Design and fabrication of the devices

The ring waveguides of the AlGaAsOI resonators were designed to work within the normal dispersion regime in the C band, with dimensions of 400 nm × 1,000 nm. The width of the bus waveguide at the facet was designed to be 200 nm for efficient chip-to-fibre coupling. The fabrication of the AlGaAs microresonators was based on heterogeneous wafer bonding technology. The process is currently realized at the 100-mm-wafer scale without any strict fabrication processes such as chemical-mechanical polishing or high-temperature annealing that are not compatible with the CMOS process. It can therefore be directly adopted by current III–V/Si photonic foundries51. A Q factor >2 million can be obtained in the AlGaAsOI resonator, corresponding to a waveguide loss of <0.3 dB cm−1. The fraction of aluminium is 0.2, which corresponds to a two-photon absorption wavelength of around 1,480 nm. The epitaxial wafer growth was accomplished using molecular-beam epitaxy. A 248-nm deep-ultraviolet stepper was used for the lithography. A photoresist reflow process and an optimized dry etch process were applied in waveguide patterning to minimize waveguide scattering loss. More fabrication details can be found in refs. 52,53. The SiPh PIC, including its Si modulators and Si–Ge PDs, was fabricated on a 200-mm SOI wafer with a Si-layer thickness of 220 nm and a buried oxide layer thickness of 2 μm using CMOS-compatible processes at CompoundTek Pte in a one-to-one 200-mm-wafer run with its standard 90-nm lithography SOI process. The waveguide loss in this SiPh platform is approximately 1.2 dB cm−1 in the C band. In our experiment, lensed fibres with different mode field diameters were selected for the AlGaAsOI and SOI chips; the coupling loss is about 3–5 dB per facet for AlGaAsOI waveguides and about 2–3 dB per facet for Si waveguides.

Characterizations of the building-block units

The FSR of the 144-μm-radius rings utilized in this study is about 90 GHz. The microcomb shows advances both in start-up and stabilization. During the dark-pulse generation, a much smaller abrupt power change occurs when the comb transits from continuous-wave states to dark-pulse states, indicating the elimination of the well known triggering problem in bright soliton generation. Compared with general bright solitons, the dark pulse is inherently tolerant to thermal effects that usually make bright soliton states difficult to access54. For long-term stability measurement, the comb spectra and comb line power of a free-running dark-pulse comb are recorded by a high-resolution optical spectrum analyser (OSA) every 5 min.

More details are presented here for the SiPh devices shown in Fig. 2. The opto-electrical BW of the depletion-mode Si-based MZMs was measured by a vector network analyser (Keysight N524), with the typical results of >30 GHz. The on-chip phase compensation units are MZI-based titanium nitride (TiN) microheaters. The resistance is approximately 200 Ω. The TiN metal layer is about 1 μm above the Si layer, ensuring a heating efficiency of about 20 mW π−1. Meanwhile, a deep trench process is utilized to isolate each microheater to diminish thermal cross-talk. For the on-chip true-time delay line, we adopted a 2-μm-wide multimode Si waveguide for low-loss transmission. Euler curves were used in the spiral waveguide for adiabatic bending. For a 60-ps Si delay line, the total loss is <0.5 dB, with a delay-time variation of <3% among 8 tested devices. For the vertical epitaxial Ge PD, the responsivity declines with the increasing on-chip power. A saturated point of about 20 mW could be reached when the power is further increased. Microring filters employed for WDM could be tuned by microheaters, with which a 180-GHz channel spacing can be obtained under 20-mW power dissipation. The CMOS drivers for signal amplification before injection into the Si MZM (not used in the high-bit-rate (>50 Gbps) signal transmission experiment) show a 3-dB gain BW of about 24 GHz.

The performance of other building-block devices is presented in Extended Data Fig. 1. The linewidth of the DFB laser used as the pump is measured by a delayed self-heterodyne method55. The measurement and Lorentzian fitting result are shown in Extended Data Fig. 1a, exhibiting a laser linewidth of about 150 kHz. For the SiPh devices, the 3-dB BW of the Si–Ge photodiodes is shown in Extended Data Fig. 1b, indicating an approximately 30-GHz S21 parameter. Such a non-optimized BW accounts for the penalty in Fig. 3c. Structure design for a lower resistor-capacitor time constant could further increase the operation BW. For on-chip monitoring, the asymmetric MMI-based 10:90 power splitter56 is employed in the system, as shown in Extended Data Fig. 1c. The symmetry of the multimode region is broken by removing the corner of the MMI (marked with a red dashed rectangle), which causes a dramatic redistribution of the optical field, thus leading to an uneven power splitting by changing the width of the cut-off corner. We randomly chose four identical 1:9 MMIs and tested the power splitting ratios. The results were found to be close to the design target (dashed line), exhibiting good consistency, as shown in the bottom panel of Extended Data Fig. 1c. Moreover, the grating couplers used in this work (Extended Data Fig. 1d) show a roughly 2-dB coupling efficiency difference across the operation band (1,535–1,565 nm).

