3D integration enables ultralow-noise isolator-free lasers in silicon photonics – Nature

3D mode transitions

Using a series of adiabatic tapers⁵⁶, the optical mode is transferred over a vertical distance exceeding 4.8 μm between an InP/Si hybrid mode in the laser active region, to the ULL SiN waveguide layer of the ultrahigh-Q resonator. The mode is transferred first from the InP/Si hybrid waveguide into a Si waveguide and subsequently from the Si waveguide into a SiN RDL. As these InP, Si and SiN layers are fabricated either in contact (in the case of InP and Si) or in close proximity (in the case of the Si and SiN RDL), the optical mode transfers rapidly between them and their tapering lengths are less than 300 μm in total. In particular, the InP to Si rib waveguide transition can be very short (around 25 μm), as InP and Si feature similar refractive indices⁴⁵. The Si rib waveguide with 231-nm etch depth is subsequently tapered to a 200-nm width to transfer the mode to the thin Si waveguide with 269-nm thickness. The thin Si tapers from around 3 μm to 150 nm to match the effective index of the RDL SiN waveguide for efficient Si–SiN power transfer. To span the vertical distance separating the SiN RDL and the ULL SiN layer on which the high-Q resonators reside, the optical mode is gradually evolved from the upper SiN to the lower SiN layer. The RDL SiN and ULL SiN waveguides feature identical core thicknesses of 100 nm, so that their effective indices are readily matched. The RDL SiN waveguide width is thus tapered from 2,800 nm to 200 nm, while simultaneously widening the ULL SiN waveguide width from 200 nm to 2,800 nm, over a distance approaching 1 cm in length. This scheme enables efficient power transfer (<1-dB insertion loss) from the RDL SiN waveguide to the ULL SiN waveguide.

In weakly confined ULL SiN waveguides, the optical mode extends significantly into the silicon dioxide (SiO₂) cladding. Previous work³¹, in which a ULL SiN waveguide was heterogeneously integrated in close vertical proximity with an InP/Si hybrid waveguide, resulted in a relatively high propagation loss of 0.43 dB cm⁻¹. In this work, a 3D layer transition enables the ULL SiN waveguide to be buried deeper within the SiO₂ cladding, such that impurities originating from the back-end heterogeneous integration process do not influence the ULL waveguide performance.

To motivate the addition of a SiN RDL in such a 3D layer transition, we compare the performance of a Si-to-SiN waveguide transition with a SiN-to-SiN waveguide transition. The optimal length of an adiabatic coupler transition is given by \({L}_{{\rm{opt}}}\approx \frac{3}{2}\frac{1}{\kappa \sqrt{{\epsilon }}}\), in which κ represents the coupling coefficient between the waveguides in the coupling region, and ϵ = 0.01 represents the tolerance of power transferred to the undesired, anti-symmetric system mode (that is, loss)⁵⁶. This optimal (minimum) length is calculated as a function of vertical separation in Extended Data Fig. 1a, which demonstrates that a SiN-to-SiN layer results in a more efficient (shorter) transition for vertical separation exceeding 2 μm.

Thus, the inclusion of a SiN RDL beneath the Si waveguide provides improved vertical coupling efficiency, enabling the ULL SiN waveguide to be buried deeper below. The SiN RDL is further motivated by additional performance and fabrication concerns. Efficient power transfer between Si and SiN waveguides requires a very narrow Si width to match the propagation constants of the respective waveguides, as shown in Supplementary Fig. 1b. Such narrow Si waveguides feature significant sidewall-roughness-induced scattering loss, limiting the length of such structures. Furthermore, the combination of narrow width and long length of a Si-to-SiN transition capable of spanning several-micrometre distance would yield a fragile structure that is susceptible to damage during the fabrication process. As such, the close proximity of the SiN RDL to the Si waveguide enables a short Si-to-SiN transition, improving process yield. In this work, the SiN-to-SiN transition length was chosen to be excessively long (approaching 1 cm) to enable flexibility in the choice of vertical separation while retaining transition efficiency.

To experimentally evaluate the achievable transition efficiency from the RDL SiN to the ULL SiN, two such layer transition structures were placed within a racetrack resonator. In contrast to a cut-back approach, in which multiple identical structures are cascaded in series to extract an aggregate insertion loss, a resonator-based measurement technique enables insertion loss of a structure to be measured independently of fibre-to-chip coupling losses, resulting in a more accurate measurement. Previous work has demonstrated this approach to accurately measure insertion losses well below 0.1 dB (ref. ⁵⁷). The transmission spectrum of the resonator was measured and fitted to extract the internal round-trip loss, as shown in Supplementary Fig. 2. From this measurement, the insertion loss was inferred to be below 0.03 dB per transition. However, the resonator-based test structure was fabricated on a separate wafer that did not undergo any heterogeneous integration process and featured a narrower spacer thickness of roughly 3.5 μm. Thus we conservatively expect the insertion loss of the RDL-to-ULL transition within the heterogeneous laser to be well below 1 dB.

