Microwave filters are indispensable building blocks in radio frequency (RF) receivers and are paramount to myriad applications, including radar, wireless communication, and sensing [1–3]. Due to the electronic bottleneck, traditional electronic microwave filters are facing significant challenges for meeting the exponentially increasing capacity in next-generation RF systems [4,5]. For instance, 6G wireless networks in the near future are expected to support an ultrahigh data transmission rate of 100 Gb/s in which the higher RF spectrum region such as mm wave (30 to 300 GHz) shall be harnessed [6,7], while the valid operation frequency of existing electronic filters is commonly restricted below 10 GHz [8]. As a promising alternative, microwave photonic filters (MPFs), which perform filtering operations in the optical domain, are capable of offering wideband tunability along with high stability, low latency, and strong immunity to electromagnetic interference [9]. Such outstanding advantages have propelled abundant explorations on MPFs over the past decades [10–14]. Nevertheless, conventional MPFs are implemented based on bulk devices or benchtop instruments and hence are plagued by the limitations in system cost, space, and power consumption.

Recently, the integration of MPF systems onto a chip-scale platform has been driven by advances in nanophotonic fabrication techniques, which is termed “integrated MPFs” [5]. Considerable progress has been made on integrated MPFs that explore different schemes based on microring resonators (MRRs) [15,16], Mach–Zehnder interferometers (MZIs) [17], waveguide Bragg gratings [18], microcombs [19,20], on-chip stimulated Brillouin scattering [21,22], etc. These demonstrations prove the inherent superiorities of integrated MPFs, in terms of compact footprint, light weight, and scalability. However, in most reported integrated MPFs, only the passive components are incorporated on chip, while all necessary active devices, including lasers, modulators, and photodetectors (PDs), are off-chip and based on fiber connections. The overall systems of such integrated MPFs are still complex, with high power dissipation and sensitivity to environmental fluctuations, thus imposing restrictions on their deployments in real RF systems. Therefore, a higher degree of photonic integration is becoming an urgent pursuit for the implementation of integrated MPFs.

To date, few attempts to create an MPF in high photonic integration have been executed based on monolithic platforms, including indium phosphorus (InP) and silicon platforms. The InP platform can enable monolithic integration of all necessary active and passive devices, on which the first fully integrated MPF was realized [23]. However, issues related to large propagation loss severely limit the accessible spectral resolution on this platform (e.g., 3 dB bandwidth is 2.5 to 5.5 GHz in Ref. [23]). Compared with the InP platform, the silicon photonics platform offers moderately low loss, together with small size, high-volume production, and compatibility with electronic circuits [24]. More recently, MPFs in a high integration level based on a silicon platform have also been demonstrated [25,26]. For example, a silicon highly integrated bandpass MPF consisting of a phase modulator, a micro-disk resonator, and a PD was reported; this MPF features a passband bandwidth of 1.93 GHz, a frequency tunable range within 3 to 10 GHz, and an extinction ratio of 15 dB [25]. Nevertheless, a chip-scale laser source has always been absent for silicon-integrated MPFs due to an indirect bandgap.

Despite these achievements, several pivotal problems still persist, hindering the follow-up development of integrated MPFs. First, the realized filtering performances as indicated above are relatively limited and not competitive with those of discrete MPFs. Generally, toward next-generation RF applications, the integrated MPF is supposed to meet rigorous filtering characteristics, including a sub-GHz bandwidth [4], large rejection ratio ($>40\mathrm{dB}$ [5]), and high frequency band ($\ge 18\mathrm{GHz}$ [27]). Second, the former MPFs in high photonic integration have only realized a single filtering function. While in the modern frequency-agile RF environment, the reconfigurable microwave filters, especially those with bandpass/band-stop switchable response have gained growing necessity due to potential benefits for improving versatility and reducing volume [28–30], which have not been demonstrated in all-integrated format. Moreover, all integrated MPFs reported until now are implemented based on individual photonic platforms (most are InP, Si or ${\mathrm{Si}}_3{\mathrm{N}}_4$), since the integration materials have their own strengths and weaknesses, providing all essential components with required characteristics of a complete MPF are fairly challenging on a single platform [31]. Typically, Si and ${\mathrm{Si}}_3{\mathrm{N}}_4$ platforms are ideal candidates to realize high spectral resolution due to low loss; however, laser integration on these substrates has remained a longstanding difficulty. Fortunately, however, hybrid integration has recently been recognized as a feasible avenue to construct efficient bridges among different material platforms [32], which mainly relies on optical connections in the form of direct facet-to-facet coupling [33], micro-optics [34], or photonic wire bonding [35]. Hybrid integration of multimaterial platforms carried out in transceivers [36,37], quantum photonics [38], frequency synthesizers [39], and so forth, however, remains unexplored for creating a completely integrated MPF.

