1. INTRODUCTION

There are strong demands to reduce the cost, footprint, and power consumption of the photonic components that route data traffic between and within data centers. These demands can be met by integrating various devices on a single chip in the form of photonic integrated circuits (PICs). In particular, transmitters based on PICs can reduce costs compared with transmitters based on discrete devices, because the cost of a transmitter is mainly determined by how it is assembled [1]. From this perspective, PICs based on silicon photonics (SiPh) are promising solutions. This is because, with SiPh technology, we can use mature complementary metal-oxide-semiconductor (CMOS) fabrication technology [2–5] to build compact and low-loss optical components such as optical filters, switches, and spot-size converter (SSCs). However, since silicon is an indirect-bandgap material, heterogeneous integration of III-V compound semiconductors is required to fabricate lasers and high-efficiency modulators with SiPh-based PICs. In particular, heterogeneous integration of directly modulated lasers (DMLs) is very important because DMLs provide low power consumption, low cost, and a small footprint.

Accordingly, direct epitaxial growth on silicon has attracted much attention [6,7], but fundamental difficulties remain, such as lattice mismatch, differences in the coefficient of thermal expansion (CTE), and the formation of antiphase boundaries (APBs). Recently, III-V quantum dot (QD) lasers directly grown on silicon with optical output power higher than 100 mW have been demonstrated because QDs are less sensitive to threading dislocations [8,9]. However, a buffer layer thicker than 2 µm is still needed, which makes them difficult to optically connect with SiPh devices. In contrast, recent direct growth techniques using selective growth masks as dislocation filters allow for the monolithic integration of dislocation-free III-V layers on silicon without the need for a thick buffer layer [10–14]. Although a variety of optically pumped lasers have been demonstrated based on this scheme, electrically pumped lasers with sufficient output power for data-center applications have yet to be reported. This seems to be correlated to the difficulty of both growing ternary or quaternary compound semiconductor material and fabricating a p-i-n junction under the structural limitations due to the selective masks.

Another possible solution is wafer-scale heterogeneous integration using direct- and adhesive-bonding techniques [15,16]. These technologies have been used to develop heterogeneously integrated lasers in which the III-V active region is optically connected to silicon waveguides [17–19]. In addition, micro-transfer-printing that allows for manipulation of micrometer-size III-V layer coupons has attracted attention as an effective way to integrate III-V optical devices on silicon-photonics components [20,21]. As a bias light for SiPh modulators, these lasers provide sufficient output power with a narrow linewidth. On the other hand, the power consumption and footprint of DMLs on silicon are still large compared with the vertical-cavity surface-emitting lasers (VCSEL) widely used in data-center applications [22–25]. To reduce the power consumption, it is important to reduce the cavity length because the power consumption of a DML is proportional to the active-region volume. However, miniaturizing a laser requires large optical confinement in the active region and a large grating-index coupling coefficient, which are difficult to achieve in conventional heterogeneously integrated lasers.

In this context, we have proposed and demonstrated membrane buried-heterostructure (BH) DMLs on silicon. An InP-based membrane BH sandwiched between low-refractive-index materials enables us to increase the coupling coefficient, which makes it possible to provide a small mirror loss with a short cavity. This structure also provides a large optical confinement factor in the active region. Therefore, a reduction in active region volume that provides high modulation efficiency, low energy consumption, and small footprint can be achieved. The BH also improves the thermal impedance because the thermal conductivity of InP is 10 times larger than that of quaternary compound semiconductor materials [26]. By reducing the cavity length to 75 µm, we have achieved 25.8 Gbit/s non-return-to-zero (NRZ) signal modulations with energy cost of ${\sim}{100}\;{\rm fJ/bit}$ [27]. To fabricate these lasers, the multi-quantum-well (MQW) layer sandwiched between InP layers is directly bonded to a silicon substrate with thermal oxide (${{\rm SiO}_2}/{\rm Si}$ substrate). Then the MQW layer is removed down to the InP layer, except for the laser active region. An InP layer is regrown on the remaining InP layer to fabricate the BH [27–31]. The key to achieving the regrowth after the bonding is to keep the total III-V thickness less than the critical thickness, ${\sim}{430}\;{\rm nm}$, which is calculated based on the difference in the CTE between silicon and InP [31].