Device fabrication

Fabrication of the SiN waveguides was performed at Tower Semiconductor, a commercial CMOS foundry, on a 200-mm-diameter Si wafer with 15-μm-thick thermal SiO₂. Low-pressure chemical vapour deposition SiN with 100-nm thickness was deposited and patterned using deep ultraviolet (DUV) stepper lithography and reactive ion etching to form the ULL waveguide layer. Tetraethyl orthosilicate-based oxide was deposited on the ULL layer, annealed at 1,150 °C, and underwent chemical mechanical polishing to form an approximately 4-μm-thick spacer layer³⁵. To form the RDL waveguide transition, another 100-nm-thick low-pressure chemical vapour deposition SiN was deposited and patterned by the same process. The adiabatic RDL taper was defined and etched on this layer. Additional tetraethyl orthosilicate-based oxide was deposited, annealed and underwent chemical mechanical polishing to leave around 500-nm-thick SiO₂ on top of the RDL. The processed 200-mm wafer was then transferred out of the foundry for subsequent processing. The wafer was cored into 100-mm wafers to be compatible with an ASML 248-nm DUV stepper. Diced silicon-on-insulator pieces with 500-nm-thick Si device layer were bonded on the polished SiO₂ surface using plasma-activated direct bonding. The Si substrate was removed by mechanical polishing plus deep Si Bosch etching. The buried SiO₂ layer was removed by buffered hydrofluoric acid. The fabricated Si/SiN RDL/SiN ULL wafer was then ready for the heterogeneous InP process on Si similar to our previous studies³¹. In general, Si waveguides and tapers were patterned with a DUV stepper whereas the grating was patterned with electron beam lithography with a period of 240 nm. The Si layer underwent several patterned etches with different etch depths. The first etch had a 231-nm etch depth to form Si shallow etched rib waveguides in the InP/Si and phase-tuner sections. Then the Si gratings and thin Si tapers were formed respectively with 269-nm etch depth. Si outgassing channels were patterned later and etched with an etch depth of 500 nm in the area that had no Si waveguides. The Si etch was reactive-ion-etched with a mixed etching gas of C₄F₈/SF₆ and the etch depth was controlled by an etch monitor Intellemetrics LEP400. After Si processing, InP dies with the layer stack shown in Fig. 1c were bonded on the fabricated Si circuits, with the InP substrate removed by mechanical polishing and 3:1 hydrochloric acid:deionized water. A thin layer of p-type contact-metal Pd/Ge/Pd/Au was formed using a lift-off process. The p-InP mesa was etched using CH₄/H₂/Ar, with a SiO₂ hard mask. The dry etch was monitored using an etch monitor and stopped at the AlInGaAs  quantum well (QW) layer. After another round of QW layer lithography, the QW layer was etched using a mixed solution of H₂O/H₂O₂/H₃PO₄ 15/5/1. An n-type InP mesa etch followed the QW etch to complete the mesa definition with the same etching gas CH₄/H₂/Ar. The excess Si on top of the SiN devices was removed using a XeF₂ isotropic gas etch. The entire chip was passivated using low-temperature deuterated SiO₂ (ref. ⁵⁸) followed by the contact-metal window opening through CF₄-based inductively coupled plasma etching. The n-type contact-metal Pd/Ge/Pd/Au and another layer of Ti/Au on top of the p-type contact metal were deposited and formed. Proton implantation was performed to define the current channels. Ti/Pt was deposited as heaters for the phase tuner on Si and resonance tuner on SiN. The chip went through another round of SiO₂ deposition and contact via opening. The Ti/Au probe metal was deposited to finish the wafer fabrication. The fabricated 100-mm-diameter 3D PIC wafer was diced and polished to expose the SiN edge couplers for fibre-coupled device characterization. The detailed process flow charts can be found in Extended Data Fig. 7.

Impurity depth profiling analysis using secondary ion mass spectrometry

To analyse the impurity distribution along the depth direction (depth profiling), a secondary ion mass spectrometry system (also known as ion microprobes, CAMECA IMS 7f) was used to analyse the devices. In the measurement, a raster area of 50 μm × 50 μm was swept with the primary beam (for ionization and sputtering) and secondary ions generated only in the centre area of 20 μm × 20 μm were collected by the instrument filtering aperture to prevent impacts from other layers at the edge of the hole drilled. To obtain conductivity required for secondary ion mass spectrometry, 20 nm of gold was deposited onto device surfaces. Reference devices NIST SRM 610 and 612 (ref. ⁵⁹) (National Institute of Standards and Technology Standard Reference Materials (NIST SRM)) were used for the calibrations of elementary concentrations. The measurement was implemented in a vacuum level of 3 × 10⁻⁹ torr. For the positive-ion measurements, O⁻ ions were the primary beam. For the negative-ion measurements, Cs⁺ ions were the primary beam, and the electron beam was also engaged to neutralize the sample to avoid charging effects. The results are plotted in Extended Data Fig. 3.