In this work, we address all the aforementioned points and demonstrate an all-integrated high-performance MPF. This integrated MPF is realized through hybrid integration of an InP chip-scale laser and a monolithic silicon chip consisting of a dual-drive Mach–Zehnder modulator (DDMZM), a high-$Q$ MRR, and a PD. The filtering response of this integrated MPF can be flexibly switched between band-stop and bandpass functions by controlling bias DC voltage of the DDMZM. For the band-stop filtering, a wide-frequency tunable range of 3 to 25 GHz and a narrow 3 dB bandwidth of 380 to 450 MHz were achieved, along with a high rejection ratio of $>40\mathrm{dB}$. For the bandpass filtering, the frequency tunable range and 3 dB bandwidth were measured as 3 to 21 GHz and 360 to 470 MHz, respectively. In regard to the RF link performances, an RF gain of $-28.2\mathrm{dB}$, a noise figure of 51.2 dB, and a spurious-free dynamic range (SFDR) of $99.7\mathrm{dB}·{\mathrm{Hz}}^{2/3}$ were obtained. Compared with former MPFs in a similar photonic integration degree, the spectral resolution is greatly promoted by nearly one order of magnitude, while the tunable range is extended more than twice. Notably, thanks to such unparalleled filtering performances, high-resolution RF filtering of actual microwave signals is given, which, to our best knowledge, has never been performed by a highly integrated MPF before.

The architecture of the hybrid integrated MPF is illustrated in Fig. 1(a), incorporating an InP waveguide-integrated distributed feedback (DFB) laser and a monolithic silicon chip onto the submount. The silicon chip consists of three key devices: a DDMZM, a high-$Q$ racetrack MRR, and a PD. Figure 1(b) presents the operation principles for the bandpass/band-stop switchable filtering of the integrated MPF. A continuous-wave (CW) light with a center angular frequency of ${\omega }_c$ is generated by the InP laser and then is routed to the DDMZM. The input RF signal with an angular frequency of ${\omega }_{\mathrm{r}\mathrm{f}}$ is applied to the DDMZM with the aid of a 90° RF hybrid. By adjusting the phase difference ${\phi }_{\mathrm{DC}}$ between two MZI arms via a thermal phase shifter, both the equivalent phase modulation (EPM, ${\phi }_{\mathrm{DC}}\text{=}0$) and the asymmetric double-sideband modulation (AS-DSB, $0<{\phi }_{\mathrm{DC}}<\pi /2$ or $\pi /2<{\phi }_{\mathrm{DC}}<\pi$) schemes can be acquired. This provides the ability to switch the filtering response between bandpass (based on EPM) and band-stop (based on AS-DSB). Subsequently, the modulated optical spectra are tailored by the resonance notch of the high-$Q$ MRR. Finally, the processed optical spectra are translated to corresponding RF filtering response by means of the PD. A detailed analysis can be found in Appendix A.

The monolithic silicon photonics chip is designed in-house and fabricated on a 220 nm SOI platform leveraging CMOS-compatible processes by Advanced Micro Foundry, as shown in Fig. 1(c). The footprint of the whole chip is about $1.34\mathrm{mm}\times 3.18\mathrm{mm}$. The DDMZM is based on a carrier-depletion scheme with a pair of 2.5 mm long electro-optical (EO) phase shifters. Traveling-wave electrodes are utilized to support high-speed operation. Two TiN microheaters with a length of 130 μm are placed on the two arms of the MZI to control the bias point of the DDMZM. The high-$Q$ MRR is based on rib waveguides with a 90 nm slab and is designed to have an all-pass racetrack configuration with a free spectral range (FSR) of $\sim 50\mathrm{GHz}$. To improve the $Q$-factor, the width of straight waveguides is set as 2 μm for eliminating scattering loss from the rough sidewalls, while the width of bent waveguides (20 μm in radius) is 450 nm to enable low-loss bending. To avoid exciting high-order optical mode, 40 μm long adiabatic tapers are used for the connection of 2 μm waveguides and 450 nm waveguides. A TiN microheater is applied to tune the wavelength shift of the resonance peaks. The PD is designed as a vertical p-doped-intrinsic-n-doped (PIN) structure and realized by germanium (Ge) epitaxy growth technique. Besides, in order to monitor the optical signals at different positions of the on-chip optical link, several asymmetric optical couplers with a splitting ratio of 1:9 [40] are inserted to extract 10% power of the propagation light for monitoring.