Recently, however, PICs need to support high modulation speeds in uncooled or semi-cooled operations and wavelength division multiplexing (WDM) with a wavelength range of 35.6 or 60 nm [32]. To meet these requirements, the bandgap of the MQW layer used as the active region for DMLs has to be optimized for each channel in the transmitter array. Although bonding approaches allow integrating different III-V layers on silicon, the bonding of multiple III-V dice with different bandgaps is required, which increases both fabrication and assembly costs.

To overcome this drawback, III-V compound semiconductors with different bandgaps should be grown simultaneously in a single epitaxy, for which selective area growth is a promising solution. In this paper, we describe the fabrication and the characterization of a heterogeneously integrated multiwavelength laser array fabricated by using selective area growth of InGaAlAs MQWs on a thin InP layer directly bonded to a ${{\rm SiO}_2}/{\rm Si}$ substrate (InP-on-insulator substrate). We first confirmed the crystal quality of the InGaAlAs MQWs grown on the InP-on-insulator substrate without the selective area growth technique. Since quaternary compound semiconductors are sensitive to strain compared to InP, we carefully control the residual strain in the active regions to be around zero at the growth temperature [33]. The photoluminescence (PL) and x-ray diffraction (XRD) characteristics of MQW are comparable to those for the same structure grown on an InP substrate. Then, we demonstrate selective area growth of InGaAlAs MQWs, where we are able to control the PL peak wavelength to fabricate laser over ${\gt}$100 nm with no significant degradation by changing the space and width of the ${{\rm SiO}_2}$ mask. We fabricated eight-wavelength membrane distributed feedback (DFB) lasers with lasing wavelengths ranging from 1272.3 to 1310.5 nm. All the lasers exhibited 25.8-Gbit/s direct modulation with NRZ signal. We also demonstrate 28-Gbaud pulse amplitude modulation with four-amplitude-level (PAM-4) operations. These results indicate the applicability of the epitaxial growth using an InP-on-insulator platform as a fabrication technology of III-V/Si PICs for the future PIC-based transmitters.

 

Fig. 1. Epitaxial growth of InGaAlAs MQWs using InP-on-insulator substrate. (a) Schematic of the epitaxial growth procedure using InP-on-insulator substrate fabricated by wafer bonding of InP epitaxial wafer to thermally oxidized silicon substrate, together with the growth progression. (b) Photoluminescence (PL) intensity maps for 12-period MQWs grown on InP substrate (top) and on InP-on-insulator substrate (bottom). The residual strain varied from ${-}{1290}$ to 670 ppm.

Download Full Size | PDF

2. EPITAXIAL GROWTH ON AN InP-ON-INSULATOR SUBSTRATE

First, we performed the epitaxial growth of InGaAlAs MQWs on an InP-on-insulator substrate to confirm their crystal quality. Figure 1(a) schematically depicts the epitaxial growth procedure. The fabrication of the InP-on-insulator substrate began with an epitaxial growth of an InP buffer layer, an InGaAs etch stop layer (100 nm), and an InP layer (50 nm) on a 50 mm standard InP (001) substrate. The epitaxial InP wafer and a 50 mm silicon (001) wafer with a 2 µm thick thermal ${{\rm SiO}_2}$ layer were treated with oxygen plasma and were directly bonded to each other at room temperature under atmospheric pressure. The post-annealing was carried out at 200°C while pressure of 1.0 MPa was applied. The InP substrate was mechanically polished and chemically etched using an InGaAs etch stop layer to leave the 50 nm thick InP layer on the ${{\rm SiO}_2}$. A typical root mean square (RMS) value of the InP surface is less than 0.5 nm [31]. Then we grew an InP buffer layer, an InGaAlAs MQW layer, and an InP cap layer using metal-organic vapor phase epitaxy (MOVPE). The detailed epitaxial growth progression is also described in Fig. 1(a). The growth temperature was around 630°C, and the pressure was 6.7 kPa under hydrogen carrier gas. The V/III ratio was ${\sim}{200}$, and the growth rate was ${\sim}{800}\;{\rm nm/h}$. The precursors for indium, gallium, aluminium, arsenide, and phosphide were trimethylindium (TMI), triethylgallium (TEG), trimethylaluminium (TMA), arsine (${\rm AsH}_3$), and phosphine (${\rm PH}_3$), respectively.