The sample area measured here is a pure waveguide region without the top laser structure but experienced the full back-end-of-line process. The appearance of boron atoms indicates the boundary of the lower thermal oxide cladding layer because the substrate Si wafers are of p-type (resistivity of about 100 Ω cm) to accelerate thick thermal oxidation. The appearances of both Si–N clusters and C–N clusters indicate the thin SiN waveguide layer because nitrogen atoms begin to appear in large amounts. The coincidence of B, Si–N and C–N traces cross-verify each other and gives the SiN waveguide depth position of 5.9 μm.

Additional laser characterization

The lasers are characterized on a temperature-controlled copper stage with a precision temperature controller (Vescent SLICE-QTC) for device characterization at 20 °C. We screened the lasers before the self-injection-locking characterization, phase-noise measurement and so on. The laser light-current measurement results are shown in Extended Data Fig. 4a, which exhibit an approximately 74-mA laser threshold, influenced by the DFB grating strength. Compared with typical laser light-current behaviours, one difference of such a laser-resonator device is that with the increase of laser gain current, the recorded power would see several dips in the light-current curve when the laser output power is filtered by the ring resonator. The spacing of the resonance dips is determined by the ring resonator FSR (30 GHz in this work) when the laser wavelength is swept across multiple resonances during the gain-current increase. It has to be noted that in this light-current sweep, the laser gain current is stepped at 1 mA so not every resonance can be matched and recorded.

We can thus lock the laser to different resonances by tuning the laser gain current. Besides, the thermal tuning of ring resonances allows the continuous tuning of the SIL laser wavelengths across the DFB laser wavelength. This capability is critical in microwave generation as microwave frequency can be synthesized precisely based on the laser gain and ring resonance controls. We lock two SIL lasers at two resonances with over 3-nm-wavelength space and the laser spectra are shown in Extended Data Fig. 4b. This wavelength separation promises >375-GHz millimetre-wave generation if a fast PD is available. More importantly, the phase noise will be the same as low carrier frequencies as it is determined by the laser phase noise.

Laser self-injection locking

By tuning the laser wavelength to a resonance from the ring, the backscattered light from the ring locks the laser wavelength to the resonance provided that the phase of the backscattered light arriving at the laser is an integer multiple of 2π of the forwards laser output phase. In other words, the wavelengths of the laser and the resonance are matched in the frequency domain whereas the phases of the laser and the backscattered light are matched in the time domain, as shown in Fig. 2a. Matching the wavelengths is performed by tuning either the laser gain current or the ring heater current, whereas matching the phases is done by tuning the phase-tuner current. Both laser gain current and phase-tuner current are driven with low-noise laser current sources (ILX Lightwave LDX-3620) to ensure stable and low-noise operation. Detection of the self-injection-locking state is assured by observing not only the decrease in the output power from the ring when the laser wavelength hits the resonance but also the decrease in the linewidth of the self-heterodyne beat as the self-injection locking takes place, as shown in Fig. 2b. The self-heterodyne interferometer set-up consists basically of a Mach–Zehnder interferometer (made from two 3-dB couplers) with a polarization controller and a short delay line in one of its arms and a fibre-coupled acoustic-optic modulator (Gooch & Housego 27 MHz) in the other arm, as shown in Fig. 2b. The beat frequency from the self-heterodyne interferometer is detected using a PD (Newport 1811) before it is sent to an electrical spectrum analyser (ESA) (Rohde & Schwarz FSWP). The phase-tuner current is roughly adjusted during the self-injection-locking process to allow the locking to occur and finely tuned afterwards to ensure stable self-injection locking. It is worth mentioning here that the self-injection-locking state can last for hours without even packaging the laser chip. This can be attributed to the integration of the laser and the resonator on the same chip, which reduces the phase fluctuation between the laser and the backscattered light from the ring.

