dBm). Thus, the fiber-coupling loss, defined as the difference between the output power directly detected using the PD (i.e., almost all output power from the front facet) and that coupled to the HNAF, was 2.7 dB. This value is almost identical to those obtained for SSC-integrated membrane lasers on SiO₂/Si [13], [21]. Increased output power at the high bias current range also indicates that the self-heating effect was suppressed. This fact also evidences high thermal dissipation at the active region due to the thin SiO₂ intermediate layer as is confirmed in Fig. 1(b).

Fig. 5.

L-I curves at 25 °C. The power directly detected using the large-diameter PD (red) and the power coupled to the HNAF (blue) are shown.

The lasing spectrum at 80 mA is shown in Fig. 6, where the inset shows the lasing wavelength as a function of the input electrical power. The side mode of the DFB stopband at shorter wavelength was selected, and the single-mode lasing was thus achieved with the side-mode suppression ratio of 55 dB. The actual κ was estimated from the spectral fitting to be 250 cm⁻¹, which is as designed. No significant reflection peak was observed in the DFB stopband, which means that the reflection (e.g., possibly occurring at the boundary between the buried 2-μm-thick SiO₂/SiC, the InP tapered edge, and the output SiO_x facet) was sufficiently small. Note that, if there is significant to moderate reflection at the output edge, Fabry-Pérot modes appear in the DFB stopband [14], in contrast to the spectrum in Fig. 6. From the inset in Fig. 6, the lasing wavelength was continuously changed, which indicates that there is no mode hopping. Next, we evaluate the thermal resistance R_th of the fabricated laser with the intermediate SiO₂ layer thicknesses d_SiO2 of 30 nm, comparing the values of previously fabricated lasers (without the CMP process and SSC) with d_SiO2 of 10 [17] and 40 nm [14]. R_th can be derived from the experimental results, relying on R_th = dT/dP = (dλ/dP)/(dλ/dT), where dT, dP, and dλ are changes in temperature, input electrical power, and lasing wavelength, respectively. Here, since the active length L in this work differs from that in previous works [14], [17], we evaluate R_th for unit active length (1 μm). In our InGaAlAs-based MQW laser, dλ/dT is typically 0.09 nm/K [14]. Based on this and dλ/dP obtained from the slope from inset of Fig. 6 (8.1 × 10⁻³ nm/mW), the R_th for unit active length for the present device with d_SiO2 = 30 nm and L = 120 μm was evaluated to be 10.8 × 10³ K/W, whereas the previously fabricated devices (L = 80 μm) for d_SiO2 of 10 and 40 nm were evaluated as 7.0 × 10³ and 11.6 × 10³ K/W, respectively. Thus, the R_th of 10.8 × 10³ K/W for d_SiO2 of 30 nm in this work is reasonable.

Fig. 6.

Lasing spectrum at 80 mA. The inset shows the lasing wavelength as a function of the input electrical power.

Fig. 7 shows the small-signal response at 80 mA, measured with a light component analyzer (Keysight N4373D). We achieved a f_3dB of 57 GHz for the DR laser with enlarged active length of 120 μm. This value is almost identical to that for the active length of 50 μm [14]. This means that we were able to maintain high f_r thanks to the high-thermal dissipation with the ultrathin SiO₂ layer beneath the active region and reduce the damping effect with the cavity redesign. Furthermore, the large bandwidth suggests that the RC time constant does not limit the bandwidth at least up to ∼60 GHz, even with the enlarged active length. This is due to the small pad-to-electrode capacitance of the lateral p-i-n structure. Consequently, we were able to increase the fiber-coupled output power while maintaining a large bandwidth capable of >100-GBaud modulation.

Fig. 7.

Small-signal (S₂₁) response at 80 mA. The dashed line shows the S₂₁ response of -3 dB.

Finally, we performed high-speed transmission testing under a fiber-amplifier-free condition. The experimental setup for evaluating the bit-error rate (BER) is shown in Fig. 8. Radio-frequency (RF) signals up to 112-GBaud PAM-4 were generated using an arbitrary waveform generator (AWG) (Keysight M8199A) at 200 GSa/s with an analog f_3dB of 70 GHz. An external 66-GHz amplifier with an 11-dB gain (SHF M827B) was used to amplify the signal from the AWG. The signal was combined with bias current from a direct-current (DC) source using a 65-GHz bias tee and injected into a device under test (DUT) through a 67-GHz RF probe via a short RF (<30-cm) cable. After passing through an optical isolator and transmitting through a standard-single-mode fiber (SSMF), the optical light was detected with an in-house uni-traveling-carrier (UTC) PD module with a f_3dB of ∼90 GHz. Note that a part of optical light was fed into an optical power meter using a 13-dB coupler. The received optical power (ROP) detected by the UTC-PD was controlled using a variable optical attenuator (VOA). Due to the lack of a trans-impedance amplifier (TIA), the electrical signal after the UTC-PD was amplified with an external electrical amplifier (SHF M827B) and then captured using a real-time digital storage oscilloscope (DSO) with a sampling rate of 256 GSa/s and bandwidth of 110 GHz. The captured signals were processed offline.