Since the CTE of InP and silicon are different, thermally induced strain is applied to III-V layers during the growth. To prevent the formation of dislocations, the total III-V layer thickness should be kept less than the critical thickness determined by the thermally induced strain [28,31]. The thermally induced strain is given by the product of the differences in CTE between the III-V layer and substrate and the differences in growth and bonding temperatures. Here, for simplification, the CTE of InP and silicon are considered. The thermally induced strain applied to InP is estimated to be approximately 830 ppm (compressive) at growth temperature and approximately ${-}{340}\;{\rm ppm}$ (tensile) at room temperature, when bonding and growth temperatures are set to 200ºC and 630ºC, respectively. Accordingly, the critical thickness is calculated to be ${\sim}{410}\;{\rm nm}$ [31]. Another key step in growing a high-quality active region is to compensate for the compressive thermally induced strain of ${\sim}{830}\;{\rm ppm}$ by controlling the strain in the active region. We control residual strain in MQWs by increasing the tensile strain in the quantum barrier layer, whereas the quantum wells (QWs) have compressive strain of ${\sim}{1.2}\%$ to obtain high material gain, which is almost the same QW condition as in our previous DMLs [27–30].

To evaluate how the residual strain affects the crystal quality, we grew an InP buffer layer (120 nm), 12-period QWs (${\sim}{200}\;{\rm nm}$) with various residual strains, and an InP cap layer (30 nm) on an InP-on-insulator substrate and InP substrate. We doubled the number of QWs compared to that of our conventional laser structure (typically six periods [33]) to promote the formation of dislocations. Figure 1(b) shows the microscale PL intensity maps with different residual strain. Here we used a PL microscope with focused laser excitation at a wavelength of 1064 nm. The strains in the MQWs were characterized by XRD. The spots in Fig. 1(b) indicate the presence of threading dislocations, and we did not observe a dependence relationship between residual strain and threading dislocation density (TDD) for the MQWs grown on the InP substrate. On the other hand, for the MQWs on the InP-on-insulator substrate, the TDD significantly increased when a compressive strain was induced, whereas it was almost the same when tensile strain between ${-}{1290}$ and ${-}{240}\;{\rm ppm}$ was induced. This indicates that the residual tensile strain successfully compensated for the thermally induced strain in the MQWs. The estimated TDD (${5.2} \times {{10}^3}/{{\rm cm}^2}$) for MQWs with tensile residual strain is comparable to that of the etch-pit density (EPD) of commercially available InP substrate. This clearly demonstrates that the residual strain in MQW is a critical parameter for obtaining a high-quality active region on an InP-on-insulator substrate.

