Nonlinear optics as a means to achieve ultrafast all-optical signal processing has been extremely powerful, and has focused on photonic integrated circuit platforms based on highly nonlinear materials such as silicon^[1-3]. The range of signal processing operations that can be performed with nonlinear optical techniques is very large, and includes optical logic^[4], optical temporal demultiplexing at ultra-high bit rates from 160 Gigabits/s^[5] to well over a terabit per second^[6], optical performance monitoring based on slow light^{[7, 8]}, signal regeneration^{[9, 10]}, and many other functions^[11-16]. Complementary metal oxide semiconductor (CMOS) compatible platforms are extremely important since they can exploit the extensive global infrastructure established to fabricate computer chips. Since all CMOS compatible material platforms are centrosymmetric, the 2nd order nonlinear response is zero and so nonlinear devices in these platforms have been based on the 3rd order nonlinear susceptibility that accounts for processes including optical third harmonic generation^{[11, 17-21]} as well as the instantaneous intensity dependent refractive index termed the Kerr nonlinearity (n₂)^{[1, 2]}. The efficiency of all-optical devices based on the Kerr nonlinearity depends on the waveguide nonlinear parameter (γ), and although silicon waveguides and nanowires have achieved extremely high nonlinear parameters (γ), they also display quite high nonlinear optical loss arising from indirect two-photon absorption (TPA) in the telecommunications wavelength region (near 1550 nm), since the indirect bandgap of silicon is only 1.17 eV which is much less than twice the photon energy at 1550 nm. Moreover, worse than the intrinsic TPA is the effect of the resulting generated free carriers^[2] since the carrier recombination time in silicon can be very long (microseconds), enabling significant carrier densities (and hence loss) to build up. Even if the effect of these carriers can be minimized through the use of p–i–n junctions to sweep out the carriers and greatly reduce the effective carrier lifetime, the intrinsic material nonlinear figure of merit [FOM = n₂/(βλ), where β is the TPA and λ the wavelength] of silicon is only 0.3 in the telecom band. This is too low to achieve high all-optical performance and is a fundamental property of silicon’s band structure and therefore cannot be improved. While TPA can be turned around and actually used as a positive tool for some nonlinear functions^[22-24], for most processes silicon’s low FOM in the telecom band poses a significant limitation. This has been the motivation for research into alternative nonlinear optical platforms, such as chalcogenide glass^[25-34], for example. However, although many of these platforms offer significant advantages, most of them are not CMOS compatible, which ultimately is a fundamental consideration to achieve widespread and efficient manufacturing.

New CMOS compatible platforms for nonlinear optics were introduced in 2007/8 that exhibit negligible TPA in the 1550 nm wavelength regime. These include silicon nitride^{[35, 36]} and high-index doped silica glass, trade-named Hydex^[37-47]. In addition to having negligible nonlinear absorption even up to many Gigawatts of power, these platforms have a moderate Kerr nonlinearity, resulting in an extremely high nonlinear FOM with a nonlinear parameter (γ) that is high enough to realize significant parametric gain. Following the first report of a micro-resonator based optical frequency comb source driven by the Kerr nonlinearity in 2007^[48], the first fully integrated optical parametric oscillators based on micro-ring resonators (MRRs) were reported in 2010^{[36, 37]} in these new CMOS compatible platforms of SiN and Hydex. Since 2010 the field of integrated micro-combs, or Kerr combs has become one of the largest and most successful fields in optics and photonics^[47]. They are new and powerful tools to accomplish many new functions on an integrated chip, due to their very high coherence, while also offering flexible control of their wavelength spacing. Optical micro-combs are produced via optical parametric oscillation driven by modulational instability, or parametric gain, in integrated ring resonators. This offers huge advantages over more conventional approaches including discrete multiple laser wavelength sources. Many breakthroughs have been reported with Kerr micro-combs, from innovative mode-locked lasers^[49-52] to quantum optical photonic chips^[53-61], ultrahigh bandwidth optical fibre data transmission^[62-64], optical neural networks^[65-67], integrated optical frequency synthesizers^[68] and more, and have been extensively reviewed^{[47, 69-76]}. The success of these new CMOS platforms has motivated the search for even higher performing CMOS compatible platforms such as amorphous silicon^[77] and silicon rich silicon nitride^[78], searching for the ideal combination of low linear and nonlinear loss together with a high nonlinearity.

A significant application for microcombs has been in signal processing functions for telecommunications and RF/microwave systems, ranging from RF photonic applications to signal generation and processing for radar systems^[79-128]. RF photonics is attractive because of its ultra-high bandwidth as well as low transmission loss and strong immunity to electromagnetic interference. There are different RF photonic approaches including methods that map the optical filter response onto the RF domain including devices based on stimulated Brillouin scattering^[88-95], that have achieved high RF resolution — down to 32 MHz, as well as a stopband discrimination > 55 dB. A different but equally key approach to reconfigurable RF transfer functions for signal processing has been transversal filter methods ^[96-100]. These operate by generating progressively delayed and weighted replicas of an RF signal multicast onto many optical carriers, which are subsequently summed via photo-detection. Transverse filters can achieve a wide range of RF functions solely by varying the tap weights, and so this approach is very attractive for advanced dynamically adaptive RF filters. Discrete diode laser arrays^[101] and fibre and integrated Bragg grating arrays and sampled gratings^[103] have been used to generate the required taps. However, while offering advantages, these methods have increased complexity and footprint, limiting performance due to the limited number of wavelengths. Alternative methods such as electro-optic (EO) or acousto-optic (AO) combs^{[102, 104, 105]}, can help overcome this, but they require many high bandwidth modulators and high-frequency RF sources.

