1. Introduction

The scale, complexity and performance of silicon photonic integrated circuits (PICs) has revolutionized optical communications in the past decade [1]. Perhaps the most surprising aspect of this revolution is the fact that silicon does not possess many desirable optical properties (apart from a high refractive index) and the silicon photonics revolution was primarily driven by the availability of a foundry fabrication infrastructure, courtesy of the microelectronics industry, that could be applied to optics [2–4]. Over the past two decades, a wide variety of component designs have been optimized and their fabrication process perfected for silicon [5] and it is hard to foresee a similar investment of resources in any other material platform. On the other hand, there are a number of application areas in which silicon’s lack of desirable optical properties proves a severe limitation to achieving system performance. These limitations include the absence of a direct bandgap, lack of a $\chi ^{(2)}$ nonlinearity to build fast electro-optic devices and zero piezoelectric response which makes it challenging to design acousto-optic devices. As a representative example, one of the key challenges facing transceivers for exascale systems [6] is avoiding the $\sim$ 3 dB penalty for coupling light from the III-V laser die to the silicon PIC [7]. Electro-optic frequency comb [8] based coherent communication systems [9] will also benefit greatly from monolithic integration of lasers and modulators. On the quantum photonics side, one of the outstanding problems facing linear optic implementations of quantum computing is implementing feed-forward routines on a chip [10], which requires (monolithically) interfacing fast, low loss modulators with efficient single photon detectors. Despite the outstanding performance improvements of carrier based depletion modulators, electro-optic modulators present the only near-term solution that can satisfy both the bandwidth ($\sim$ 40 GHz) and loss requirements (< 3 dB) necessary for scalability [11]. Other application areas where alternative material platforms are worth exploring are: cryogenic photonic circuits for interfacing superconducting digital circuits with the outside world [12] and integrated acousto-optics, which requires a piezoelectric material for exciting acoustic waves [13–15].

Gallium arsenide (GaAs) presents a viable alternative to silicon for these applications as it possesses all the desirable optical properties that silicon lacks: a direct bandgap, a $\chi ^{(2)}$ nonlinearity, and a (weak) piezoelectric coefficient [16]. More importantly, GaAs has a refractive index that is almost identical to silicon ($\lambda$ = 1.55 ${\mu }$m), making it easy to port a variety of optimised photonic designs and fabrication process flows between the platforms. In contrast to other electro-optic platforms like lithium niobate, it has a higher refractive index allowing compact component design, which is key to monolithic systems integration. GaAs also provides a natural route towards incorporating active gain media like quantum dots and wells, which are promising for applications in both classical and quantum [17] photonics. Traditionally integrated photonics in GaAs has suffered from the low index contrast achievable between GaAs and the AlGaAs buffers which serve as waveguide cladding layers. The low index contrast leads to large mode sizes and bend radii which make photonic integration challenging [16]. In addition, the reduced optical power density (due to larger mode area) makes it difficult to adequately exploit the nonlinear coefficients for frequency conversion and EO modulator applications [18,19]. In recent years, there has been tremendous progress in the development of thin films of GaAs on low index media (particularly silicon oxide and nitride) by wafer bonding [20] and a wide variety of devices showing impressive nonlinear performance have been demonstrated [21–25] . While wafer bonding is indeed a promising route towards building III-V PIC platforms, unlike in the case of silicon-on-insulator, it is not necessary. This is mainly due to the ease of obtaining defect-free (lattice matched) growth of III-V thin films on foreign substrates, allowing efficient film release [26]. A III-V PIC process that avoids wafer bonding and can combine arbitrary device complexity with high efficiencies is therefore worth investigating and is the main goal of this paper. In the long term, we also believe the waveguide losses in such a suspended platform are expected to be lower on account of avoiding the excess dissipation at the bonded interface.

In this work, we show that a multi-step fabrication process, derived from a silicon photonics foundry platform (Cornerstone) [27], can be applied to build large scale photonic integrated circuits in suspended GaAs. The suspension of the GaAs layer, achieved by selective wet etching of the underlying AlGaAs film, is necessary to achieve the requisite high refractive index contrast [28]. By moving to a tethered rib waveguide geometry and judicious choice of etch release holes, all the components of a standard (passive) PIC platform, in particular grating couplers, waveguides, ring-resonators and waveguide splitters (multi-mode interference couplers), can be adapted to the suspended GaAs platform without compromising performance efficiency. The passive devices reported in this work serve as a key building block for the development of active devices, in particular, efficient integrated electro-optic and acousto-optic modulators, which are currently under development.