Based on the above results, we grew a six-period InGaAlAs MQW with tensile residual strain on an InP-on-insulator substrate and on an InP substrate as a reference. Figures 2(a) and 2(b) compare the crystal quality of the two wafers with respect to their PL and XRD, respectively. The peak PL intensity was 7.2 times greater than that of the reference because both the excitation and collection efficiency were enhanced due to the underlying ${\rm InP}/{{\rm SiO}_2}/{\rm Si}$ interfaces. The slight spectral broadening from 30.0 to 33.4 meV seems to be due to the enhancement of excitation energy density, and the spectral shapes and peak PL wavelength are almost the same [33]. The steepness of the XRD peaks, which represents both the uniformity and periodicity of MQWs, are also almost the same and well agreed with the numerical simulation. The evaluated residual strain in the MQW is ${-}{680}\;{\rm ppm}$ (tensile). Figure 2(c) shows the surface morphology measured with an atomic-force microscope (AFM). The winding step-terrace structure indicates the strain distribution due to the direct bonding, but no dislocations are observed. The PL intensity map shown in Fig. 2(d) also shows that there are neither dark spots nor dark lines over the area of ${500} \times {500}\;\unicode{x00B5}{\rm m}^2$, corresponding to the TDD of less than ${4.0} \times {{10}^2}/{{\rm cm}^2}$. These results clearly demonstrate the superior quality of the MQW, even though it was grown on an InP-on-insulator substrate.

 

Fig. 2. Characterization of epitaxially grown InGaAlAs MQWs on InP-on-insulator substrate. (a) PL spectra for six-period MQWs grown on InP-on-insulator substrate (solid blue line), InP substrate (dashed red line), and normalized PL spectrum for the MQW grown on InP substrate (dashed green line). (b) X-ray diffraction (XRD) characteristics for six-period MQWs grown on InP-on-insulator substrate and InP substrate, along with numerical simulation. (c),(d) Atomic force microscope (AFM) image and PL intensity map for six-period MQWs with InP-cap layer grown on InP-on-insulator substrate, respectively.

Download Full Size | PDF

 

Fig. 3. Selective-area epitaxial growth on InP-on-insulator substrate. (a) Schematic drawing of selective epitaxy of InGaAlAs MQWs with different bandgaps. (b) Schematic of the top view of the selective-area growth mask. (c) PL peak wavelength shift caused by the selective mask. (d) PL spectra for MQWs with different mask widths.

Download Full Size | PDF

3. FABRICATION OF WIDE-WAVELENGTH-RANGE MEMBRANE LASERS ON SILICON

A. Selective Epitaxial Growth of MQWs

To support WDM systems, each laser in the array should be made of an MQW that has an optimized PL peak wavelength. This makes it difficult to fabricate PICs because, as long as we use bonding approaches, the bonding of multiple III-V dice with different bandgaps is required. Therefore, a selective-area epitaxial growth technique using InP-on-insulator substrate, which makes it possible to grow MQWs with different bandgaps, is desirable for integrating multiwavelength DMLs on silicon.