Turnkey dark-pulse microcomb generation

The turnkey microcomb generation test setup is shown in Extended Data Fig. 2a, with either an ECL or a DFB laser as the pump. Slow laser-frequency detuning is enough for microcomb generation, which can be realized by adjusting the cavity length via tuning the lead zirconate titanate voltage of the commercial ECL or changing the laser current of DFB, respectively. After the comb generation, the spectra are recorded; meanwhile, the total power of the generated comb lines is measured in real time. A pre-calibration process is required to ensure the laser frequency will locate at the comb accessing range ultimately. For the ECL-pumped dark-pulse comb (Extended Data Fig. 2b), a 1-Hz square wave is used as the trigger signal, which tunes the pump wavelength about 0.3 nm away from or into the resonance. For the DFB-pumped experiments (Extended Data Fig. 2c), when a laser is turned on, there is always an automatic frequency ramping-up process owing to the injected carrier and the warming of the cavity, which can directly initiate the microcomb generation as long as the lasing frequency of the final stable state lies within the range of the access window of the coherent state. In our experiment, the laser current is switched between two values with a period of 6 s (1 s for the ‘off’ state and 5 s for the ‘on’ state). Both results show immediate on–off behaviours of microcomb generation along with the low-speed control signal. It is noted that there is some power ripple of the DFB-pumped comb in the first few seconds, which is due to the temperature vibration caused by thermoelectric cooler, after which the comb state is stabilized. The comb is reproducible in several consecutive switching tests, with great robustness.

Details of data transmission experiments

In our experiment, the microcomb is first pumped by a commercial tunable laser (Toptica CTL 1550), then by a DFB laser chip for a higher degree of integration, where an optical isolator is deployed between the DFB laser and the AlGaAsOI microresonator to eliminate the reflection. When tuning the pumping wavelength from the blue side to a certain detuned value at around 1,552.5 nm, both configurations generate dark pulses with 2-FSR comb spacing. The detailed experimental setup for data transmission is shown in Extended Data Fig. 3a. For the comb spectrum with large power fluctuations, an additional amplification process is required owing to the insufficient gain of those low-power channels, which introduces extra system complexity and power consumption on the transmitting side. In this work, owing to the strong thermal effect, the avoided mode-crossing (AMX) strength of the AlGaAs microresonator can be thermally pre-set to obtain a coherent microcomb with a less disparate power distribution across the operation band. Thus, only a notch filter is required to attenuate the central three comb lines for the subsequent equalized comb amplification. The comb is amplified by an EDFA and then split into odd and even test bands39,57,58 by a wavelength-selective switch (Finisar Waveshaper 4000s). A Si modulator and a lithium niobate (LN) modulator (EOspace, 35-GHz BW) are deployed at the odd and even bands, respectively. Ten comb lines in each test band are simultaneously modulated. The modulators are driven at a 32-Gbaud or 50-Gbaud symbol rate. The differential PAM-4 signal is generated by a commercial pulse pattern generator (Anritsu PAM4 PPG MU196020A). The insertion loss of the SiPh (LN) modulator is 13(8) dB. The SiPh modulator undergoes a relatively high loss (including the edge coupling loss of about 2 dB per facet), which results in a power difference between the two test bands. The modulated test bands are then combined by a 50:50 power coupler and launched into another WSS for comb power equalization. At the receiving side, each WDM channel encoded by the Si modulator is sequentially filtered out and measured. Eye diagrams are produced by a sampling oscilloscope (Anritsu MP 2110A) with a 13-tap transmitter and dispersion eye closure quaternary (TDECQ) equalizer (accumulation time, 8 s). The BERs are measured online by an error detector (Anritsu PAM4 ED MU196040B) with 1-dB low-frequency equalization and a decision-feedback equalization. Extended Data Fig. 3b shows the 100-Gbps PAM4 eye diagrams for each of the 20 channels.

It is worth noting that the performance is underestimated. In our proof-of-concept test configuration, ten channels in each test band are modulated at the same time. Considering two-photon absorption in Si waveguides, the maximum input power for the Si modulator is about 13 dBm, which results in only 3-dBm optical power per single lane. Moreover, considering the extra penalty introduced by the WSS for power equalization, unnecessary in real-word transmission scenarios, the OSNR for each channel can be at least 10 dB higher. Thus, a better transmission result is attainable.