Kerr micro-combs^{[36, 37]} have been very successful in their applications to RF systems, providing advantages over other methods of supplying multiple wavelengths for RF photonic systems. They have achieved extremely high bandwidth data communications as well as a wide range of microwave signal processing devices^[107-128]. Their comb spacings can be much wider than electro-optic combs, and in many ways EO and micro combs are complementary. EO combs excel at finer spacings from 10’s of megahertz to 10–20 GHz, while integrated micro-combs typically have much wider spacings from 10’s of GHz to 100’s of GHz and even THz. Larger comb spacings have much wider Nyquist zone for large RF bandwidth operation, whereas smaller spacings provide many more wavelengths or RF “taps”, although at the expense of a smaller Nyquist zone. Micro-combs provide more wavelengths while still having a large free spectral range (FSR), all in a small footprint. For RF transversal filters the number of wavelengths determines the number of channels for RF true time delays and RF filters^{[85, 121]}. Systems such as beamforming devices^[112] can also be improved in quality factor and angular resolution. Other approaches to filtering include RF bandwidth scaling^[125] that yields a particular bandwidth for each wavelength channel, where the total bandwidth (maximum RF signal bandwidth) will depend on the number of wavelengths which is dramatically increased with micro-combs.

Recently^[121], we reviewed transversal filtering and bandwidth scaling methods based on Kerr micro-combs, as applied to RF and microwave spectral filters, followed by a review of temporal based signal processing^[129]. In this paper, we review recent progress in microwave and RF photonic based channelizers based on Kerr micro-combs. We cover devices achieved with both widely spaced micro-combs with an FSR of 200 GHz^[113] as well as results obtained with record low FSR micro-combs with a spacing of 49 GHz, operating via soliton crystals^[122]. We highlight their potential and future possibilities, contrasting the different methods and use of the differently spaced micro-combs. While 200 GHz Kerr micro-combs have been successful for RF channelizers, achieving high levels of performance, high versatility and dynamic reconfigurability, their large comb spacing limited the number of channels as well as the RF bandwidth that can be achieved without the use of thermal tuning. This is an important consideration because RF channelizers need many components such as optical amplifiers and spectral shapers that are only available at telecom wavelengths (1530–1620 nm). The limitation in the number of channels has restricted the overall RF bandwidth, frequency resolution, and dynamic reconfigurability of micro-comb based RF channelizers. To overcome this, we focused on RF channelizers based on record low FSR combs, with a spacing of 49 GHz, in order to achieve over 90 channels in the C-band^[122]. This represents the highest value for any micro-comb based RF channelizer. Our results confirm the feasibility of achieving high performance reconfigurable transversal RF filters for signal processing with reduced footprint, complexity, and cost. We first review the microcombs used to realize the RF channelizers, followed by the results for 200 GHz microcomb based channelizers and finally by the 49 GHz soliton crystal based systems.

The formation of micro-combs is a complex process that arises from a combination of a high nonlinear parameter together with low nonlinear as well as linear loss, and finally from careful engineering of the dispersion. Micro-combs have been realized in a variety of material platforms^[47] including magnesium fluoride, silicon nitride, silica, and doped silica glass^{[47, 70, 85]}. In 2008, efficient four wave mixing (FWM) operating at low (milliwatt) continuous-wave (CW) power levels was reported in a MRR with a 575 GHz FSR spacing, in a ring resonator with a low Q-factor of around 60 000^[39]. This was the first report of any form of nonlinear optics at milliwatt CW power levels in a silica glass-based platform. In 2010 this was followed by the first report of an integrated Kerr micro-comb^{[37, 38]}. Another key breakthrough came in 2017^{[116, 117]} with reports of integrated Kerr micro-combs with record small FSR’s under 50 GHz. This significantly increased the available number of channels, or wavelengths, to more than 80 in the telecommunications wavelength C-band spanning 1530–1565 nm. In addition to their low FSR spacings, these Kerr micro-combs operated in a different mode to dissipative Kerr soliton (DKS) states^[66-73], in a process that has been called soliton crystals^{[130, 131]}. Many other innovative states have been reported in micro-combs, including microcombs with extremely low threshold powers^[132], dark solitons^[133], laser-cavity solitons^[134] and others^[135-140].

The devices that were used in the work reviewed here were fabricated in Hydex glass^{[37, 38]}, a CMOS compatible platform categorized as high index doped silica. Micro-ring resonators with Q factors ranging from 60 000 to more than 1.5 million have been achieved, with radii ranging from 592 μm down to 48 μm, corresponding to FSRs in the range 49–575 GHz. The RF signal processors used as the basis for the work reviewed here were fabricated in MRRs with FSR spacings of 200 and 49 GHz. Films of Hydex glass (n = ~1.7 at 1550 nm) were deposited via plasma-enhanced chemical vapour deposition (PECVD) and photolithographically patterned with deep ultraviolet stepper mask aligners. Waveguides with very low surface roughness were created by reactive ion etching, after which an upper cladding of silica glass (n = ~1.44 at 1550 nm) was grown. We use both lateral and vertical bus-ring coupling schemes, with a typical gap of about 200 nm. Vertical coupling can be controlled by film growth more accurately than by lithography. Hydex glass has many advantages including a very low linear loss (~ 0.06 dB·cm⁻¹), a reasonably large nonlinear parameter (0.233 W⁻¹·m⁻¹), and probably most importantly, negligible TPA even up to many gigawatts cm⁻² of power. We achieved high Q factor resonators of ~1.5 million for both 49 and 200 GHz spaced MRRs. After packaging with fiber pigtails, the coupling loss can be as low as 0.5 dB per facet with the aid of on-chip mode converters.

For the 200 GHz devices, combs were generated by amplifying the CW pump power to more than 1 W (> +30 dBm). Next the wavelength was tuned from blue to red relative to the resonance wavelength, targeting a TE resonance near ~1550 nm. As the offset between the cold resonance and pump wavelength became sufficiently small, the intracavity pump power reached a threshold when modulation instability gain yielded oscillation^[47], first generating primary combs with their spacing determined by the MI gain peak, which is a function of dispersion as well as the intra-cavity pump power. As the detuning changed further, microcombs with a spacing equal to the FSR appeared. While these were not soliton states, nonetheless we found that they were more than adequate for our applications to microwave signal processing. We found that it was not necessary to strictly operate in soliton states such as the dissipative Kerr solitons (DKSs)^[66]. This is an important issue since, while much progress has now been made around DKS solitons^[140], they nonetheless intrinsically require quite complicated pump tuning dynamics, generally employing both simultaneous amplitude (power “kicking”) and wavelength sweeping, including in reverse directions, to be able to “kickstart” the solitons from chaos. Our early work on micro-combs was based on these 200 GHz FSR combs that operated in this partially coherent state. As mentioned, while not rigorous soliton states, they were still low noise enough and managed to avoid the chaotic regime^[47] — we found them to be more than adequate for our RF work. They were used to successfully demonstrated many different RF signal processing functions^{[85, 107, 109-111]}.