2. Fabrication process flow

Figure 1 shows a schematic illustration of the main process steps for a grating coupler fabricated using this process. The process starts by patterning the grating coupler teeth and defining the outline of the waveguides and the tethers (GRATING etch) in GaAs (Fig. 1(a)). The patterning is carried out using electron beam lithography with hydrogen silsesquioxane (HSQ) resist and the GaAs layer is etched using a standard Ar/Cl$_{2}$ chemistry with etch thickness monitored using an ellipsometer to ensure precise etch depths (within $\pm$ 5 nm) are achieved. This is followed by a RIB etch step (Fig. 1(b)), where the GaAs layer is etched a further 50 nm to define the rib waveguides. The grating coupler region is protected with HSQ resist during this step. The two layers are registered with respect to each other using a set of alignment marks defined during the GRATING etch step. A FULL etch step is next carried out by etching the remaining 100 nm of the GaAs layer (+ 20 nm overetch) to access the AlGaAs buffer as shown in Fig. 1(c). To suspend the GaAs layer, the Al$_{0.7}$Ga$_{0.3}$As buffer is selectively etched in a dilute (24${\%}$) hydrofluoric acid (HF) solution. To remove any remnants of etch residue, the sample is cleaned in a dilute potassium hydroxide solution (KOH:H$_2$O = 1g:4mL) and flash dried using isopropanol [29].

 

Fig. 1. Illustrative process flow schematic for building a suspended waveguide platform in GaAs. The process follows the standard three step process used to fabricate passive silicon photonic devices. The GaAs thickness (in nm) in each region is indicated. (a) A 70 nm GRATING etch is used to define the grating coupler in GaAs. (b) The GaAs is etched a further 50 nm to define the rib waveguides (RIB etch). (c) Finally, a 100 nm FULL etch is carried out to reach the AlGaAs layer. The exposed AlGaAs layer is selectively etched away in weak HF acid and the sample is covered with $\sim$ 2 ${\mu }m$ SiO$_{2}$.

Download Full Size | PDF

3. Suspended rib waveguides for low-loss propagation

A cross-section of a suspended GaAs rib waveguide fabricated in this platform is shown in Fig. 2(a). A zoomed-out SEM image of the rib waveguide suspended by tethers is shown in Fig. 2(b). The normalized total electric field of the fundamental transverse electric (TE) mode, calculated using a numerical mode solver (COMSOL Multiphysics), is overlaid for reference. The high refractive index of the GaAs layer ensures that the mode is mainly confined to the rib region, with very little leakage into the surrounding GaAs or the AlGaAs buffer. By designing rib waveguides with waveguide width $\sim$ 540 nm and total rib width $\sim$ 6 ${\mu }$m, we can ensure that the optical field has negligible overlap ($\eta _{ov}$) with the remaining AlGaAs buffer layer and is tightly confined within the GaAs waveguide. The choice of the rib width was mostly determined by the difficulty of suspending wider GaAs devices on account of the lack of intrinsic stress in the GaAs layer , which results in bowing of the suspended structures. As mentioned before, the high refractive index allows us to work with smaller rib widths without sacrificing performance (waveguide loss). More importantly, it allows us to design compact, low loss suspended waveguide bends with bend radii $\sim$ 25 ${\mu }$m in this work, and the prospect of achieving compact microring resonators with radii $\approx$ 5 ${\mu }m$. After HF acid release, the suspended GaAs film is capped with $\sim$ 2 ${\mu }m$ of silicon oxide deposited using plasma enhanced chemical vapor deposition (PECVD). The oxide film is necessary for separating the metal electrodes (required for the electro-optic devices) from the GaAs layer. In addition, they provide mechanical rigidity to the suspended films by pinning them at the corners of the etch holes (shown by the red box in Fig. 3(b)). To test this rigidity, we performed a stress test by placing the chip (in isopropanol) in an ultrasonic bath at full power for $>$ 10 minutes. After removal from the bath, the chip showed no signs of weakening or structural damage, which ensures that it can survive later processing, critical for active devices. We are able to land fiber arrays on the GaAs chip without noticeable structural damage to the devices during optical characterisation. The mechanical rigidity is key for enabling the deposition of thick metallic electrodes for active devices and provides great benefit from a packaging perspective, which is especially critical for cryogenic operation.

 

Fig. 2. (a) Cross-section SEM image of a suspended GaAs rib waveguide structure. The electric field ($|E_{norm}|$) of the propagating transverse electric (TE) mode, calculated using a numerical FEM simulation using COMSOL, is overlaid. (b) Zoomed-out cross-section showing the GaAs rib waveguide held by suspension tethers. The extent of the HF undercut can be clearly seen. (c) Larger photonic circuits can be built by interfacing components together in this rib waveguide geometry. A suspended grating coupler interfaced with a rib waveguide is shown as a representative example.