Figure 3(a) shows a schematic drawing of the selective-area growth of multiwavelength III-V layers, and Fig. 3(b) shows a schematic of the top view of a set of selective growth mask. Due to the very high selectivity of the growth rate between the semiconductor surface and ${{\rm SiO}_2}$ mask, the group-III species take additional lateral source supply paths (surface migration and vapor phase diffusion) from the mask to the semiconductor surface. This causes a change in the growth rate as well as the crystal composition depending on the mask structure. Therefore, the growth of multiwavelength MQWs with a single growth becomes possible [34–36]. On the other hand, the selective growth of multiwavelength MQWs using an InP-on-insulator is a challenge because the residual strain also changes depending on the mask design, whereas the residual strain should be controlled to be tensile for all MQWs. The vapor phase diffusion length of indium is much shorter than that of gallium and aluminium [36], so the MQW with a wider selective mask should have a higher indium content, resulting in additional compressive strain. To compensate for this, we set the composition of group-III species to provide a tensile residual strain of ${-}{2400}\;{\rm ppm}$ in MQW when a mask is not used. The epitaxial layer consists of an InP buffer layer (${\sim}{90}\;{\rm nm}$), an MQW layer (${\sim}{85}\;{\rm nm}$), and an InP cap layer (${\sim}{25}\;{\rm nm}$). Note that the residual strain and layer thicknesses vary depending on the selective mask parameters. All other growth conditions were the same as those for the non-selective epitaxial growth on InP-on-insulator substrate. We selectively grew InGaAlAs MQWs on an InP-on-insulator substrate to evaluate the PL wavelength controllability. The epitaxial layers are thicker in the vicinity of the mask, but we found that a ${\sim}{3}\;\unicode{x00B5}{\rm m}$ wide flat surface is obtained at the center of the masks. Therefore, we used the PL microscope with a focused spot size of ${\sim}{1}\;\unicode{x00B5}{\rm m}$ to evaluate the peak wavelength. Figure 3(c) shows the PL peak wavelength shift for the MQWs grown on the InP-on-insulator substrate. Here mask width ${w_m}$ ranged from 5 to 50 µm, and the distance between the masks ${w_g}$ ranged from 10 to 120 µm. The deposition of InP polycrystal on the ${{\rm SiO}_2}$ masks was not observed in any mask sizes. The PL wavelength for the reference MQW without a mask was 1206.0 nm. As shown in Fig. 3 (c), we found clear dependence of the PL wavelength shift on mask parameters. A further evaluation is shown in Fig. 3(d), where the PL spectra for MQWs with different masks are compared. Here $w_m$ was changed from 0 to 50 µm, while ${w_g}$ was kept constant at 20 µm. The distortions in the PL spectra for ${w_{m}} = {50}\;\unicode{x00B5}{\rm m}$ are attributed to the ${{\rm H}_2}{\rm O}$ atmospheric absorption peaks at around 1380 nm. We observed a monotonic increase in the PL intensity as the PL wavelength changed from 1206.0 nm (${w_m} = {0}\;\unicode{x00B5}{\rm m}$) to 1316.5 nm (${w_m} = {30}\;\unicode{x00B5}{\rm m}$). A possible explanation for this is that the thicker QWs resulted in a more efficient absorption of the excitation light. In addition, the thicker and more compressive QWs for the larger masks also cause crystal degradation when the QW becomes thicker than the critical thickness. This seems to be the reason for the decrease in PL intensity of ${w_m} = {50}\;\unicode{x00B5}{\rm m}$. Nevertheless, we achieved a PL wavelength shift of more than 100 nm with a single epitaxial growth using selective epitaxy on InP-on-insulator substrate.

 

Fig. 4. Schematic of fabrication procedure for selectively grown wide-wavelength-range membrane laser array on silicon. (a) Direct bonding of InP epitaxial wafer to silicon substrate with 2 µm thick thermal oxide. (b) Formation of ${{\rm SiO}_2}$ mask on InP-on-insulator substrate for selective-area growth. The curvy gap represents the wide separation of two channels. (c) Selective-area epitaxial growth of InGaAlAs MQW on InP-on-insulator substrate. (d) Formation of mesa stripes by dry and wet etching. (e) Selective-area regrowth of InP to fabricate BH. (f) $P$- and $n$-type doping. (g) Fabrication of surface grating and metallization. (h) Integration with spot-size convertor.

Download Full Size | PDF

 

Fig. 5. (a) Photograph and (b) PL intensity map for the entire50 mm wafer after the selective epitaxial growth of six-period InGaAlAs MQWs. (c) PL characterization of MQWs for eight-channel membrane lasers compared with reference MQWs grown on InP substrate. The inset compares the PL peak wavelength and FWHM.

Download Full Size | PDF

B. Laser Fabrication

On the basis of the selective epitaxial growth results, we fabricated eight-channel membrane lasers with eight-wavelength MQWs, which were designed for an eight-channel local area network (LAN) WDM system configuration based on 100/400 Gbit Ethernet [32]. Thus, the channel spacing was set to 800 GHz to demonstrate the lasing wavelength controllability.

