Silicon photonics based on silicon-on-insulator (SOI) substrates has been rapidly growing over the past two decades, providing various on-chip passive/active photonic functionalities (waveguide-based passive components, modulators, photodetectors, etc.) by taking advantage of mature planar CMOS technology, leading to the development of Si photonic integrated circuits (PICs) [1–4]. Moreover, III–V-on-Si (III-V/Si) heterogeneous integration techniques have been developed to address the laser source issue for Si photonics and have now been adopted in scalable state-of-the-art CMOS-compatible processes [5]. Today, Si photonics is playing a leading role in the community of integrated photonics and has found applications in a large number of areas including optical interconnects [6,7], telecommunications [8], computing [9], and so on [10]. Meanwhile, significant progress spanning the last decade on integrated nonlinear photonics has emerged as a new paradigm for both nonlinear optics research and applications. An intriguing offering of integrated nonlinear photonics is its capability of generating new classes of coherent, ultra-broadband light sources (i.e., microcombs) in nonlinear waveguides [11,12], which is not attainable from linear photonics systems. Microcombs have triggered widespread use of chip-scale nonlinear devices in applications [13] including ultrahigh-capacity coherent telecommunications [14,15], optical frequency synthesis [16], optical atomic clocks [17], quantum optics [18], etc. In the past few years, significant technological advances have enabled ultralow-loss nonlinear waveguides and ultrahigh-quality-factor ($Q$-factor) microresonators [19–23]. Those breakthroughs have essentially allowed high-efficiency nonlinear processes operating at dramatically reduced power levels of milliwatts or sub-milliwatts [19,21,24–28], eliminating bulky optical equipment such as benchtop pump lasers and amplifiers. The power requirements in nonlinear waveguides are now compatible with co-integrated III-V/Si pump lasers on a single chip [29], which represents a key step towards harnessing the on-chip nonlinear properties from individual and passive-only nonlinear devices, to mass-manufactured system-level architectures and application developments.

Therefore, it is naturally desired to embed nonlinear functionalities in relatively mature Si PICs for complete chip-scale nonlinear photonics with the ability of direct signal generation and processing. However, Si has inherent disadvantages regarding many nonlinear applications, such as its small-bandgap-induced strong two-photon absorption (TPA) and related free-carrier losses at telecom wavelengths and insignificant ${\chi }^{(2)}$ nonlinearity [30–32], thus limiting its nonlinear efficiency and available functionality at telecom wavelength bands where other photonic applications are clustered. Consequently, there is a strong need to look for an alternative nonlinear material candidate that is perfectly suited for compatibility with Si photonics integration and adding the missing piece. Among various investigated integrated nonlinear platforms from dielectrics to semiconductors [33–43], aluminum gallium arsenide (AlGaAs) is particularly attractive, due to its strong intrinsic second- and third-order nonlinearities, relatively large and tailorable bandgap, and large linear refractive index similar to Si that can ensure strong photon confinement and also efficient coupling to Si PICs. In particular, the bandgap can be engineered from about 1.42 eV to 2.95 eV by varying the alloy composition of ${\mathrm{Al}}_x{\mathrm{Ga}}_{1-x}\mathrm{As}$ [44] so that the operating wavelengths can be sufficiently below the half-bandgap wavelength, thus allowing mitigation of the most detrimental nonlinear losses induced by TPA effects [45,46]. For example, by choosing the Al composition to be around 0.2, the corresponding bandgap is 1.69 eV, which results in a TPA cutoff wavelength of 1467 nm, and thus the TPA can be ignored when operating at the telecom band around 1550 nm [42,47–49]. This advantage overcomes the intrinsic detrimental TPA effect in Si that hinders its nonlinear applications. As a consequence, AlGaAs promises ideal nonlinear optical figures of merit alongside linear Si photonics [50,51], and has long been termed as the “the silicon of nonlinear optical materials” at telecom bands [45]. Moreover, recently, the rapid growth of implementing these superior nonlinearities has been taking place by using an AlGaAs-on-insulator (AlGaAsOI) platform [42,43,52], which enables compact, high-confinement, ultralow-loss waveguides and microresonators. Thus, various on-chip ultra-efficient nonlinear applications have been unlocked including second-harmonic generation [53,54], super-continuum generation [55], Kerr frequency comb/soliton generation (threshold powers of the order of 20 μW) [21,56,57], and ultra-bright entangled-photon-pair generation [58].