More recently, we have used soliton crystal microcombs that featured not only this new mode of operation, but with a record low FSR spacing of 49 GHz^{[130, 131]}. Soliton crystals arise from mode crossings and are easier and more reliable to generate than other solitons, including DKSs, and even easier than the partially coherent states of the 200 GHz resonators. They can be generated even with simple manual control of the pump wavelength and power, without any complex pump dynamics. The underlying physics for this is based on the fact that the internal optical energy in the cavity of the soliton crystals is very close to that of the chaotic state. Hence, when soliton crystals are generated from chaotic states, there is only a small net change in the intracavity optical energy. This means that there is virtually no induced shift in the resonant wavelength – either thermally or via the Kerr nonlinear effect. It is this shift in resonant wavelength and internal energy that makes DKS states so difficult to generate since the resonance “avoids” the pump wavelength via the self-induced shifting. This same effect is also responsible for greatly increasing the efficiency of the soliton crystals, so that the energy in the comb lines are much higher relative to the pump power than the DKS states, particularly for single soliton DKS states. There is one drawback of soliton crystal states, however, and that is that their spectra are not flat – they have characteristic nonuniform “curtain” patterns. While this can sometimes mean that spectral flattening is required, this is not always the case. Indeed, this issue has not prevented soliton crystals from achieving many breakthroughs in high performance photonic RF microwave functions. Even though the resonator mode-crossing, as well as anomalous dispersion, needs to be engineered, these issues have not posed a significant barrier and high fabrication yields have been achieved^[64]. The specific micro-comb states reached was not important for RF applications, only that low RF noise and high coherence can be achieved, and we found this to be easy to implement with simple pump wavelength tuning. Soliton crystals have yielded the lowest noise levels of any of the micro-comb states that we have worked with and so we have focused on them for our RF work, highlighted by a low phase-noise microwave oscillator^[120].

The ability to detect and analyze RF and microwave signals with very large bandwidths is critical for radar systems, electronic warfare, satellite communications and more^{[79, 80, 141-150]}. RF and microwave channelizers, which effectively slice the RF spectrum into many frequency bands that are compatible with digital processing^[149], are attractive for RF signal detection and analysis. Conventional RF channelizers usually use an array of RF filters and so are susceptible to the electronic bandwidth limitation. Photonic methods are promising for RF channelizers since they can achieve high bandwidths and also have very strong electromagnetic interference immunity.

Early reports of RF photonic channelizers used methods where RF signals were transmitted over single optical wavelengths and were then physically split to achieve spectral channelizing, usually with diffraction gratings^[12], acousto-optic crystals^[151], fiber Bragg gratings^[152], or integrated photonic chips^[153]. However, all of these methods require many narrow linewidth and spectrally dense, precisely centered, filters. Hence these systems tend to be limited in spectral resolution as well as in the number of RF channels. They also tend to have a large footprint.

Recent work on RF spectral slicing using multicasting of RF signals onto many optical wavelengths simultaneously has adopted a number of techniques including stimulated Brillouin scattering^{[92, 154, 155]}, parametric nonlinear optical processes in fiber^{[156, 157]}, spectrally incoherent sliced sources^[158], discrete laser arrays^[159], or electro-optic modulator generated frequency combs^{[149, 160, 161]}. Of these, probably the most representative work^[160] is where two electro-optic frequency combs were used to achieve an RF channelizer. That work used dual EO combs with slightly different FSRs and a carefully controlled offset to effectively achieve a Vernier effect to realize multi-channel detection. In contrast, the majority of approaches to channelization are essentially simply instantaneous frequency measurements designed to detect only a single narrowband signal. In Ref. [160], the RF signal was modulated on the signal comb and could thus be channelized and directly processed via analogue to digital conversion (ADC), without the need for any optical filters. The advantage of that approach, in comparison to our approach based on microcombs, is that it doesn't need a flat-top periodic optical filter. However, a significant disadvantage is that it requires a second optical frequency comb that is carefully controlled relative to the first comb, which is very challenging.

Other approaches that use wavelength scanning devices^[162] or dispersive Fourier transformations^[163] have also been reported, but these generally face limitations of one sort or another, including limited channel numbers and limited RF spectral resolution. They also tend to be fairly complex and expensive because of the requirement for many components such as external RF sources and mode-locked lasers etc. Integrated Kerr micro-combs^{[36-38, 47, 164]}, particularly those fabricated in CMOS compatible platforms^{[37-47, 49-61, 165, 166]}, have many advantages for broadband RF channelizers compared with discrete multi-wavelength sources, such as being able to provide a much higher number of channels or wavelengths^{[107-129, 167]} with a greatly reduced footprint and lower complexity.

Here, we review recent work on broadband photonic RF channelizers realised through the use of integrated optical frequency Kerr microcombs, in combination with passive ring resonator filters. By using an on-chip Kerr comb source consisting of a nonlinear micro-ring resonator (MRR) with an FSR spacing of either 200 or 49 GHz, in combination with a passive on-chip MRR based filter with a FSR of 49 GHz and Q factor of 1.55 × 10⁶, we achieved RF channelizers with high performance. The use of a 49 GHz comb resulted in a larger number of channels since the FSR of the comb source was nearly matched to the 49 GHz passive MRR. For the 200 GHz spaced comb, only every 4th resonance of the 49 GHz passive MRR could be used. We verified the RF performance experimentally at frequencies up to 26 GHz, achieving a high spectral resolution of better than 1.04 GHz for both systems. In addition to high RF performance, this approach offers a reduced footprint, lower complexity, and potentially lower cost.