Figure 4 shows the fabrication procedure for the membrane laser array. First, a 50 mm InP-on-insulator substrate was prepared by using the direct-bonding method. Next, eight-wavelength MQWs were selectively grown by MOVPE. The crystal composition of the MQW layer and the layer structures are almost identical to the ones mentioned in Section 3.A. To control the PL wavelength for the LAN-WDM application, the mask width ranged from 7 to 17 µm and the distance between the masks was fixed at 20 µm, while the distance between lasers (laser pitch) was 250 µm. After the selective area growth, a ${{\rm SiO}_2}$ layer was deposited by using plasma-enhanced chemical vapor deposition (PECVD). The ${{\rm SiO}_2}$ mask pattern was formed by using electron-beam lithography (EBL) and reactive ion etching (RIE). Then, eight mesa stripes with different-wavelength active regions were formed by dry and wet etching. After that, the mesa stripes were simultaneously buried with undoped InP to fabricate BHs. The growth temperature was around 600°C, and the pressure was about 4.0 kPa. Lateral p-i-n structures were formed by thermal diffusion of zinc ($p$ doping) and ion implantation of silicon ($n$ doping). Then, surface gratings were fabricated by etching the InP cap layer just above the InGaAlAs/InP BHs. After that, the butt-joint InP waveguides with laterally tapered edges were formed by RIE. After the metallization, the ${{\rm SiO}_x}$ layer was deposited by electron-cyclotron-resonance (ECR) PECVD. The ${{\rm SiO}_x}$ layer was functionalized as not only the core of the SSC but also the cladding of the III-V layer. After the core structures on the tapered InP waveguides had been defined by partial etching of ${{\rm SiO}_x}$ layer with RIE, a ${{\rm SiO}_2}$ cladding layer was deposited by PECVD. Finally, contact holes were formed above the electrodes to electrically drive the lasers.

Figures 5(a) and 5(b) show a photograph and PL-intensity map for the entire wafer after the selective growth of InGaAlAs MQWs, respectively. The PL intensity was uniform except for at the wafer edge, and there were no significant voids even after the epitaxial growth. Figure 5(c) shows PL spectra for each channel, and the inset shows wavelength and FWHM of each PL. We obtained multiple active regions with PL peak wavelengths ranging from 1238 to 1276 nm with FWHMs ranging from 41.7 to 44.3 meV. Despite the different growth rate and the change in compositions due to the selective mask, the broadening of FWHM for the wider masks (longer PL wavelength) is only 2.6 meV. In addition, the FWHMs for the MQWs selectively grown on InP-on-insulator substrate are comparable to those for the references (MQWs grown on InP substrate and InP-on-insulator substrate with no selective mask). These results indicate high uniformity of both thickness and material composition of each QW in the same stack. The eight-channel DML array on silicon was fabricated using these MQWs. Figures 6(a) and 6(b) show cross-sectional bright-field scanning transmission electron microscope (BF-STEM) images of the membrane lasers for the shorter ($\lambda {0}$) and longer ($\lambda {7}$) wavelength channels, respectively. Despite the variation in total III-V thickness from about 315 to 350 nm due to the selective growth, flat membrane BHs without dislocations were obtained. Figure 6(c) shows a top-view photograph and the schematic of the laser array using these MQWs, and Fig. 6(d) shows cross sections of the membrane laser. To achieve single-mode lasing with asymmetric optical outputs from the front side of the laser arrays, we designed a 140 µm long DFB grating comprising a 120 µm long front-side section and a 20 µm long back-side section for each channel [37]. The Bragg wavelength of the back-side section was designed to be 6 nm longer than that of the front-side section so that the longer wavelength side of the stopband defined by the front-side grating is selected. The grating-index coupling coefficient was designed to be ${\sim}{600}\;{{\rm cm}^{- 1}}$. The grating pitch of each channel was individually designed on the basis of a different target wavelength and different equivalent refractive index due to the thickness variation. The active region is 0.7 µm wide and 140 µm long. To support 25 Gbit/s NRZ and 50 Gbit/s PAM-4 signals, we designed a longer cavity compared with our previously developed lasers (80 µm). In particular, the higher output power (with good modulation linearity) provided by longer cavity increases the signal-to-noise ratio of the transmission, which is particularly important for higher-order modulation formats such as PAM-4 [38,39].