Here, we further extend the integration of AlGaAs nonlinear photonics onto an SOI substrate with the goal of bridging “the silicon of nonlinear optical materials” and Si PICs. To achieve this goal, we propose a general heterogeneous architecture to combine the two types of PICs with fabrication compatibility and integration scalability. We demonstrate low-loss AlGaAs-on-SOI waveguides and microresonators with $Q$-factors over ${10}^6$, as well as low coupling loss ($<1\mathrm{dB}$) between AlGaAs and Si PICs. Efficient microcomb generation on a standard SOI platform using a telecom-band pump is reported for the first time, with a threshold power of as low as $\sim 0.25\mathrm{mW}$. Particularly, the pump and access of the microcomb are through Si waveguides. The demonstrated compatibility in both the fabrication process and power budget for the two classes of photonic functionalities unfolds the possibility of fully integrated scalable and advanced nonlinear PICs, by leveraging both the mature technology of Si photonics and superior nonlinear properties of AlGaAs.

Figure 1(a) depicts the proposed device configuration whereby light in the bottom SOI layer is coupled to the top AlGaAs layer and vice versa via evanescent coupling aided by a pair of inverse tapers of Si and AlGaAs waveguides. The two device layers are spaced by an oxide layer of certain thickness (200 nm in current case). In practice, this structure can be realized by low-temperature (typically $<400°\mathrm{C}$) wafer bonding of AlGaAs on a waveguide-patterned SOI wafer. More generally, such vertical, heterogeneous integration of AlGaAs PICs can be performed on a Si photonics substrate in which critical processes are completed for both passive and active components including waveguides, Si PN junction-based modulators, SiGe detectors, etc. Therefore, the proposed integration strategy possesses the capability of integrating high-quality nonlinear AlGaAs PICs on a fully functional Si PIC wafer in a back-end-of-line (BEOL) integration approach for large-scale and advanced PICs. It has to be noted that the BEOL integration of AlGaAs is fully compatible with the existing Si PIC process flow without modifying the thermal budget, distinguishing itself from silicon nitride-based nonlinear photonics, which normally requires high-temperature anneals. Here, for the purpose of proof-of-concept demonstration, the simple Si waveguide elements are defined for off-chip coupling access, while in the AlGaAs layer, a bus-waveguide coupled microring resonator is defined. The Si underneath the AlGaAs waveguides and resonators is removed to avoid mode leakage, resulting in a suspended structure above an air trench as shown in Fig. 1(b). Note that this air trench can also be filled with oxide, followed by chemical mechanical polishing for surface planarization. The input/output Si waveguide is shallow etched by $\sim 231\mathrm{nm}$ on an SOI substrate with a 500 nm Si layer and 1.0 μm buried oxide layer. In the coupling region, a shallow-etched Si waveguide is transitioned to a deep-etched Si waveguide with full 500 nm etch depth. In addition, cross-shaped vertical channels (VCs) are patterned on the Si layer to facilitate AlGaAs bonding with an improved yield. A 400 nm thick AlGaAs film is grown by molecular-beam epitaxy, and the aluminum composition is chosen to be 20% so that the TPA effect is negligible at wavelengths around 1550 nm, as discussed above. After bonding, the fully etched AlGaAs microrings are defined with different radii and widths in a range of 600 nm to 750 nm to ensure anomalous group-velocity dispersion (GVD) around 1550 nm, as previously demonstrated for Kerr frequency comb generation [43]. The bus waveguide is carefully designed to allow efficient coupling to the desired fundamental transverse-electric (TE0) mode. Note that initially the choice of Si thickness (500 nm) is considered to accommodate the co-integration of heterogeneous InP/Si lasers [29]. Nevertheless, the proposed integration architecture is also applicable to other SOI substrates with different Si thicknesses (e.g.,  220 nm), as well as different AlGaAs thicknesses. The field profile of the TE0 mode in the AlGaAs waveguide is shown in Fig. 1(b), indicating a tight mode confinement within the AlGaAs core. To optimize the coupling between Si and AlGaAs tapers, we performed numerical simulation for the transmission of the TE0 mode and chose the widths of Si and AlGaAs to be 200 nm and 300 nm, respectively, the length to be 200 μm, and the oxide thickness to be 200 nm with a theoretical coupling efficiency $>95\%$. Figures 1(c)–1(e) show the images of the fabricated AlGaAs/Si heterogeneously integrated device, indicating the chip-scale fabrication throughput, the realized AlGaAs/Si structure, and the coupling tapers between Si and AlGaAs.