Download Full Size | PDF

 

Fig. 3. (a) Measured fiber to fiber transmission spectrum of a representative grating coupler with parameters, period ($\Lambda$) = 660 nm, duty cycle ($\eta$) = 0.5 and etch depth ($d$) 70 nm. The fiber array was angled at 12 degrees, the laser output power set to 0 dBm and the coupling was optimised for maximum transmission. The simulated 2D fiber to fiber transmission spectrum for two air gap thicknesses (1.3 ${\mu }m$ and 1.5 ${\mu }m$) are shown by the red and magenta dashed curves respectively. An SEM image of a fabricated grating coupler before oxide encapsulation is shown in the inset. (b) Cross-section image of a GaAs rib waveguide after silicon oxide encapsulation, showing the bottom air gap is reduced to $\sim$ 1.3 ${\mu }m$, from a starting gap of 1.5 ${\mu }m$. The rigid anchoring of the structure at the ends, due to the oxide encapsulation, is indicated by the red box.

Download Full Size | PDF

4. Suspended grating couplers for efficient optical access

The standard set of passive design components available as part of a standard silicon photonics process development kit (PDK) are low-loss waveguides, grating couplers (GC), resonators and on-chip waveguide splitters and combiners. Amongst these, the grating coupler is probably the most important component, as it serves as the interface between the PIC and the outside world. A compact low-loss grating coupler is indispensable for large scale PICs as it allows in-situ device characterisation, without the need for chip-cleaving. Traditionally, GaAs based devices have relied on edge coupling as it is challenging to design efficient grating couplers when the index contrast between the core and cladding is low. While compact free space grating couplers have been optimized by the quantum dot community [30], their coupling efficiency is low and they are not suitable for building large scale PICs. On the other hand, efficient out-of-plane grating couplers have long served as the de-facto standard in the silicon photonics community. Figure 3(a) plots the measured fiber-fiber transmission spectrum of a grating coupler test structure that consists of two grating couplers linked by a suspended ridge waveguide. The device was probed using a fiber array angled at 12 deg. The separation between the couplers is 127 ${\mu }$m and the waveguide bends have a radii of 25 ${\mu }$m to give a total waveguide length of 155 ${\mu }$m. An SEM image of our suspended GaAs focusing grating coupler is shown in the inset of Fig. 3(a). The grating coupler is supported using tethers, defined during the GRATING etch step. The high refractive index of GaAs allows us to keep the light confined to the central region ($t_{GaAs}$ = 220 nm) and efficiently focus it into the rib waveguide. The critical fabrication step in suspending a grating coupler is ensuring that all of the GC is released during the wet etch step. This is ensured by providing release holes both to the side and rear of the GC, as can be seen in the inset. The grating coupler design is modified from the standard silicon designs to account for the difference in surrounding refractive indices and thicknesses. In the suspended GaAs devices, the underside cladding is air and the topside cladding is silicon dioxide. This helps reduce the insertion loss, as the guided wave is more effectively scattered towards the higher index side (the top oxide) towards the fiber, rather than towards the substrate. The main parameters affecting the coupling efficiency are the grating period ($\Lambda$), etch depth ($d$), duty cycle ($\eta$), and thickness of the AlGaAs cladding layer. For $\lambda$ = 1550 nm, the design parameters used for the device shown in Fig. 3 are $\Lambda$ = 660 nm, $d$ = 70 nm, ${\eta }$ = 0.5. The measured peak fiber to fiber coupling efficiency is $\sim$ -8 dB at $\sim$ 1550 nm, which amounts to $\sim$ 4 dB insertion loss per coupler. One of the key parameters that affects the optimal coupling efficiency is the gap between the GaAs device layer and the substrate, which is determined initially by the AlGaAs layer thickness. The starting AlGaAs thickness of 1.5 ${\mu }$m lies at the bottom of the coupling efficiency curve and any reduction in the spacing between the GaAs membrane and the substrate will improve that number. This can be seen by the simulated (2D) coupling efficiency spectra for devices with air gaps of 1.3 ${\mu }m$ (red) and 1.5 ${\mu }m$ (magenta) and a top oxide cladding of 2 ${\mu }m$ [5]. As can be seen from a zoomed-in cross-section SEM image of the waveguide in Fig. 3(b), the deposition of the top cladding oxide loads the membrane making it sag. This lowers the gap between the membrane and the substrate and ends up increasing the coupling efficiency. From our cross-section SEM images, we currently estimate the bottom gap to be $\sim$ 1.3 ${\mu }$m.