The integrated light source used here is a commercially available InP DFB laser diode die ($0.9\mathrm{mm}\times 0.25\mathrm{mm}$) based on a buried-heterostructure waveguide structure [see Fig. 1(d)], which has an emitting wavelength around 1552 nm and an output power of 95 mW. A thermoelectric cooler (TEC) is implemented to control and stabilize the operation temperature. The InP laser chip and the monolithic silicon chip were hybrid integrated with the help of micro-optics elements (see Appendix B) and assembled into a compact package, as shown in Fig. 1(e). The RF and DC pads on chip were wire-bonded to a custom-designed print circuit board (PCB) for electrical connections.

The transmission spectrum of the high-$Q$ MRR is presented in Fig. 2(a). The extracted full width at half maximum (FWHM) is 2.3 pm, corresponding to a loaded $Q$-factor of $6.7\times {10}^5$. The Lorentzian fitting shows that the intrinsic $Q$-factor and the average propagation loss are $1.63\times {10}^6$ and 0.41 dB/cm, respectively. Such a high $Q$-factor will enable the integrated MPF to realize highly selective filtering with a sub-GHz passband or stopband bandwidth. Figure 2(b) presents the phase response of the high-$Q$ MRR within a resonance, which was measured by a high-resolution optical vector analysis method (Appendix C). It can be observed that the phase shift at the resonance frequency is nearly 180°, indicating the high-$Q$ MRR is under the overcoupling state [41]. Thus, the bias phase ${\phi }_{\mathrm{DC}}$ is required to be tuned to the range of $[\pi /2,\pi ]$ for realizing a deep band-stop filter (Appendix A). The tunability of the optical spectrum was also evaluated, as shown in Fig. 2(c). By changing the DC voltages applied to the TiN microheater from 0 to 3 V, the resonance wavelength of the high-$Q$ MRR can be continuously redshifted up to 0.24 nm (30 GHz).

Then, the performances of the active devices in this hybrid integrated MPF were evaluated. The output of the InP-based DFB laser was sent into an optical spectrum analyzer (OSA, Yokogawa 6370C) to characterize the laser spectrum, as seen in Fig. 3(a). The side mode suppression ratio is higher than 55 dB around 1552 nm. The linewidth of the laser was measured by a delayed self-heterodyne (DSH) method (Appendix D), which is extracted as narrow as 150 kHz, as shown in Fig. 3(b). For the DDMZM, the static insertion loss is $\sim 4.6\mathrm{dB}$, and the modulation efficiency is approximately $1.65\mathrm{V}·\mathrm{cm}$. The 3 dB EO modulation bandwidth was measured as 26.5 GHz at a reverse-bias voltage of 2.5 V. For the PD, a detection responsivity of $\sim 0.8\mathrm{A/W}$ at 1550 nm was observed. The 3 dB optical-electro (OE) bandwidth was tested as large as 38 GHz under a bias voltage of 2 V. More detailed measurement results can be found in Appendix E. Thanks to the large bandwidths of both the DDMZM and PD, the integrated MPF is capable of supporting a rather wide operation frequency range.

Using the setup shown in Fig. 4, two experiments for measuring the RF filtering response of the integrated MPF were performed. In the first experiment, the RF spectra were measured based on the monolithic silicon chip and an external light source (EXFO, T100S-HP). In this case, a CW light was generated by the external laser and injected into the silicon chip with a power of $\sim 16\mathrm{dBm}$, subtracting a coupling loss of 7 dB. Frequency-swept RF signals with 9.5 dBm power were provided by a vector network analyzer (VNA, Keysight N5247A) and divided by a 90° RF hybrid (Marki, QH-0440) to drive the two EO phase shifters of the DDMZM. Then, the recovered RF signals from the PD were sent back to the VNA to acquire the frequency response of the integrated MPF. During the measurement, the input/output RF signals were connected to the bond pads of the silicon chip through high-speed microwave probes (Cascade Microtech, ACP40). The desired reverse bias voltages of the DDMZM and PD were provided by a multichannel DC power supply (Keysight, E36312A) and combined with RF signals using bias-tees (SHF, BT65D). In addition, for thermally tuning the DC bias phase ${\phi }_{\mathrm{DC}}$ of the DDMZM and the resonance wavelength of the high-$Q$ MRR, another DC power supply (Keysight, E36312A) was adopted.