In Fig. 2, we present the simplified fabrication process. The overall fabrication consists of three major steps: Si waveguide patterning on the SOI wafer, AlGaAs epi wafer bonding, and AlGaAs waveguide patterning [Fig. 2(a)]. The details of the process flow are shown in Fig. 2(b). For the entire process, a 248 nm deep ultraviolet (DUV) stepper was used for photolithography, which is well suited for scalable integration and volume fabrication. The fabrication started with the deposition of 200 nm oxide on a 100-mm-diameter SOI wafer using plasma-enhanced chemical vapor deposition (PECVD). The oxide layer was then patterned and dry etched with chemistries of ${\mathrm{CHF}}_3/{\mathrm{CF}}_4/{\mathrm{O}}_2$ in inductively coupled plasma reactive-ion etching (ICP-RIE). The oxide served as a hard mask, and the Si rib waveguide was formed by dry etching $\sim 231\mathrm{nm}$ of the 500 nm SOI in another ICP-RIE process with chemistries of ${\mathrm{SF}}_6/{\mathrm{C}}_4{\mathrm{F}}_8$. Next, the fully etched Si taper was realized by a second etch of the rest of Si using a photoresist mask together with the predefined oxide mask. After Si waveguide processing, the remaining oxide mask was removed by a dip in buffered hydrogen fluoride (BHF) solution, and the wafer was carefully cleaned for the following bonding process. The second step is to bond the AlGaAs film onto the patterned Si waveguide wafer. The 400 nm thick epitaxial AlGaAs film was grown on a GaAs substrate with an etch-stop layer of 500 nm thick ${\mathrm{Al}}_{0.8}{\mathrm{Ga}}_{0.2}\mathrm{As}$ in between. Here we employed a similar bonding process as reported in our previous work [21]. Different from our previous work, here we deposit 200 nm silicon dioxide on the as-grown AlGaAs surface before bonding using an optimized ICP-PECVD process, which results in a smooth deposited surface with a root-mean-squared roughness of $\sim 0.3\mathrm{nm}$. This oxide layer serves as a spacer for optimal coupling as well as for mechanical support of the suspended AlGaAs structures [as shown in Fig. 1(b)]. Additionally, in this work, we omit atomic layer deposition of a thin ${\mathrm{Al}}_2{\mathrm{O}}_3$ (alumina) passivation layer on the AlGaAs surface, as the combination of high stress in the oxide spacer layer and poor adhesion in the alumina passivation layer led to cracking and delamination of the deposited films. However, the absence of alumina passivation may result in slightly increased optical loss due to surface absorption. We note that this spacer oxide deposition could be eliminated by using chemical–mechanical planarization of the Si-photonic substrate, a technique that is widely available in standard CMOS foundries. Prior to bonding, the surfaces of the substrates were activated in ${\mathrm{O}}_2$ plasma. The bonding was carried out at a temperature of 200°C for 10 h. We note that the same process is compatible with bonding an InP epitaxial film, which will be certainly useful for future heterogeneous integration of laser sources simultaneously with AlGaAs. The GaAs substrate was removed by a combination of mechanical lapping and wet etching in a mixture of ${\mathrm{H}}_2{\mathrm{O}}_2$ and ${\mathrm{NH}}_4\mathrm{OH}$ at room temperature. The etch-stop layer was selectively etched in diluted HF. After bonding, the AlGaAs waveguides were then patterned using our previously optimized processes for low-loss AlGaAs waveguides [21] and finally cladded with a 1.0 μm PECVD ${\mathrm{SiO}}_2$.