Fig. 1 shows a schematic of the RF channelizer operating by combining a microcomb as a source (in our case combs with FSRs of either 200 or 49 GHz) with a 49 GHz passive MRR that acts as a narrow optical filter. We first discuss the system based on the 200 GHz FSR microcombs, where we generated Kerr combs in the MRR by pumping with a CW tunable laser which was then amplified by an EDFA. A tunable bandpass optical filter suppressed the amplifier amplified spontaneous emission (ASE) noise and a polarization controller was then used to optimize the polarization coupled into the waveguide. When the wavelength of the pump was tuned near a resonance of the microcomb MRR, if the power was sufficiently large, then parametric gain eventually resulted in optical parametric oscillation. Finally Kerr combs with a spacing equal to the MRR FSR δ_OFC (~ 200 GHz) were generated. In the channelizer wavelength range the frequency of the k-th (k = 2, 3, 4, …) comb line is

 (1)

where f_OFC(1) is the first comb line frequency on the red side. The microcomb lines were made uniform using a WaveShaper and then propagated through a phase modulator (PM) which multicast the RF broadband signal over all wavelengths. Finally, each of the 200 GHz comb lines, where each had the RF signal spectrum imprinted on it, were spectrally sliced, or sampled, by the second high Q MRR with a 49 GHz FSR. Hence, in the first series of experiments we had to use every 4th resonance of the passive MRR to achieve the spectral slicing, which resulted in an effective FSR with a spacing δ_MRR of ~196 GHz. Therefore, the RF spectrum on each of the 200 GHz microcomb lines was progressively sampled with a sequential step between resonances of about 4 GHz. The output channelized RF frequencies are:

 (2)

with f_RF(k) being the k-th RF channelized frequency, and f_MRR(k) the k-th center frequency of the passive MRR filter. Here, [f_MRR(1) – f_OFC(1)] is the spacing between the 1st comb line and nearest filter line, which is equal to the channelized RF frequency offset, and (δ_OFC – δ_MRR) corresponds to the channelized RF frequency step between adjacent wavelengths. While the RF channels could in principle be sampled with an optical spectrum analyzer or de-multiplexer such as an AWG (arrayed waveguide grating), as long as it had a channel spacing that matched the FSR (or a multiple) of the 49 GHz MRR, in practice AWG multiplexers and similar devices generally do not have nearly as fine a resolution as the MRRs used here – they tend to have channel widths matched to the ITU grid of 50 or 100 GHz, versus the MRRs used here that had a 3 dB linewidth of about 1 GHz. The RF signals were then measured using homodyne detection, where the comb lines that had been flattened were separated first before modulating their phase, in order for them to be able to act as local oscillators (LOs). Next, they were coupled together with the channelized optical RF sidebands and then detected coherently before post processing. In the initial work based on the 200 GHz combs^[113], however, we only measured the optical spectrum and not the actual RF waveform, to demonstrate the effectiveness of our method, postponing the channel demultiplexing and RF detection to our subsequent papers. Regardless, even in that initial work the large number of wavelengths generated by the micro-comb resulted in a large channel number, with very broadband RF channelizing compatible with digital channel bandwidths.

That first work was followed by an RF channelizer based on a 49 GHz microcomb combined with the 49 GHz passive MRR filter^[122] that resulted in an RF photonic channelizer with significantly better performance. This was a result of the fact that the two MRRs had approximately equal FSRs, at 49 GHz. The first MRR generated a soliton crystal microcomb that the next MRR acted on as a passive filter. Using the finer spaced (than the original 200 GHz) microcomb had two main benefits. First, the fact that the FSR of the comb source was ¼ of the original work, meant that it generated as many as 92 wavelength channels across the C-band (versus 20 in the initial work). This resulted in a much larger operational bandwidth (instantaneous, without thermal tuning) of 8.08 GHz – > 22 times larger than the 200 GHz comb device^[113]. Secondly, by closely matching the FSR of the second integrated high-Q MRR to the first, in order to sample the RF spectrum using the Vernier effect, a small RF channelization increment of only 87.5 MHz was realized. This effectively resulted in simultaneous and continuous (both temporally and spectrally) RF spectrum channelization, since the channel step was comparable to the passive MRR spectral resolution of 121.4 MHz. Another difference between the second device over the first one is that in the later device we used parallel phase to intensity modulation conversion (PM-IM) for all wavelength channels. This directly resulted in RF signal being generated in a very stable and compact scheme without the use of a separate local oscillator path. Finally, the operation frequency range of the channelizer could be tuned dynamically by varying the offset frequency between the passive MRR and the microcomb. Thermally tuning the passive MRR enabled us to achieve RF channelization for a very large frequency range of 17.55 GHz. In addition to yielding a very high performance in the RF domain, our approach has lower complexity, smaller footprint, and even lower cost.

The second broadband RF channelizer (Fig. 1(b)) is very similar to the previous device and also contains 3 modules. The first consists of comb generation followed by spectral flattening. Here, a MRR was driven by a CW pump to initiate parametric oscillation. The high Q factor (> 10⁶) of the MRR, together with its high nonlinear FOM and anomalous dispersion, together generated sufficient parametric gain to produce Kerr soliton crystal micro-combs. The nature of the oscillation state of the microcomb was determined primarily by the pump to resonance detuning together with the pump power. By sweeping the pump wavelength from blue to red, a range of dynamic nonlinear states could be achieved, including coherent soliton states. The comb lines can also be labelled by Eqs. (1) and (2) except that now the number of lines is much larger. N microcomb lines are generated with a spacing of δ_OFC, the optical frequency of the k-th (k = 1, 2, 3, …, 92) comb line is given by Eq. (1). An optical WaveShaper was again used to flatten the comb line powers.

In the second module, the flattened comb lines were passed through an electrooptical phase modulator, where the input broadband RF signal was multicast onto all wavelengths. Next, and as before, the replicated RF spectra were sliced by the passive MRR with an FSR of δ_MRR, where the slicing resolution is given by the 3 dB bandwidth of the passive resonator, about 1 GHz in our case. Hence, the RF spectral segments for all channels were essentially channelized with a RF centre frequency step given by Eq. (2).