Noise analysis of different pump schemes

The noise floor of the DFB and the ECL are roughly characterized in an OSA, as shown in Extended Data Fig. 4a. The laser spectra indicate that the noise of the DFB is evidently higher than that of the ECL. The combs in our experiments are pumped by the free-running DFB laser and the ECL separately, as shown in Extended Data Fig. 4b, c. With the almost same pumping power of about 10 mW, the DFB chip holds a 10-dB-higher noise floor compared with the ECL, corresponding to an equivalent OSNR reduction in each comb line. Moreover, the amplification after the comb generation would also result in OSNR degradation, which could be a potential problem when replacing the current EDFA with integrated SOAs (about 4–5-dB-noise-floor increment in a commercial EDFA and about 7 dB in commercial on-chip SOAs).The OSNR of the DFB-pumped microcomb can be further improved by employing an on-chip optical filter for comb distillation59,60 or introducing optical injection locking between the microcomb and slave lasers for low-noise amplification61. Also, increasing the pump power will lead to a higher average OSNR and more stable long-term behaviour, which is an advantage over the injection-locking-based dark-pulse generation21,62.

Setup of the dispersive delay-line MPF scheme

As the non-uniformity of delays owing to the inevitable fabrication errors will degrade the filtering performance, the second TDL-MPF approach is also implemented to further determine the optimal filtering performance: a spool of single-mode fibre (SMF) is used instead of the on-chip spiral delay lines to produce dispersive delay. Extended Data Fig. 5 shows the experimental setup of the reconfigurable MPF carried out in a dispersive delay-line configuration. Compared with Fig. 4a, most of the MPF system remains unchanged and has one main difference, which is that the on-chip true-time spiral delay lines are removed from the SiPh signal processor. The processed comb lines will propagate through a spool of 5-km SMF (as a dispersive element) to obtain a solid delay unit between adjacent taps, which can be expressed as T= δλDL (ignoring the high-order dispersion of SMF), where δλ represents the comb line spacing, D is the dispersion coefficient of SMF and the L is the length of SMF. In this scheme, the basic delay T among comb lines is generated by a single dispersive element, which can be kept as uniform value and not influenced by fabrication errors. Besides, this system is more flexible; for instance, the centre frequency of the filtering passband can be adjusted by simply change the length or dispersion coefficient of SMF.

Details of RF filter experiments

The DFB-driven dark-pulse Kerr comb exhibits 2-FSR (180-GHz) comb spacing. The initial comb source is amplified by an EDFA, and 8 comb lines in the range of 1,547–1,560 nm are selected using an optical bandpass filter before injection into a SiPh signal processor chip. The input and output coupling are achieved via grating couplers of about 40% coupling efficiency. Frequency-swept RF signals with 9-dBm power from a vector network analyzer are applied to the Si MZM in double-sideband format. The tap weighting coefficients are set by adjusting the relative detuning among the comb lines and their corresponding resonance wavelengths in the Si MRA with TiN microheaters placed on the waveguides. The output light of the Si chip is split by a 10:90 optical power coupler: 10% of the light is sent into an optical spectrum analyser (Yokogawa AQ6370C) for spectral monitoring, whereas the other 90% of the light propagates through the follow-up optical link. In the dispersive delay scheme, a spool of 5-km SMF is used to acquire the dispersive delay between adjacent comb lines (taps). Finally, the processed comb lines are beat in a 50-GHz PD (Finisar 2150R) to convert the optical signal into electrical domain. A low-noise EDFA is placed before the PD to compensate for the link insertion loss and coupling loss.

For the practical demonstrations of RF signal filtering, a 50 Gsamples s−1 arbitrary waveform generator (AWG, Tektronix AWG70001) is used to produce the desired RF input signals. To validate the BW reconfigurability of this filter, an ultrawideband RF signal is generated, spanning from 5.5 GHz to 9 GHz. To validate the FSR reconfigurability of this filter, a complex RF signal is produced that contains a 50-Mb-s−1 QPSK spectrum modulated at 3.6 GHz and a 50-Mb-s−1 QPSK spectrum modulated at 7.2 GHz. The RF outputs from the AWG are amplified by a linear electrical driver (SHF 807C) before routing to the Si MZM. The filtered RF signals are detected by a signal analyser (Keysight N9010B) for spectrum measurement. A similar FSR multiplication of the MPF has been reported previously and explained by temporal Talbot effects63. However, the crucial Talbot processor used in these MPF systems is based on more complex discrete devices, which will increase the power dissipation and make the system less stable.

Unlike the conventional waveshaper based on bulky liquid-crystal spatial light modulators64, one of the remarkable advantages of the chip-scale add-drop microring resonator (MRR) array used in our work is the rapid reconfiguration of RF filtering responses. The reconfiguration operation on filtering spectra is realized by adjusting the shaping profiles of comb lines, through the TiN microheater placed on the waveguides. To explore the maximum reconfiguration speed, a standard electrical square-wave waveform is generated by a function waveform generator (RIGOL, DG2102) to drive a single MRR channel. The output of the MRR is received by a photodetector (Thorlabs DET08CFC/M), and then recorded by a digital oscilloscope (RIGOL, DS7014 10 GSa s−1). Extended Data Fig. 6 shows the measured switching temporal response. As seen in Extended Data Fig. 6b, c, the 90/10 rise and fall times are 15 μs and 53 μs, respectively. Therefore, the fastest response speed for the reconfiguration operation is approximately 19 kHz.



Source link