 

Fig. 6. Cross-sectional bright-field scanning transmission electron microscope (BF-STEM) image of the selectively grown lasers on silicon for (a) the shortest wavelength channel and (b) the longest wavelength channel. (c) Top-view photograph (left) and the schematic (right) of the membrane laser array and (d) schematic of the cross sections of the membrane laser.

Download Full Size | PDF

4. DEVICE CHARACTERIZATION

A. Static Characteristics

We measured static lasing characteristics at 25°C under continuous-wave conditions. The output light for each channel was launched into a high-numerical-aperture fiber (HNAF), which was fusion-spliced to a standard single-mode fiber by butt-coupling the facets of the HNAF and SSC. Figure 7(a) shows optical output powers and bias voltages of the eight lasers versus injected current (L-I-V characteristics). We achieved lasing for all eight channels, and the fiber-coupled output powers were greater than 1.5 mW. The kinks in the L-I curves are due to the mode hopping caused by the imperfect design of asymmetric grating. We consider that the relatively long back-side grating (20 µm) induced the mode hopping, and it can be suppressed by optimizing the grating length. Figure 7(b) shows the lasing spectra with bias currents from 15 to 23 mA, together with the corresponding PL spectra measured just after the selective growth of MQWs. The deviations between the measured and designed lasing wavelengths are less than 2 nm, and the mean channel spacing is 860 GHz. In addition, the detuning between the PL peak and lasing wavelength is almost constant (${\sim}{35}\;{\rm nm}$) for all channels, which is designed to provide gain spectra peak at around lasing wavelength. These results indicate that the selective-area epitaxial growth on the InP-on-insulator substrate enables precise controllability of the lasing wavelength of the multiwavelength lasers with optimal gain mediums.

 

Fig. 7. Static characterization of selectively grown membrane lasers on silicon. (a) Optical output powers and bias voltages of eight lasers versus injected current under continuous-wave operation at 25°C. (b) Lasing spectra of eight lasers at 25°C together with the corresponding PL spectra measured just after the selective growth of MQWs.

Download Full Size | PDF

B. Dynamic Characteristics

Finally, we evaluated the dynamic characteristics of the selectively grown membrane laser array at 25°C. We directly modulated the lasers with 25.8 Gbit/s NRZ signals using a ${{2}^{31}}{-}{1}$ pseudo-random bit sequence (PRBS). We did not use any pre-emphasis or equalization. The eye diagrams observed with a sampling oscilloscope for all eight channels are shown in Fig. 8(a). The bias currents were 15.0 to 26.5 mA, the bias voltages were 1.84 to 2.65 V, and the voltage swings were 0.89 to ${1.58}\;{{\rm V}_{p - p}}$. We achieved 25.8 Gbit/s direct modulation with a dynamic extinction ratio greater than 4.0 dB for all eight channels.

 

Fig. 8. Dynamic characterization of selectively grown membrane lasers on silicon. (a) Eye diagrams for the eight-channel membrane lasers directly modulated with 25.8 Gbit/s NRZ signal. (b) 56 Gbit/s (28 Gbaud) PAM-4 back-to-back bit-error-rate (BER) characteristics of the shorter (lane 0) and longer (lane 7) wavelength devices, together with the equalized eye diagrams.