For the 49 GHz comb based device, we used phase modulation combined with notch filtering, using the transmission through port of the second passive MRR, to convert the RF/optical signal from phase modulation to intensity modulation. Phase modulation first produced both lower and upper sidebands that had opposite phases. One of the sidebands was then rejected by the notch resonance, leaving the optical carrier to beat with the other unsuppressed sideband on photodetection. Thus, the process resulted in converting the modulation format from phase modulation to single sideband intensity modulation without a local oscillator. Hence, we effectively realized N bandpass filters in parallel, each with a spectral resolution bandwidth of ∆f, which was set by the resonant linewidth of the passive MRR. The RF centre frequencies f_RF(k) were given by the relative offset between the optical carrier frequency and the passive resonance frequencies [f_MRR(k) – f_OFC(k)] (see Eq. (2)). Therefore, the input RF spectrum was channelized, or demultiplexed, into N channels located at f_RF(k) and each having a ∆f bandwidth.

As mentioned, adopting this method eliminated the requirement for a separate local oscillator path to be able to achieve homodyne coherent detection. Therefore, this approach is more stable and compact than methods that require interference paths^[113]. Finally, the optical signals on the different wavelength channels were de-multiplexed and then converted into the RF domain by a photodetector array, thus producing N parallel channelized RF signals, each having a ∆f spectral width, which in our case was in the operation bandwidth range of typical ADCs. Following this, the RF channelized signals were converted into digital signals using an ADC array and finally analyzed further with digital domain tools.

Both MRRs (Fig. 2) were based on a Hydex glass, a high-index doped silica glass platform that has CMOS compatible fabrication processes. First, high-index (n = ~1.70) Hydex glass films were deposited using PECVD, then patterned using photo-lithography and then reactive ion etched to produce waveguides that featured very low surface roughness. Finally, low index Hydex (n = ~1.44 at 1550 nm) was grown, again by PECVD, to produce the upper cladding layer. The advantages of Hydex for nonlinear optical parametric oscillators (OPOs) are very low linear propagation loss (~0.06 dB·cm⁻¹), a reasonably high optical nonlinearity (~233 W⁻¹·km⁻¹), and most importantly, essentially zero two-photon absorption even at extremely high intensities of ~25 GW·cm^−2[38]. The low linear loss yielded Q factors for the ring resonator that exceeded one million. A scanning electron microscope (SEM) image of the 200 GHz MRR cross section is shown in Fig. 2(c). The radii of the MRR used for the microcomb generation was 200 GHz and the passive MRR used for spectral slicing were ~135 and ~ 592 μm, corresponding to FSRs of ~1.6 nm (~200 GHz) and ~0.4 nm (~49 GHz), respectively. The device was packaged with fibre pigtails with a fibre–fibre insertion loss of ~1.5 dB for the 49 GHz MRR and ~3.5 dB for the 200 GHz MRR.

The pump laser power was boosted to 500 mW and then the wavelength tuned from blue to red in order to generate the microcombs. Initially, with the pump wavelength tuned near the MRR resonance, the primary microcombs were generated first (Fig. 3(a)), which featured idler and signal frequencies in the L and S-bands with a pump – signal spacing of 19 FSRs (3.8 THz) governed by the modulational instability gain profile. When the pump to adjacent resonance wavelength offset was tuned further, the parametric gain peaks widened, followed by the appearance of secondary comb lines with a spacing equal to the FSR, driven by cascaded FWM. The final soliton crystal Kerr combs were produced with the pump wavelength at 1548.58 nm. Figs. 3(b) and 3(c) show that the ensuing microcomb was relatively flat over the full C and L bands and was more than 200-nm wide over all S, C, L, and U bands, featuring 20 wavelengths across the C-band and more than 60 wavelengths across the combined C and L-bands. The large comb spacing translated into a large Nyquist zone, with an RF bandwidth over 100 GHz. This is very difficult to achieve with either externally-modulated comb sources or mode-locked lasers.

While the 200 GHz comb did not have a spectral profile that indicated operation in the single soliton state, this did not pose a limitation. Our theoretical analysis indicated that the Kerr comb was working in a partially coherent state that featured a stable amplitude and phase. Indeed, the spatial patterns appeared to reflect operation in soliton molecule states^{[130, 131, 168-170]}, which are low noise states, similar to cavity solitons.

Since the comb source served both as an optical carrier for RF sideband generation and as a LO for RF receiving (for the 200 GHz comb), the channelized RF optical sidebands and LOs were inherently coherent, and so rigorous phase-locked comb lines were not needed. Indeed, this is an advantage compared with RF coherent receivers^[171]. Similarly, any coherence between adjacent wavelengths is also not needed since the wavelengths do not beat together.

The 20 C-band channels of the 200 GHz comb were shaped by a Finisar 4000S WaveShaper to equalize the power of the wavelength channels. A feedback control path accurately read and shaped the power of the comb lines as monitored by an optical spectrum analyser, which was then compared to the theoretical weights which were uniform in our case. This generated a feedback signal error to the WaveShaper so as to calibrate the system and perform accurate comb flatness and shaping to within ± 0.3 dB (Fig. 3(d)). Note that passive gain flattening filters can be used to flatten the spectrum instead of the WaveShaper. A phase modulator then multicast the RF signal onto all wavelengths, where Figs. 3(e) and 3(f) show the optical spectra of the flattened comb lines that were modulated with the RF signal. The 20 equalized channels were limited to the C-band only because of the operating wavelength range of the WS. In practice the number of wavelengths can be increased easily to 60 with the use of a C/L band WS. Next, each wavelength channel was spectrally sliced in order to extract the RF signal, by using the second MRR (49 GHz). Figs. 4(a) and 4(b) show the 49 GHz MRR drop-port transmission over a range of 5 nm, with an FSR of ~49 GHz (~0.4 nm) over 152 modes (denoting the mode at 193.294 THz as 0). The single resonance transmission (Fig. 4(c)) has a 124.94 MHz FWHM, or Q = 1.549 × 10⁶, with a 1.04 GHz –20 dB bandwidth. This equals the RF resolution, which is compatible with state-of-the-art A/D^[141] modules, and so our broadband RF channelizer is digital-compatible.