We first characterized the transmission spectral response of our devices. TE polarized light from a tunable laser was coupled into one Si waveguide through a lensed fiber, and the output was collected by another lensed fiber from the other Si waveguide edge coupler, as denoted in Fig. 1(a). The transmitted power was detected with a high-speed photodetector, and the response was recorded using an oscilloscope. The spectral resolution of our measurement is 0.01 pm. The input power can be further adjusted with a variable optical attenuator. Figure 3(a) shows the transmission spectra of several waveguides with the same input/output Si edge couplers connected by a series of Si-AlGaAs tapers (two, four, and six tapers) and a pure Si waveguide (zero tapers) for reference. We extracted the taper loss at the wavelength of 1550 nm with a linear fit of the transmission as a function of the number of tapers in dB scale and show the results in Fig. 3(b), indicating a Si-AlGaAs taper loss of 0.87 dB, e.g.,  a coupling efficiency of 82%. For comparison, we also show the simulated taper transmission in the inset of Fig. 3(a) with a theoretical coupling efficiency of 95% at 1550 nm. Additionally, the effect of lateral misalignment between the tapers is also simulated, indicating a tolerance of at least 100 nm for $<1\mathrm{dB}$ loss increment. Improvements on the taper dimension control as well as alignment accuracy can be done to further improve the coupling efficiency.

The waveguide dispersion and the $Q$-factors of AlGaAs microrings were obtained by measuring the transmission of the waveguide-coupled resonator devices. For the GVD characterization, using a microring with free-spectral range (FSR) of 0.5 THz and a nominal width of $\sim 650\mathrm{nm}$, we measure the resonance frequencies of the TE0 mode family as a function of the relative mode number $\mu$, relative to a reference resonance at ${\omega }_0$ around 1550 nm. The resonance frequencies ${\omega }_{\mu }$ of the modes can be expanded as ${\omega }_{\mu }={\omega }_0+\mu D_1+\frac{1}{2}{\mu }^2{D}_2+\dots$, where $D_1/2\pi$ is the FSR around ${\omega }_0$, and $D_2/2\pi$ is related to the second-order dispersion and is positive (negative) corresponding to anomalous (normal) GVD. By fitting the data as shown in Fig. 3(c), $D_2/2\pi$ is extracted to be 62.8 MHz, indicating the anomalous dispersion for the AlGaAs waveguides. The intrinsic $Q_0$ are measured to be $0.97\times {10}^6$ and $1.12\times {10}^6$ for a small ring (radius of 12 μm and FSR of 1 THz) and a large ring (radius of 143 μm and FSR of 90 GHz), respectively, as shown in Figs. 3(d) and 3(e). Here the $Q$-factors are lower than that of our previous work [21,43], which is likely due to the bottom spacer oxide passivation and can be improved by optimizing the passivation quality. Nonetheless, the demonstrated $Q$-factor reveals a propagation loss as low as 0.5 dB/cm in the AlGaAs-on-SOI waveguides, which is also on par with the state-of-the-art result in Si PICs.