Download Full Size | PDF

In addition, we also modulated the lasers with PAM-4. For this measurement, grey-encoded PAM symbols were generated at 56 giga-samples/second by an arbitrary waveform generator (AWG) with an analogue bandwidth of ${\sim}{25}\;{\rm GHz}$. The length of the PRBS was ${{2}^{17}}$ bits. A root-raised cosine filter with 0.01 roll-off was applied for pulse shaping. The signal was then amplified electrically through a linear radio frequency (RF) amplifier with 22 dB gain and 55 GHz bandwidth. The modulated signal was added to the DC current via a bias-tee (65 GHz bandwidth) and injected to the membrane laser via an RF probe (40 GHz bandwidth). The output light was launched into the fusion-spliced HNAF by butt-coupling it to the SSC. The photodetection was carried out using a commercial 40G p-i-n photodiode integrated with a transimpedance amplifier. The signal was then stored at 160 giga-samples/second using a real-time digital storage oscilloscope for offline processing. For digital signal processing at the receiver, a 50-tap, half-symbol, linear least-means square equalizer was used. No optical amplifier or any complicated signal processing technique, such as nonlinear equalization, was used in these measurements. Figure 8(b) compares the 56 Gbit/s (28 Gbaud) back-to-back bit-error-rate (BER) characteristics of the shorter wavelength channel (lane 0) and longer wavelength channel (lane 7), together with the eye diagrams. The bias current, bias voltage, and the voltage swing for lane-0 (lane-7) were 24.9 mA (23.0 mA), 2.29 V (2.36 V), and ${1.89}\;{{\rm V}_{p - p}}$ (${1.51}\;{{\rm V}_{p - p}}$), respectively. We observed clear eye openings for 56 Gbit/s PAM-4 modulation for both the longer and shorter wavelength channels without using any complicated signal processing such as nonlinear equalization [38,39]. The obtained BER is lower than the forward error correction limit of ${2.4} \times {{10}^{- 4}}$ required for data-center networks [32]. These results indicate the lasers’ potential for use in data centers and short-reach optical communications.

5. CONCLUSION

We have demonstrated a wide-wavelength-range directly modulated laser array using multiwavelength MQWs selectively grown on an InP-on-insulator substrate. The key to achieving the high crystal quality is to keep the total III-V thickness less than the critical thickness, as well as to control the residual strain at the growth temperature. Although each laser has a different material composition and different thickness due to the selective epitaxy, the lasing wavelengths were successfully controlled with a wavelength range of over 38.2 nm, and the lasers can be directly modulated up to 25.8 Gbit/s (NRZ) and 56 Gbit/s (PAM-4). This proves that both the crystal quality and material gain coefficient of the MQWs selectively grown on the InP-on-insulator substrate reach the required levels for the state-of-the-art data-center communications. The capability of integrating the wide-wavelength-range directly modulated lasers on silicon with a selective-area epitaxy paves the way for establishing III-V/Si WDM-PICs for future data-center networks.

A next important step towards the fabrication of PICs will be to demonstrate the scalability of this technique using a larger InP-on-insulator wafer. There is no fundamental limitation up to 150 mm because InP wafers are commercially available, but the current standard wafer size for silicon-photonics technologies is 200 or 300 mm. This mismatch can be resolved by using die-to-wafer bonding technologies with relatively large InP coupons. Note that no precise bonding alignment is required in this case, because all the III-V device fabrication processes are carried out using lithographic markers on a silicon substrate. In addition, the epitaxial lift-off and Smart Cut technologies are promising ways to reduce wafer cost because they enable InP donor wafers to be recycled [40,41]. An epitaxial growth using InP membrane on silicon wafer fabricated by Smart Cut has already been demonstrated [42], but the crystal quality of the III-V layers should be improved to use it as a fabrication method of optical devices. Moreover, direct heteroepitaxial growth technique with dislocation filters is another promising candidate to fabricate InP-on-insulator wafers at low cost [10,11].

Finally, integrating various III-V devices with SiPh devices is also important. This can be achieved by replacing the thermally oxidized wafer used in this work with a silicon-on-insulator (SOI) wafer. Membrane BH lasers optically coupled to a silicon waveguide have already been fabricated by using a directly bonded MQW layer [43,44], which confirms the feasibility of this integration scheme. Altogether, we believe that membrane BH lasers on silicon and their fabrication scheme using InP-on-insulator substrate will pave the way to integrating III-V/Si photonics.