To demonstrate the operation principle of the RF channelizer, we show both the 200 GHz comb and the 49 GHz passive MRR transmission spectra, marking every 4th resonance that lines up with the 200 GHz comb in Fig. 5(a). We also see that for the 20 C-band comb channels, the relative offset between every 4th resonance of the 49 GHz MRR and the 200 GHz comb lines shifts linearly in the range 3.89–88.65 GHz, and this determines the RF spectral sliced frequency. This 200 GHz comb lines were measured to be at an FSR of δ_OFC = 200.44 GHz while four times the 49 GHz MRR lines were δ_MRR = 196.01 GHz. Thus the channelized RF frequency increment between channels was δ_OFC – δ_MRR = 4.43 GHz.

We measured the frequency response of the RF channelizer up to 20 GHz. This required only 4 of 20 channels in Fig. 5, equal to 4 frequencies in the range 1.7 to 19.0 GHz. We measured the transmission of the passive 49 GHz MRR with an optical spectrum analyzer (Fig. 6), illustrating how successive RF optical sidebands featuring progressively shifting RF frequencies were channelized at different wavelengths. Optical-carrier RF tones (Fig. 6(a)) at frequencies of 1.7, 6.3, 11.2, and 15.8 GHz were contained in 4 wavelengths, yielding an RF frequency step of ~ 4.8 GHz, consistent with Fig. 5. The leaked power of channel 4 arose primarily from extraneous optical carrier tones, but these can be easily decreased by optimizing the modulation format, for example by using carrier-suppressed single sideband modulation.

In order to fully exploit the 1 GHz RF resolution of our system and fill in the gaps created by the ~ 4.8 GHz RF step size, we temperature tuned^{[47, 171]} the passive MRR to continuously vary the relative spacing between the comb and the MRR filter wavelengths (f_MRR(1) – f_OFC(1)). This thermal tuning could be accomplished with a millisecond response time. Figs. 6(a)–6(d) show the output optical spectra at temperatures ranging from 24.0 to 25.5 °C. As seen in Fig. 6(e), the RF channelized frequency offset varied by ~1 GHz/°C. By combining the finer resolution of the temperature tuning with the 4.8 GHz RF step size over 60 wavelengths in the C/L-bands, this device achieved a very broad RF bandwidth > 100 GHz.

Fig. 7 shows the channelized radio frequencies of the various wavelengths extracted from Fig. 6, across the RF range 1.7–19.0 GHz, restricted only by the bandwidth of the RF modulator. Fig. 8 shows the extinction ratio of the channels (ER - the ratio of the signal channel optical power to the maximum power of the remaining channels) deduced from Figs. 6(a)–6(d), showing an ER as high as 20 dB. The optical power non-uniformity of the channels arose predominantly from the spectral gain profile of the amplifier. This can be reduced greatly with the use of gain flattening filters. The ER was also limited by excess noise of the 2nd EDFA when adapting the optimized RF modulation format in order to increase the optical signal-to-noise-ratio (OSNR). To address this lower noise EDFAs could be used to enhance the performance. The thermal controllers had an accuracy of 0.01 °C, contributing to an error of about 50 MHz for the channelized RF frequencies, which could be reduced with more precise thermal control or through use of a feedback loop.

Although thermal tuning does not result in the channelization of the full RF spectrum simultaneously, it nevertheless allowed us to show the potential of our work. In fact, thermal tuning is not necessary, as long as the frequency increment between nearest wavelength channels (resulting from the FSR difference between the microcomb and passive MRRs) is comparable to the spectral resolution (the –20 dB linewidth of the passive MRR), which can be achieved through accurate lithographic control of the MRR FSRs. One should, for example, design δ_OFC = 200 GHz for the passive MRR so that δ_MRR = 199 GHz, still with a 1 GHz 20 dB bandwidth. Hence, the RF spectrum could be fully and simultaneously channelized and detected with a 1 GHz resolution without the use of thermal tuning. While the product of channel number and resolution – the RF bandwidth – would be a bit less than the device demonstrated here, at 60 GHz with 60 wavelengths over the C and L bands, this could be increased by using a lower FSR comb (100 GHz versus 200 GHz for example). Finally, our approach generates RF outputs by using homodyne detection. In contrast, down-conversion of the RF channelized signal would require the use of heterodyne detection with a 2nd optical comb having a specifically designed FSR and offset, which has been demonstrated by recent breakthroughs in dual-comb spectrometers^{[160, 172-174]}. By using dual integrated microcombs with carefully designed offsets and FSRs for heterodyne detection^[175-177], broadband RF signals can be channelized and then down-converted to digital bandwidths for direct A/D conversion and post processing. This approach offers a very attractive solution for photonic integrated RF receivers.

In this section, we summarize the results for the photonic RF channelizer based on two MRRs having approximately matching FSRs at 49 GHz. This approach achieves significantly improved performance over the previous device based on a 200 GHz spaced microcomb. Here, the soliton crystal micro-comb is generated by the first MRR with a spacing equal to the FSR of 49 GHz, whereas the second MRR serves as a filter with approximately the same channel spacing. This has two main benefits. First, the smaller FSR of the soliton crystal comb source provides up to 92 wavelengths across the C-band. This results in a greatly increased instant RF bandwidth of 8.08 GHz – more than 22× that of the previous device^[113]. By using high-Q integrated MRRs with approximately matching spacings to both generate the comb and filter the RF spectrum via the Vernier effect, we achieved a RF channelization step between channels of 87.5 MHz. This led to a channelization of the continuous RF spectrum since the channelization increment was less than the 121.4 MHz RF spectrum resolution. We also employed parallel phase to intensity-modulation conversion (PM-IM) over all wavelengths — which directly generated an RF output in a stable scheme and compact footprint without needing any independent local oscillator paths. Lastly, the total RF operation frequency range of the channelizer could be tuned dynamically by varying the frequency offset between the passive MRR and active microcomb. The passive MRR was thermally tuned to achieve RF channelization across a 17.55 GHz bandwidth. Apart from achieving high performance, our design yields a lower complexity, a smaller footprint, and potentially even lower cost.