5. Suspended multimode interference couplers for power distribution

To build scalable PICs, it is critical to have the ability to split and recombine light. Linear networks of waveguide splitters form a key building block for optical implementations of quantum information processing [31], deep neural networks [32] and optical phased arrays [33]. A 2x2 multimode interference coupler (MMI) is the standard building block that underpins these linear networks. Figure 4(a) shows an SEM image of 2x2 MMI fabricated using the suspended GaAs platform and based on a standard silicon foundry design [27]. A zoomed-in image of the same device is also shown. The MMI center region length and width are 45 ${\mu }$m and 6 ${\mu }$m respectively with the rib width being 12 ${\mu }$m. The waveguides are tapered from a starting width of 540 nm to 1.5 ${\mu }$m over a taper length of 20 ${\mu }$m. The input (and output) ports are separated by a center-to-center distance of $\sim$ 2 ${\mu }$m. The measured transmission spectrum of the best-performing MMI is shown in Fig. 4(b). The plot shows the transmitted power from the two output ports (labelled by red and magenta arrows in the inset microscope image). The grating coupler transmission spectrum from Fig. 3(a) is overlaid for reference. From Fig. 4(b), it is clear that the transmission spectrum of the MMI is dominated by the grating coupler transmission characteristics. This is expected, since the 1-dB insertion loss bandwidth of the MMI is $\sim$ 100 nm, whereas the 1-dB transmission bandwidth of the grating couplers is $\sim$ 30 nm. To accurately quantify the insertion loss and bandwidth, a series cascaded MMI structure is necessary. On the other hand, the single MMI measurement reported here clearly shows that power splitting in the two output arms is roughly equal ($\sim$ 50:50) and the excess insertion loss introduced by the suspended MMI (over the 3 dB due to power splitting) is < 0.5 dB, primarily limited by our fiber alignment. With further design and fabrication optimization, these devices can be used to effectively split and re-combine light on a GaAs chip, exactly analogous to silicon. The MMI devices also give a sense of the the scale and complexity of the devices that can be engineered. As silicon photonics and before it silicon microelectronics have shown, once a set of robust building blocks have been demonstrated, circuits of arbitrary complexity can be synthesized by connecting these building blocks together in the desired order.

 

Fig. 4. (a) SEM image of a fabricated suspended 2x2 multimode interference (MMI) coupler. A zoomed-in image of the MMI is shown on the right. (b) Measured transmission spectrum through the two output ports (labelled 1 and 2 and indicated by the red and magenta curves). The reference fiber to fiber transmission spectrum (from Fig. 3(a)), without the MMI, is shown in blue. A microscope image of the complete MMI with the input and output ports labelled is shown in the inset.

Download Full Size | PDF

6. Suspended microring resonators for resonant optical recirculation

The final component of the passive devices toolkit for suspended GaAs photonics are microring resonators. High quality factor dispersion engineered microring resonators have served as the foundation of a variety of experimental advances in photonic sensing, frequency combs and on-chip generation of single photon pairs by spontaneous four wave mixing [34]. Figure 5(a) shows an SEM image of a suspended microring resonator fabricated using this platform. The ring radius is 25 ${\mu }$m and the waveguide resonator gap was designed to be 325 nm. Figure 5(b) shows the measured transmission spectrum of the resonator showing a series of TE resonances. The TE mode selectivity is determined by the polarisation selectivity of the grating coupler. Figure 5(c) shows a narrow band transmission scan around one of the resonances. Overlaid is a fit to the spectrum using a Lorentzian lineshape. The extracted quality factor ($Q_{t}$) of the device is $\sim$ 15000. The $Q_{t}$ of the device in Fig. 5 is primarily dominated by coupling to the waveguide. By increasing the waveguide resonator gap, we can reduce this coupling and achieve higher $Q_{t}$. Figure 6(a) shows the transmission spectrum of a microring resonator with 375 nm waveguide resonator gap. Figure 6(b) shows the normalized transmission spectra for two selected modes of the two microrings (indicated by the red boxes in Figs. 5 and 6). The two devices have identical radii (25 ${\mu }m$) but different waveguide resonator gaps of 325 nm (red curve, spectra from Fig. 5(c)) and 375 nm (blue curve) respectively. The $Q_{t}$ of the 375 nm gap device is $\sim$ 31560 indicating that waveguide coupling plays a significant role in determining our measured quality factor.