The frequency response of the integrated MPF operating as a band-stop filter was measured first. A resonance notch at 1552.27 nm was selected to filter the upper sideband of the AS-DSB optical signals. Through precisely controlling the DC bias phase of the DDMZM to enable full cancellation of two beat sidebands, a deep RF frequency notch was realized at $\sim 3\mathrm{GHz}$. In order to validate the frequency tunability, the resonance wavelength of high-$Q$ MRR was tuned through a microheater. As shown in Fig. 5(a), with the electrical power applied to the microheater sweeping within 0 to 17.4 mW, the center frequency of the notch band can be continuously tuned from 3 to 25 GHz (S-band to K-band). It should be noted that the lower-frequency limit is restricted by the valid operation range of the RF hybrid, while the upper frequency limit is determined by the half-FSR of the high-$Q$ MRR. Therefore, the frequency tunable range could be flexibly extended via further optimization. Figure 5(b) depicts the 3 dB bandwidths and rejection ratios as functions of RF frequencies. It can be observed that the rejection ratios maintain a high level of $>40\mathrm{dB}$ over the entire frequency tunable range (3 to 25 GHz), and the 3 dB bandwidths are extracted within 380 to 450 MHz. The long-term stability for maintaining these high performances was also evaluated and is given in Appendix F.

Next, the integrated MPF operating as a bandpass filter was demonstrated and evaluated. By setting ${\phi }_{\mathrm{DC}}$ equal to zero, the filter function was rapidly switched to bandpass filtering with a response time of $\sim 48\mathrm{\mu s}$ (Appendix G). Figure 6(a) illustrates the frequency responses of the bandpass filtering. As can be seen, the center frequency of the passband can be adjusted in a wide frequency range of 3 to 21 GHz with the electrical power changing from 0 to 15.1 mW. Figure 6(b) represents the 3 dB bandwidths and rejection ratios versus different RF frequencies. The 3 dB bandwidths are located at 360 to 470 MHz and are in good conformity with those of the band-stop filtering. In parallel, the rejection ratios are $>10\mathrm{dB}$ over the entire frequency tunable range (3 to 21 GHz), which are poorer than those in Fig. 5(b). The reduced rejection ratios arise from the insufficient cancellation between the $\pm 1\mathrm{st}$ order sidebands, given that the ideal EPM relationship is difficult to be maintained over a broad frequency range, considering the phase unbalances originated from the RF hybrid, cables, and fabrication errors of devices.

In the second experiment, the filtering response was evaluated and validated experimentally based on the hybrid all-integrated MPF with package. That is, the InP chip-based laser was turned on, and the microwave signals were input/output of the packaged MPF module through the peripheral RF connectors. The other measurement settings remained unchanged compared with those in the first experiment. In this way, evident bandpass filtering and band-stop filtering profiles were also able to be observed, as shown in Fig. 7(a) and Fig. 7(b), respectively. However, the filtering functionalities are only valid across a limited frequency range, and the performance metrics are far worse than those achieved in Figs. 5 and 6. The noise power spectral density was also evaluated at 3 GHz with 2 mA generated photocurrent, which is around $-151\mathrm{dBm/}\mathrm{Hz}$.

The damaged filtering performances are due to the severe internal RF crosstalk [23], mostly induced by the RF packaging operation. As seen in Fig. 7(c), when the chip-based InP laser is turned off and turned on, the two obtained $S_{21}$ curves only have difference within a low frequency range (DC-6 GHz). This unexpected phenomenon is because the distance between the RF transmission lines for the DDMZM and PD is relatively small; even if no optical signal exists, a portion of microwave signal from the input RF lines (connected to DDMZM) will be directly coupled into the output RF lines (connected to PD) [42]. Furthermore, considering the inevitable alignment errors in lens assembly, the optical coupling loss between the InP laser and the silicon chip is large (12.3 dB, see Appendix B), which will be converted into the loss in the RF domain. Therefore, we could not access to the full filtering performances of the integrated MPF by this measurement approach.

Nevertheless, the proposed hybrid integration scheme to realize an all-integrated MPF is practicable. With further improvements in RF packaging and optical coupling, much higher filtering performances can be envisioned. For example, the RF crosstalk can be effectively eliminated by placing the RF pads of DDMZM and PD at opposite chip edges to have larger spatial distance in future tape-out. Based on our current chip layout, to fully exploit the truly entire functionalities of this integrated MPF, the setup of the first experiment was adopted in subsequent measurements.

