Micro-transfer printed high-speed InP-based electro-absorption modulator on silicon-on-insulator

Electro-absorption modulators (EAMs) are valuable components in photonic integrated circuits (PICs), especially for application in high-speed optical communications and sensing.^1–4 The modulators are required to have high bandwidth, a large extinction ratio (ER), low power consumption, and low operating voltage. In contrast to Mach–Zehnder Modulators,^5,6 EAMs have a very compact footprint. The intensity (electro-absorption) and phase of light (electro-refraction) are modulated by using Pockels, Frank–Keldysh (FKE), and quantum-confined Stark effects (QCSEs).^7–9 These latter effects shift the absorption edge of a material, essentially rendering previously non-absorbing wavelengths to become absorbent while also changing the index of the material below the bandgap according to the Kramers–Kronig relations.^10,11 Silicon-on-insulator (SOI) is the predominant platform for PICs as its high refractive index contrast allows for sub-micron waveguides, and, more practically, mature foundry fabrication processes result in high yield at a projected low cost, especially for the rapidly expanding data center market.^12–15 While SOI is an ideal platform for compact passive components such as waveguides and couplers, the performance of its active components suffers from various drawbacks. For instance, silicon is an indirect bandgap semiconductor unsuitable for amplification and lasing. Its primary modulators employ the carrier-induced refractive index change, which is known to be stronger in III–V materials.^16–18 There have been examples of Ge–Si EAMs^19,20 with both lateral and vertical junctions. However, they have not produced comparable performance metrics to their III–V EAM counterparts on their respective monolithic platforms, suffering from low ER, and needing very precise doping over a sub-micrometer-wide waveguide region. Due to these limits on active silicon devices, heterogeneous integration of III–V onto SOI is a promising strategy as it coalesces the high-performing III–V active devices with the lower cost platform and guiding capabilities in silicon.^21,22 III–V and Si waveguides are suitable for cross coupling as their refractive index contrast is relatively small (⁠ $Δ n ≃ 0.25 – 0.5$⁠) at the communication wavelength bands. This allows for easier coupling compared to other materials like lithium niobate (n = 2.2)²³ or silicon nitride (n = 2.0).²² Previous work^24–27 demonstrated high-performing InP EAMs on SOI using wafer bonding. An alternative promising integration method is that of micro-transfer printing (μTP), which has been gaining interest in recent years.^28–30 It is a “pick and place” technique with sub-micron placement precision capable of transferring arrays of devices from a source to a target substrate. It can potentially increase throughput and reduce costs by utilizing all the devices on the source substrate, and it may be possible to reuse the expensive substrate. While EAMs on InP substrates are a well-characterized, high-performing mature technology, it has not been engineered for heterogeneous integration via transfer printing.

In this work, we design and demonstrate a high bandwidth heterogeneous III–V InP EAM directly bonded using transfer printing to a 220 nm SOI waveguide circuit using a coupon-type structure with a footprint of 80 × 500 μm², which can be reduced in length. The length of the EAM waveguide device itself ranges from 220 to 300 μm. The integrated device exhibits a 25 dB ER over a 30 nm spectral range from 0 to −8 V and an insertion loss (IL) between 3.1 and 6.5 dB. The high-speed characteristics are measured, and a 50 Gbps data rate demonstrated.

The EAM epitaxial structure³¹ is shown in Table I and consists of a multiple quantum well (MQW) region in between a thick P-doped and a thin N-doped InP layer. The N-type cladding is thin to enable efficient coupling between the SOI waveguide and the EAM active region. It is also the layer where the coupon is bonded to the silicon waveguide. An InGaAs and InAlAs dual sacrificial layer is incorporated to undercut the coupons for transfer printing,^32,33 thereby ensuring a smooth and flat release surface capable of direct bonding. The intrinsic region has a thickness of 420 nm, and the calculated well-barrier band offsets were 226 and 30 meV for the electron and holes, respectively. The material was designed for high-speed applications, which included a small junction capacitance (⁠ $≈$ 110 fF for a 2 × 100 μm² device). The graded confinement on the P-side enhances hole transport, while the thick N-side AlInGaAs confinement compensates electron to hole transit times and increases intrinsic thickness to enhance high-speed characteristics. The heavy hole carrier transit time is calculated to be <15 ps,³⁴ which would lead to a bandwidth of over 100 GHz. The material was characterized with a 1 mm diameter circular PIN diode used to measure the bias-dependent absorption spectra under surface normal incidence. The characteristic response is plotted in Fig. 1, demonstrating the QCSE absorption modulation.