Fig. 1(b) shows the setup of the broadband RF channelizer. As before, the microcomb was generated by driving the first MRR with a CW pump to induce parametric oscillation. The high 1 million Q factor of the MRR together with the high nonlinear FOM, and engineered waveguide anomalous dispersion, all contributed to achieving the parametric gain needed to produce the soliton crystal Kerr microcombs. The particular operation mode of the frequency comb was governed by the pump to resonance offset as well as the pump power. By tuning the optical pump laser wavelength from the blue to the red, a range of dynamic nonlinear states, including coherent solitons, could be achieved. As before, for a microcomb with N wavelengths at a spacing of δ_OFC, the optical frequency of the k-th (k = 1, 2, 3, …, 92) comb line is given by Eq. (1). As before, a WaveShaper optical spectral shaper flattened the channel powers. For the second module, the equalized comb lines were modulated via an electro-optic phase modulator to multicast the RF signal across all wavelengths. Next, the replicated RF signals imprinted on all comb lines were sampled by the passive MRR with an FSR of δ_MRR, and with a resolution determined by the 3 dB resonator linewidth. As before, this resulted in a staggered RF frequency step for the RF spectral segments across the wavelength channels described by Eq. (2).

We employed notch filtering and phase modulation in this device (using the transmitted signal from the through port of the passive MRR) to perform conversion from phase to intensity modulation. First, the phase modulation produced opposite phases for the lower and upper sidebands. Next, one of the sidebands was suppressed by the filter resonance notch, leaving the other unsuppressed sideband to beat with the optical carrier upon photodetection. This essentially converted the modulation format from phase to single-sideband intensity modulation. This process therefore resulted in N bandpass filters, each with a spectral resolution of ∆f determined by the passive MRR linewidth. The microwave center frequencies f_RF(k) were given by the offset between the passive resonances and optical carriers [f_MRR(k) – f_OFC(k)] (Eq. (2)). Consequently, the RF input signal is demultiplexed into N segments at f_RF(k), each with a bandwidth ∆f. This method doesn’t need separate physical paths for the local oscillator in order to achieve homodyne coherent detection, and therefore it is much more compact and stable than the first approach^[113]. Lastly, the optical wavelengths were de-multiplexed and then changed back into the RF domain individually via an array of photodetectors, producing N channelized RF signals simultaneously, each with a ∆f spectral width within the bandwidth of typical ADCs. This was then converted into digital signals with an ADCs array and processed with digital tools.

Both the passive and microcomb MRRs were fabricated in Hydex. The dispersion of the microcomb 49 GHz FSR MRR was tailored to be anomalous in the C-band to enable parametric oscillation. Further, a mode-crossing at ~1556 nm was engineered which could initiate the background wave required for soliton crystal generation. During comb generation, the pump power was boosted to ~ 2 W while the wavelength was swept manually from blue to red. As the detuning between the pump wavelength and the microcomb MRR’s resonance became small enough to ensure sufficient modulation-instability gain in the microcomb MRR, primary combs were initiated. As the detuning was changed further this was followed by soliton crystal microcombs. We observed distinctive ‘curtain’ soliton crystal optical spectra (Figs. 9(a) and 9(b)), a resulting from the interference of the circulating tightly-packed solitons around the ring^{[130, 131]}. The microcombs had a spacing equal to the microcomb MRR FSR of 48.9 GHz, producing up to 92 wavelength channels in the C-band. Soliton crystals are low-noise and coherent and, due to their high intracavity power, they can be easily generated just via manual tuning of the pump wavelength — much easier than DKS solitons.

Next, the power of the soliton crystal micro-comb wavelengths were equalized with a WaveShaper. The comb lines’ powers were accurately shaped by using a feedback control path monitored with an optical spectral analyzer that compared the theoretical channel powers, generating an error signal that was then used to set the WaveShaper loss, with the flattened comb spectrum shown in Fig. 9(c). The 92 flattened microcomb lines were then passed through a phase modulator in order to produce optical carriers, thus multicasting the input RF signal onto all wavelengths. The passive MRR then spectrally sliced the RF replicas. We used phase modulation combined with notch filters (the passive MRR through-port transmission) to translate the passive MRR high-Q resonances into the RF domain, after which the input RF signal spectrum was channelized into N = 92 channels, each centered at f_RF(k), and each having a bandwidth of ∆f.

To quantify these experimental parameters, we measured the passive MRR transmission spectrum (Fig. 10(a)) with an incoherent broadband optical source, consisting of amplified spontaneous emission from an EDFA. Both TM and TE polarized resonances were observed, where it is seen that the channelized RF frequency f_RF(k), calculated from the spacing between the comb line and adjacent passive resonance, successively decreases from the blue to red. The resulting RF channelized frequency (Fig. 10(b)) shows a total operation bandwidth of 8.4 and 3.8 GHz for TE and TM resonances, respectively, with a RF channelization step of 41.4 MHz (TM) and 89.5 MHz (TE) per channel. Since the channelization resolution – the linewidth of the passive resonances – was ~120 MHz, we used the TE passive resonances to obtain a larger bandwidth with lower adjacent-channel interference. By aligning the passive MRR’s input light polarization to the TE mode, the TM passive resonances weren’t excited and so they did not degrade the performance of the device. The dual polarization modes and birefringence of the MRRs are discussed in another work^{[114, 115]}.

The RF spectrum was channelized into segments by the passive MRR, with each RF carried on a different wavelength, which were then de-multiplexed into spatial parallel paths and separately detected. We used a Vector RF Network Analyzer to characterize our channelizer performance. The experimental RF transmission spectra for 92 channels (Figs. 11(a) and 11(b)) verify the success of our method. Any imbalanced powers of the RF channels can be equalized by factoring the RF transmission spectra into the feedback comb shaping, with the error signal being generated by the difference in the ideal channel weights with the RF, rather than the optical, power. The SNR of the channelizer was 23.7 dB which was extracted from the RF transmission spectra. This can be improved by increasing the extinction of the periodic optical notch filter. For the MRRs, this can be realized through adjusting the coupling coefficient from the waveguide bus to the MRR.