Fig. 8.

Experimental setup for evaluating BERs.

The BER versus ROP for back-to-back (BTB) and 2-km SSMF transmissions of 100-GBaud PAM-4 signals are shown in Fig. 9(a). A 9-tap linear equalizer was used to compensate for distortions. The equalized optical eye diagrams for BTB and after 2-km transmission are shown in Fig. 9(b) and (c), respectively. The bias current, bias voltage, and peak-to-peak voltage were 80 mA, 3.04 V, and 2.77 V, respectively. The BERs were less than the 6.7%-overhead hard-decision forward error correction (HD-FEC) threshold of 3.8×10⁻³ [22]. We observed lower BERs after the 2-km transmission than for the BTB, which was due to the negative dispersion of the SSMF at the lasing wavelength of ∼1292 nm and was also observed in our previous work [14]. The energy cost was calculated to be 1.3 pJ/bit. Next, we further increased the symbol rate and transmission reach. The BERs versus ROP for BTB, 2-km, and 10-km SSMF transmissions of 112-GBaud PAM-4 signals are shown in Fig. 10(a); 21-tap linear and 21-tap nonlinear equalizers were used to compensate for the linear and nonlinear distortions according to the digital signal processing in [23]. The bias current, bias voltage, and peak-to-peak voltage were the same as those used for 100-GBaud PAM-4 measurements. We achieved BERs lower than the HD-FEC threshold even after the 10-km SSMF transmission. Fig. 10(b) and (c) respectively show the equalized optical eye diagrams for BTB and after 10-km transmission, which exhibit clear eye openings. The net data rate and energy cost were 210 Gbit/s and 1.2 pJ/bit, respectively. The previous record on the modulation speed for DMLs without the PPR effect was 56-GBaud PAM-4, achieved using InGaAlAs lasers with a ridge-shaped buried-heterostructure fabricated on a InP substrate [24]. We demonstrated >100-GBaud modulation using the DML without the PPR effect for the first time. The keys to achieving this include (i) the improved fiber-coupled output power made possible by the novel method of integrating SSCs on SiC, which increases the signal-to-noise ratio; (ii) the sufficiently large bandwidth achieved by the high heat dissipation with ultrathin SiO₂ on SiC; and (iii) the cavity design that keeps damping effects low for the 120-μm-long active region.

Fig. 9.

(a) BER vs. ROP for the 100-GBaud PAM-4 signal transmission. 100-GBaud PAM-4 equalized optical eye diagrams (b) for BTB and (c) after 2-km SSMF transmission.

Fig. 10.

(a) BER vs. ROP for the 112-GBaud PAM-4 signal transmission. 112-GBaud PAM-4 equalized optical eye diagrams (b) for BTB and (c) after 10-km SSMF transmission.

Aiming at larger throughput by increasing the bandwidth, the next step is to add optical feedback and then induce the PPR effect for the SSC-integrated membrane laser on SiC. This can be done by applying an optical reflection coating on the SiO_x core or by integrating a DBR mirror in front of the DFB section [25], [26], both of which are under investigation.

SECTION V. Conclusion

We have integrated SiO_x-based SSCs for membrane lasers on a SiC substrate to improve fiber-coupling efficiency. The SSCs were successfully fabricated on a partial 2-μm-thick SiO₂/SiC region, where a part of the SiC substrate was removed, filled with SiO₂, and flattened by CMP processes. On the other hand, the membrane lasers were fabricated on an ultrathin SiO₂/SiC region (non-removal region of SiC) to ensure the high-thermal dissipation. As a result, we succeeded in decreasing the fiber-coupling loss to 2.7 dB and obtaining the fiber-coupled output power of 7.3 dBm for a DR laser with a 120-μm-long active region. Thanks to the cavity design that reduces the damping effect for the enlarged 120-μm-long active region, we were able to maintain a high 3-dB bandwidth of 57 GHz. We demonstrated fiber-amplifier-free 112-GBaud PAM-4 signal transmissions over 10 km, even without the help of the PPR-based bandwidth enhancement. Consequently, the DML capable of 200-Gbit/s per lane is promising for 800 Gigabit Ethernet and beyond.

ACKNOWLEDGMENT

The authors would like to thank Y. Shouji, Y. Yokoyama, Y. Sayama, M. Hosoya, K. Ishibashi, and J. Asaoka for assistance with device fabrication. The authors would also like to thank S. Hirata for assistance with measurements.