To further highlight the capability of efficient nonlinear applications in the developed AlGaAs-SOI platform, we performed a frequency comb generation experiment, in which the pump laser was injected to the AlGaAs ring through the input Si waveguide edge coupler, and the comb spectrum was collected from the output Si waveguide edge coupler. For the ring with an FSR of 1 THz, we pumped a resonance at $\sim 1550\mathrm{nm}$ ($Q_0\sim 0.81\times {10}^6$) and recorded the output spectra at different pump powers in the AlGaAs bus waveguide. With a pump power of about 0.25 mW, the primary sidebands of optical parametric oscillation are clearly observed at locations corresponding to multiples of the resonator FSR, as shown in Fig. 4(a). When increasing the pump power to 0.4 mW, a primary comb at a $3\times$ FSR spacing emerges and covers a wavelength range of $\sim 175\mathrm{nm}$, as shown in Fig. 4(b). With a pump power of 3 mW, a complete comb at single FSR spacing is formed, spanning a broad wavelength range from 1450 nm to 1700 nm, as shown in Fig. 4(c). This is also evidence of chaotic comb generation, and better control of thermo-optic effects could lead to coherent soliton microcombs at similar power levels [56]. While several techniques could be helpful to overcome the thermal problem to achieve coherent mode-locked microcombs including using an auxiliary laser heating for temperature compensation [59] or operating the microring at very low temperatures to suppress the thermal effect [56], a more practical approach is to engineer a resonator with normal dispersion, which will reduce the impact of thermal effects on soliton formation, and has been demonstrated in our recent work [57] and can also be employed in future designs for the AlGaAs-SOI platform.

Additionally, we estimated the comb threshold power theoretically, using the expression [60] $P_{\mathrm{th}}\approx 1.54\left(\frac{\pi }{2}\right)·\frac{Q_c}{2Q_L}·\frac{n^{2}V}{n_2\lambda Q_L^2}$, where $\lambda$ is the pump wavelength, $n$ is the modal refractive index, $n_2$ is the nonlinear refractive index, $V$ is the mode volume, and $Q_c$ and $Q_L$ are the coupling and loaded quality factors of the resonator, respectively. Given a critical coupling condition with $Q_c=2Q_L=Q_0$ and the measured intrinsic $Q_0$, the theoretical threshold power is 0.07 mW, which is lower than that observed in the measurement. This can be attributed to lower average $Q$ factors over all resonances and the stepwise wavelength tuning approach with less resolution. Lower threshold powers are thus expected by improving the overall $Q$ factors and using fast wavelength sweeps. Finally, we also tested the ring with 143 μm radius (FSR of 90 GHz) at the resonance of $\sim 1550\mathrm{nm}$ ($Q_0\sim 1.0\times {10}^6$) and show the comb spectrum at the pump power of $\sim 15.8\mathrm{mW}$ in Fig. 4(d), in which the comb lines are almost complete over the whole span, suggesting that such power levels can sustain a coherent soliton microcomb with an electronically detectable repetition rate.

In summary, we demonstrated a general architecture for efficient integration between two important integrated photonics platforms, i.e., a mature SOI waveguide platform and an emerging nonlinear III-V platform. We showcase efficient ${\chi }^{(3)}$-based optical frequency comb generation seamlessly integrated with Si PICs, while the same platform can be extended to incorporate other functions including ${\chi }^{(2)}$-based nonlinear photonics (e.g.,  second-harmonic generation). Importantly, the achieved low-loss AlGaAs-on-SOI enables efficient microcomb operation for Si PICs at power levels of only a few milliwatts, which is easily achievable with state-of-the-art standard III-V/Si heterogeneously integrated lasers. Our demonstration bridges mature SOI PICs and the superior nonlinear functionalities of III-V photonics, in addition to existing lasers, modulators, photodetectors, etc. The present result may provide a viable route towards fully functional nonlinear chips for miniaturized systems hosted in Si photonics and open new perspectives for related research and applications.

Acknowledgment. We thank Justin C. Norman, Chenlei Li, and Joel Guo for helpful discussions.