The RF channel centre frequencies are shown in Fig. 11(c), exhibiting a channel increment of 87.5 MHz/channel, with a total instant RF bandwidth of 8.08 GHz. This agrees with theory based on the transmission spectrum of the MRRs. The average resolution of the RF channels, derived from the RF transmission spectrum 3dB bandwidth for each channel, was 121.4 MHz. This resulted from the high Q factors of the passive MRR. This high resolution greatly relaxed the demands on the ADCs bandwidth and verifies that our approach is compatible with the relatively small bandwidths of digital components.

The device based on the 49GHz soliton crystal microcomb had moderately higher adjacent-channel crosstalk. This arose mainly because of the slight difference between the RF resolution (121.4 MHz) and the channelization step (87.5 MHz). Ideally these should be equal. Nonetheless, while this situation of a channel step being less than the resolution is not perfect, any residual channel crosstalk could be eliminated by discarding redundant channels. Fine tuning the FSR of the passive MRR during fabrication would increase the RF channelization step so that it could be designed to be equal to the RF resolution. This would simultaneously result in a larger instantaneous bandwidth of 121.4 MHz × 92 = 11.17 GHz. Moreover, a much flatter passband and higher roll-off than the natural Lorentzian profile of a single MRR can be realized by using high-order optical filters^[178]. Doing this would also reduce the adjacent-channel crosstalk. Finally, the filtering resolution can be increased by using higher Q factor MRRs. Finally, for our waveguide platform the dispersion was low enough that it did not affect the FSR across the wavelength range of interest, which was constant as seen in Fig. 4(a). Further, for the channelizers in general, it is only the relative FSR shift that matters and so even for larger values of dispersion, as long as the microcomb source and passive filter have similar dispersions, the RF step size will not be affected. This is verified in our case by the fact that the relative frequency shifts across the C-band that we observed were extremely linear with wavelength (Figs. 5 and 11), indicating that dispersive effects here were negligible. Of course, in cases where a completely different platform is used for the passive filter compared to the microcomb ring resonator (such as silicon-on-insulator (SOI)), then any dispersion that is large enough to affectthe FSR spacing across the wavelength region would need to be accounted for and if necessary compensated for.

Lastly, by thermally tuning the passive MRR, the soliton crystal channelizer’s operation bandwidth can be made tunable over a wider range. The offset between the comb and filter MRRs’ resonances (f_MRR(1) – f_OFC(1)) can be thermally controlled on a timescale of milliseconds. As shown in Fig. 11(d), we shifted the instantaneous operating band of the channelizer from 1.006–9.147 to 9.227–17.49 GHz, continuously covering a total RF bandwidth of 16.48 GHz, by tuning the chip temperature of the passive MRR. Although thermal tuning does modify the passive MRR FSR slightly, this effect was very small. Indeed, it was a factor of ~3955 × (193.4 THz/48.9 GHz) smaller than the relative change in the offset (f_MRR(1) – f_OFC(1)). The change in the comb line to MRR offset (ie., f_MRR(1) – f_OFC(1)) resulting from thermal tuning was 8.221 GHz, while the corresponding change in the passive MRR FSR (RF channelization step) was only 2.1 MHz, and so this did not affect the device performance.

Finally, even though the RF instantaneous operation bandwidth could be moved across a wide range with thermal tuning, the RF operational lower frequency was limited to ~ 600 MHz, arising from several effects. Thermal locking of the center filter resonance to the optical carrier wavelength^[39] results in a lower limit for the RF response. In addition, the intrinsically low spectral roll-off characteristic of Lorentzian lineshapes, that the passive MRRs featured here, leads to a degradation in the phase-to-intensity modulation. This can be improved by using high-order filters having steep roll-offs^[178], while at the same time the thermal issues can be reduced by improving the level of control. The highest RF frequency of our channelizers was governed by the Nyquist frequency of our microcomb, which was 48.9 GHz/2 = 24.45 GHz for the soliton crystal devices and 100 GHz for the 200 GHz microcomb device. Hence, the 200 GHz system had a much higher potential maximum RF operation frequency of 100 GHz — although this involves a compromise with the number of comb lines in the useable wavelength range.

We demonstrate broadband RF channelizers based on CMOS-compatible integrated optical frequency comb sources. Broadband 200 GHz-spaced Kerr combs as well as soliton crystal combs with a 49 GHz spacing are employed. Both combs provided a record large number of wavelength channels as well as a large RF operation bandwidth. The 200 GHz comb-based device has a large potential upper bandwidth of 100 GHz, although with a limited number of channels in the C-band, typically < 20. The soliton crystal device, on the other hand, because of its lower FSR of 49 GHz that closely matched the FSRs of the MRR passive filter, yielded up to 92 wavelengths over the C-band, resulting in high performance but a correspondingly lower instantaneous RF bandwidth of 8.08 GHz. A high 121.4 MHz RF spectral resolution was realized by using a high- Q passive MRR for the optical periodic filter for both devices. For the soliton crystal device, we used phase to intensity modulation format conversion to achieve stable signal detection without any requirement for separate local oscillator paths. Dynamic adjustment of the operation RF frequency range was accomplished for both devices with the use of thermal control of the passive MRRs, achieving RF operation of up to 17.49 GHz for the soliton crystal device. This device also resulted in a broadband channelization of RF frequencies in the range 1.7–19 GHz with a high 1 GHz spectral resolution, achieved by thermal tuning of the on-chip passive MRR with a Q factor of 1.549 × 10⁶. These micro-comb based RF channelizers are highly attractive for broadband RF channelization with large channel numbers, high resolution, compact footprint, and potentially low cost. They feature many parallel channels and are very attractive solutions for broadband instantaneous signal detection and processing. They represent a key step towards fully integrated photonic receivers for modern RF systems.