 

Fig. 5. (a) SEM image of a suspended GaAs microring resonator with radius 25 ${\mu }m$ and waveguide resonator gap 325 nm. A zoomed-in image of the ring is shown on the top for reference. (b) Measured transmission spectrum of the resonator showing a series of TE resonant modes ($P_{in}$ = -6 dBm). The TE polarisation is determined by the grating coupler. (c) Zoomed-in scan of one of the resonances, fitted with a Lorentzian lineshape, giving a measured $Q_{t}\sim$ 15000.

Download Full Size | PDF

 

Fig. 6. (a) Measured transmission spectrum of a 25 ${\mu }$m radius microring resonator with 375 nm waveguide resonator gap (b) Normalized transmission spectra plotted as function of offset from the resonance center wavelength for two microring resonators with identical radii (25 ${\mu }m$) and different (designed) waveguide resonator gaps (325 nm and 375 nm). The resonant modes plotted here are indicated by the red boxes in Figs. 5 and 6.

Download Full Size | PDF

We can estimate the propagation losses of the waveguide using the measured quality factors of the microring resonators. The measured total quality factor ($Q_{t}$) of the microring can be written as $Q_{t}^{-1} = Q_{i}^{-1} + Q_{c}^{-1}$ wherein $Q_{c}$ and $Q_{i}$ represent the coupling and intrinsic quality factors [5]. To extract the intrinsic loss of the device ($Q_{i}$) which corresponds to the waveguide loss coefficient ($\alpha _{i}$), we need to determine whether the microring $Q_{t}$ is dominated by intrinsic loss, coupling loss or if the two losses are comparable.

To do this, we can start by looking at the two extreme regimes. If we assume the resonator is dominated by intrinsic losses ($Q_{i}\ll Q_{c}$), we can approximate $Q_{t} \sim Q_{i} = 2{\pi }n_{g}/({\lambda }{\alpha _{i}})$, where $n_{g}$= 3.84 is the resonator group index and $\lambda$ = 1.55 ${\mu }m$. This gives us the (intrinsic) waveguide loss coefficient $\alpha _{i}\sim$ 21.4 (45) dB/cm.

On the other hand, if the resonator $Q_{t}$ is dominated by coupling losses ($Q_{c}\ll Q_{i}$), we can approximate: $Q_{t} \sim Q_{c} = -{\pi }Ln_{g}/({\lambda }log_{e}|t|)$, where $L$ is the round-trip length and $t$ is the resonator point-coupling transmission coefficient. Using the measured transmission minima [35], $T_{min}={(t-a)^2/(1-ta)}^2$ where $a$ is the single-pass amplitude transmission coefficient, we can estimate the (coupling-dominated) waveguide loss coefficient ($\alpha _{c}$) as $\alpha _{c} = -log_{e}(a^2)/L$. From our measured transmission spectra, we estimate $\alpha _{c}$ to be $\sim$ 20.16 (37.68) dB/cm for the $Q_{t} \sim$ 31560 (15000) rings respectively.

From the two calculations, it is clear that $Q_{i}\sim Q_{c}$ for the $Q_{t}\sim$ 31560 microring. This indicates we are close to critical coupling, which is also confirmed by the transmission extinction $\sim$ 20 dB. Assuming critical coupling ($Q_{i}= Q_{c}$), $Q_{t} = Q_{i}/2$ and the estimated waveguide loss coefficient ($\alpha ) \sim \alpha _{i}/2$ = 10.7 dB/cm. The $\alpha$ should be treated as an upper bound on the loss as we believe our devices are over-coupled but need more measurements on varying waveguide resonator gaps to confirm this hypothesis. To reduce the waveguide loss to the levels routinely achieved in silicon, there are several fabrication process improvements that can be applied. In particular, we can significantly improve our post-release device cleaning procedures to remove residual contaminants from the HF acid undercut and incorporate surface passivation techniques that have shown to significantly improve the $Q_{t}$ in GaAs devices [36].

7. Conclusions

In summary, we have demonstrated that the complexity of the standard (passive) silicon photonics process can be readily transferred to more interesting optical materials, in particular GaAs. The similarity of the refractive indices of the two materials ensures that high device performance can be readily ensured without requiring extensive component re-design. Bringing the scale and complexity of silicon photonics to more interesting optical platforms will be revolutionary for device applications in wide-ranging areas from quantum photonics to cryogenic photonic circuits. Moving forward, we will extend this platform to demonstrate low-loss long spiral ($L_{wvg}{\sim }$ cm) waveguides, and active electro-optic and acousto-optic devices.