The longitudinal EAM device is designed in a waveguide configuration that evanescently couples to and from the SOI waveguides using adiabatic tapers that narrow down to 200 nm tips. This is achieved by directly printing the part-fabricated EAM on top of the SOI waveguide and using the taper to couple the fundamental TE mode from the SOI to the EAM active region by adiabatically changing the mode overlap.^36,37 The target circuits on a 220 nm SOI platform included two in/out grating couplers, which taper down to a single-mode 500 nm wide waveguide. For the EAM coupling section, the SOI waveguide is tapered up to 1 μm to increase alignment placement tolerance for the printed coupons. The 1 μm SOI waveguide has a simulated effective index of 2.76 at 1550 nm. When the effective index of the tapered EAM surpasses this, the mode overlap in the MQW region increases with the light becoming highly confined in the EAM active region. The MQW mode overlap is 20 to 22.5% for a 1 to 3 μm wide EAM, respectively (assuming no misalignment offset). A top-down schematic of the device design is shown in Fig. 2(a), along with the simulated cross-sectional mode profile at the (b) taper tip and (c) the 2 μm wide EAM mesa. Through simulations, using the Lumerical software suite,³⁸ the taper is designed to allow a single-mode transition from the SOI to the EAM. Figure 2(d) shows the simulated coupling efficiency with taper length where 99% coupling is obtained after 60, 70, and 100 μm long linear tapers for 1, 2, and 3 μm wide ridge modulator waveguides, respectively. Efficient design of the taper regions is essential to these devices as longer taper regions increase the capacitance while not effectively contributing to modulation. This is dependent on fabrication tolerances, as while a 1 μm wide device clearly needs the shortest taper, its printing accuracy must be very precise (<200 nm). Increasing the length of the taper would increase alignment tolerances, but at the cost of device performance. Some other methods can be used to also increase tolerances such as a narrower taper tip,³⁹ but adding to fabrication complexity. Device misalignment can excite higher-order modes, which may be an issue in more complex circuit configurations.

The EAM coupon fabrication consisted of five lithographic levels. The ridge waveguide was defined using e-beam lithography (EBL), and together with optical lithography, structures were dry etched to the N-layer using a SiO₂ hard mask and a Cl₂ based etch. The N-type metal was then deposited and annealed. The coupon boundary was then defined. For practical reasons relating to the available μTP stamp and ease of printing, the coupon length was 500 μm for all EAM device lengths. For shorter devices (and all device variations shown in this Letter), the length could be reduced. The tethers were defined with photo-resist, and the devices were undercut using a FeCl₃ solution.³² The coupons were picked with a polydimethylsiloxane (PDMS) stamp, which breaks the tethers, and are then printed to the SOI target, where the stamp is pushed into contact and peeled off to leave the coupon bonded to the SOI. These outlined process steps are visualized in Fig. 3(a). The accuracy of printing placement was measured to be ±0.5 μm. With an array of EAM devices printed, the residual tether resist was removed, and the coupons were collectively processed to interconnect the devices to bond-pads. SU8 was spun on, exposed, and etched back to planarize the devices and open the N-type contacts. The bond-pads were then deposited using a bilayer lithography process and electron beam evaporation of Ti/Pt/Ag/Au (50/20/500/150 nm). An optical image of a completed device is shown in Fig. 3(b), an electron micrograph in Fig. 3(c), and a focused ion beam (FIB) cross section of the completed EAM on SOI in Fig. 3(d).

The SOI circuit was fabricated using EBL and a fluorine-based dry etch for both the waveguide and gratings. The waveguide layer was fully etched (220 nm), while the gratings were partially etched by 75 nm. The measured grating coupler loss is 5.2 dB/coupler, and the waveguide propagation loss is 8 dB/cm. This corresponds to a peak loss of 12 dB for the full SOI circuit. Figure 4(a) shows the in/out reference spectrum used to normalize the EAM device characteristics.

Devices with ridge widths of 1, 2, and 3 μm were measured each with a modulation length of 100 μm and their respective taper lengths (outlined earlier). The static parameter characterizations were performed using an Agilent 8166A tunable laser source (TLS) to inject light from cleaved single-mode fibers (SMF-28) into and out of the SOI circuit at a 12° angle to gratings. A Thorlabs power meter was used to measure the output power, while a Keithley 2602A applied the bias to the device. The estimated absorption, α, was calculated using the measured normalized transmission spectra, T, through the device with length L, where $T = e − α L$ (L = 240 μm for a 2 μm wide ridge), and presented in Fig. 4(b). The QCSE shift is evident, and a 30 nm spectral bandwidth can be seen with the absorption reaching 400 cm⁻¹ at −8 V. Figure 4(c) shows the transmission with applied reverse voltage for wavelengths from 1550 to 1580 nm. From this, the ER was measured to be 30 dB at 1550 nm, and the ER shows a 25 dB extinction from 0 to 8 V up to 1580 nm. An ER of 10 dB is measured with −4 V at 1550 nm. The IL measured in the transparent region was between 3.1 and 3.5 dB for the 2 μm wide ridge. Compared to the simulation, 2.86 dB of IL is calculated for the measured misalignment of 0.37 μm, and the slight excess measured loss is attributable to waveguide loss due to sidewall roughness and/or P-doped InP region absorption. Both the 1 and 2 μm wide devices performed with similar ER and ILs indicating appropriate MQW mode overlap. The 3 μm wide mesa has slightly better ER, due to the need for longer tapers but suffers in terms of high-speed characteristics. Using ER/IL as a figure-of-merit, we obtain for a 1 μm device (L = 220 μm), where ER (dB/μm) = 0.136 and IL (dB/μm) = 0.018 that ER/IL = 7.56, which perform similarly to III–V monolithic modulators, and on average outperform Ge–Si EAMs.⁸ More information on the static characteristics can be found in Ref. 35.