SIL laser characterization

The dynamics of the phase-tuner influence on self-injection locking is investigated by sweeping its applied electrical power (Keithley 2604B) over three 2π injection-locking periods while recording the ESA spectrogram of the detected self-heterodyne beat of the SIL laser (Fig. 2c, top). The spectrogram of the injection-locking periods, which is depicted in Fig. 2c, demonstrates stable SIL periods (dark blue regions) followed by chaotic regions (light blue) and then unlocked regions. The laser power is also detected on an oscilloscope (Tektronix MSO64B) during the phase tuning over only one period, which clearly shows the mentioned behaviour (Fig. 2c, bottom). Another important parameter is the frequency range at which the self-injection locking persists. It can be obtained by either sweeping the laser frequency over the ring resonance or sweeping the ring resonance over the laser. We selected the second scheme by sweeping the current of the ring heater using a triangle signal applied to the current source (Keithley 2604B). To detect the change in laser linewidth during sweeping, a beat is made using a 3-dB coupler between the SIL laser and a narrow-linewidth fibre laser. The beat is optically amplified with an erbium-doped fibre amplifier (Amonics AEDFA-IL-18-B-FA) and sent to the fast PD (Finisar HPDV2120R) that is connected to the ESA, as shown in the lower branch of Fig. 2b. The recorded spectrogram during the resonance sweep is shown in Fig. 2d. The laser frequency noise and resultant fundamental linewidth are taken from a commercial laser phase noise analyser (OEwaves OE4000) that internally performs averaging over the measured phase noise. We have not observed significant differences in our noise spectrum between 1 kHz and 1 MHz for several measurement runs, which are very stable, and believe that the noise spectrum in this range is dominated by the thermorefractive noise of the resonator by comparing with the simulation results (Fig. 2e). As a comparison, the delayed self-heterodyne set-up uses two PDs to receive the heterodyne beat⁶⁰, which has been previously used for ultralow-noise laser linewidth characterizations³⁶ and allows for a more detailed analysis of statistical measurement errors.

Laser feedback sensitivity measurement

Figure 3b schematically depicts the experimental configurations for analysing the laser feedback sensitivity. The coupled laser emission is sent to a 90/10 fibre beamsplitter, after which 90% of the coupled power will be used for external optical feedback. The feedback loop consists of an 8-m-long single-mode fibre, a three-port optical circulator, a polarization controller and a variable optical attenuator (VOA) that allows for an attenuation ranging from 0 to 40 dB (EXFO MOA-3800). It should be noted that the laser feedback sensitivity also depends on the polarization of the reflected field, which must be adjusted to maximize the feedback influence before running the analysis. The remaining 10% of laser output is utilized for the feedback sensitivity characterization. After passing through an optical isolator, it is transferred either to a phase noise analyser (OEWaves OE4000) for frequency-noise characterization or a delayed self-heterodyne set-up for the electrical spectrum-based laser coherence check.

In this study, the feedback strength is determined by the reflected power (P_refl) and the output power (P_out) through the following relationship:

All losses from the feedback loop should be considered to calculate the reflected power, thus the feedback strength. After optimizing the set-up, the fibre-chip coupling loss is −3 dB (round-trip coupling loss is −6 dB), the total losses from the 90% beamsplitter, the optical circulator, the insertion loss of the VOA, the polarization controller and the fibre is −4.05 dB. The feedback strength that accounts for the attenuation of VOA can thus be tuned from −10.05 dB to −50.05 dB.

To further reduce the loss from set-up and thus maximize the feedback strength as large as possible, we use a 100% fibre back-reflector (BKR, Thorlabs) to replace the configurations after the 90% beamsplitter port. The loss from the feedback loop is then reduced to −0.9 dB, and the maximum feedback strength is −6.9 dB.

Microwave-signal generation

Two lasers are SIL to two 30-GHz-FSR ring resonators that are shifted in frequency by 10 GHz without tuning on the ring resonance. Although each ring resonator can be tuned over 30 GHz by applying around 0.5 W of electrical power to the heater on the ring, we used only one ring heater for tuning. The tuning range for the first FSR is −10 GHz to 20 GHz, whereas the next full FSR tuning leads to 20 GHz to 50 GHz tuning by locking the second laser to the next resonance and hence covering the full 50 GHz. As shown in Fig. 4b, the two lasers’ outputs are collected from the chip by the fibre V-groove array and sent to a 3-dB fibre coupler, then the erbium-doped fibre amplifier before beating on a fast PD (Finisar HPDV2120R) connected to an ESA. A small portion of the laser output (1%) is sent to an optical spectrum analyser (Yokogawa AQ6370C) for monitoring if the lasers are on the intended resonance. Although our chip could be used to generate any arbitrary microwave frequency over 50 GHz, it is used here to generate microwave frequencies at steps of 1 GHz over the full 50 GHz for demonstration (Fig. 4c). An offset phase-locking servo circuit (Vescent D2-135) is used to improve the long-term stability of the generated microwave signals for frequencies up to 10 GHz, by locking the phase of one of the lasers to the other one. The feedback signal from the servo control box is sent to one of the laser’s ring heaters to lock its phase to the phase of the second laser. Stable microwave signals are thus generated with low-phase-noise characteristics of the SIL lasers and the Voigt fitting is plotted in Extended Data Fig. 6.