The temperature dependence of the transmission of the EAM was examined by mounting the sample on a thermoelectric cooler. As expected, there is a reduction in the ER and a significant increase in the IL. The temperature sensitivity is significantly reduced for longer wavelengths however, as 1550 nm is on the absorption edge. The transmission for 1550 nm over reverse bias at varying temperatures can be seen in Fig. 4(d). The measured current–voltage (I–V) and capacitance–voltage (C–V) characteristics are plotted in Fig. 5. The dark current remains under 100 nA up to a reverse bias of −8 V.³⁵ The series resistance for the 1, 2, and 3 μm devices was 45, 41, and 20 Ω, respectively. The capacitance is effectively constant for each device for applied voltages below −0.5 V. The intrinsic region is not fully depleted for higher biases, indicated by the slight increase in the capacitance. The measured capacitance has a noticeable parasitic contribution, with a consistent ∼100 fF extra for each device, when compared to the junction capacitance calculation. This is most likely attributed to the bond-pads.

To evaluate the electro-optic (EO) bandwidth of the fabricated EAM, a high-speed radio frequency (RF) signal from a vector network analyzer (VNA) was applied to the EAM in conjunction with a DC bias through a bias-tee. The schematic of the experimental arrangement is shown in Fig. 6(a). The cable, bias-tee, and photodiode response is calibrated out from the link response obtained. The bandwidth measurement is carried out in the load-free configuration where an unterminated probe is used to drive the modulator, with extracted data presented in Fig. 6(b). The extracted cutoff frequency was up to 40 GHz for a 1 μm wide mesa. The DC bias for each device was −3.5 V, and the small signal swing was 500 mV. To generate the eye diagrams, a Keysight M8199B arbitrary waveform generator was used, and the transmitted light was measured using a 100 GHz detector and visualized with a Keysight UXR1104A real-time oscilloscope. A TLS with 13 dBm input power was coupled to the circuit, and an erbium-doped fiber amplifier (EDFA) was used after the output to overcome in/out losses. For a 1 μm wide device in back-to-back transmission, open-eye diagrams were measured for 12.5, 25, 40, and 50 Gbps at 1550 nm with an on–off keying non-return to zero (OOK-NRZ) modulation signal. The received signals were post-processed using a feed-forward equalizer, and corresponding diagrams are shown in Figs. 6(c)–6(h). The eye diagrams had a dynamic extinction ratio (DER) from 5.6 to 3.1 dB from 10 to 50 Gbit, respectively. The input swing was 2.7 V, but with pre-compensation, the expected voltage swing to the device would be less. The signal-to-noise ratio (SNR) was 16.7 dB but reduced at higher baud rates, being 10.6 and 8.3 dB at 40 and 50 Gbps. Taking data transmission at 1560 nm, where the $Δ α$ is maximum, the (SNR) was 16 and 11 dB for 40 and 50 Gbps, respectively.

In conclusion, a high-performance μTP III–V EAM on 220 nm SOI has been designed, fabricated, and measured. The EAM has a maximum ER of 30 dB and an IL of 6.5 dB at 1550 nm. The spectral bandwidth is over 30 nm at −8 V. These devices can reach a cutoff frequency of 40 GHz in a lumped configuration with 50 Gbps demonstrated. The heterogeneous integration is completed using direct bonding of μTP. In the future, these devices can be impedance matched and include traveling wave configuration. The bond-pad contacts could also be included on the coupon.⁴⁰ Given the scalable integration method, this device could be used as an alternative to GeSi EAMs. InP MQWs have can deliver high-performance phase modulation, which can be integrated using the same method as proved here. The EAMs can be integrated in a scalable manner with other devices onto SiPh platforms, allowing for compact functionality. Ultimately, the dense integration of multiple transfer printed active devices for high-performing PICs represents the main goal for future work.

This work was supported by SFI through IPIC (12/RC/2276-P2), (15/IA/2864), (22/FFP-A/10930), and the PIADs programme. The authors would also like to acknowledge the Tyndall National Institute fabrication facility and open-access lab together with Brendan Sheehan for experimental software.