In addition to the filtering functionalities, the RF link performance is also an important issue, which should be considered for the integrated MPF, mainly including RF link gain, noise figure, and spurious-free dynamic range (SFDR) [43]. The link gain G is defined as the RF power ratio of the output signal to the input signal, which was measured by a VNA after calibration. For the noise figure (NF), the expression can be given by [5] $$\mathrm{NF}=P_N-G+174,$$(1)where $P_N$ is the detected electrical noise power, mainly resulting from the relative-intensity noise of the laser and the shot noise and thermal noise of the on-chip PD. The $P_N$ was extracted by measuring the RF power spectrum using an electrical spectrum analyzer (ESA, Keysight N9010B) at a resolution bandwidth of 1 Hz. Figure 8(a) displays the RF link gain and noise figure at various frequencies, which were both measured under band-stop filtering function. The maximum link gain reaches $-28.2\mathrm{dB}$ at 3 GHz, which is corresponding to a noise figure of 51.2 dB. It can be seen that the link gain gradually drops with the increase of the RF frequency, arising from the inherent declines in RF responses of the DDMZM and PD.

In order to obtain the SFDR, a standard two-tone test was carried out. Two RF tones at 3 and 3.01 GHz were generated by two analog signal generators (ASGs, Keysight E8257D and Anritsu MG3695A), combined by an RF coupler, and then applied to the DDMZM. The RF tones are located at the passband of the integrated MPF. An ESA was used to study the RF spectrum detected by the on-chip PD. The powers of the fundamental components (FUND, 3 and 3.01 GHz) and the third-order intermodulation components (IMD3, 3.02 and 2.99 GHz) were extracted under different input RF power, as shown in Fig. 8(b). Given a measured noise floor of $-151\mathrm{dBm/}\mathrm{Hz}$, the SFDR for IMD3 was calculated as $99.7\mathrm{dB}·{\mathrm{Hz}}^{2/3}$, which is higher than that of previous highly integrated MPFs [23,25] but still far from the target performance ($>120\mathrm{dB}·{\mathrm{Hz}}^{2/3}$ [31]) required for practical applications. The promising strategies to improve the SFDR will be discussed later.

To testify that our proposed integrated MPF holds the capacity to handle practical tasks in real-world RF systems, we performed high-resolution switchable RF filtering of analog signals toward two typical application scenarios. The first scenario is that the received signal of interest is accompanied by a spectrally closed strong interferer. In such case, the integrated MPF should operate as a highly selective notch filter to suppress the interferer for maximizing the signal-to-noise ratio (SNR) [28]. In the demonstration experiment, a targeted signal at 15 GHz and an unwanted interferer at 15.25 GHz were provided by two ASGs and combined by an RF coupler and then act as input of the integrated MPF. After passing through the integrated MPF, the output RF spectrum was measured and analyzed by an ESA. As shown in Fig. 9(a), if the inputs are located away from the filter stopband, the targeted signal has a power of $-42.3\mathrm{dBm}$ and the interferer has $-31.8\mathrm{dBm}$ (10.5 dB stronger than signal). Subsequently, the stop band of the integrated MPF is tuned to suppress the interferer. The filtered RF spectrum is depicted in Fig. 9(b). It can be observed that the interferer is drastically rejected with an ultrahigh suppression ratio of 43 dB, while the desired signal is well retained with a low attenuation of only 3.6 dB.

The second scenario is that multiple frequency bands simultaneously exist in RF environments, while each time only one channel signal is supposed to be isolated and processed. Under this situation, a center-frequency tunable bandpass filter is needed to carry out the channel selection [30]. To simulate this real-scene, two RF signals at 11 GHz as channel 1 and at 13 GHz as channel 2 with equal power were generated to serve as input of the integrated MPF. The integrated MPF was switched to the bandpass operation mode by adjusting ${\phi }_{\mathrm{DC}}$. By tuning the center frequency of the passband, we can flexibly select one channel to pass through, while the other channel is blocked. As shown in Fig. 10(a), when the center frequency of the filter passband is tuned to 11 GHz, the signal at channel 1 passes, while the signal at channel 2 is rejected. The power difference between channel 1 and channel 2 is up to 19 dB. If the filter passband is shifted to 13 GHz, in this way, the signal at channel 2 is selected, while the signal at channel 1 is effectively blocked with a suppression ratio of 16.5 dB, as depicted in Fig. 10(b).

These results above highlight the advanced filtering performances of the integrated MPF, in combinations of high operation frequency, high-resolution selectivity, and large rejection ratio. Meanwhile, the filtering response is proved to be agilely switched between the band-stop and bandpass for coping with dissimilar type of signals, which is expected to be quite useful for reducing RF system volumes and improving versatility.

To benchmark the overall performance of the proposed integrated MPF in this work, we make a comparison among state-of-the-art integrated MPFs and electronic microwave filters, as displayed in Table 1. First, we observe that the spectral resolution and suppression ratio achieved in our work are competitive with those of electronic microwave filters, while the frequency tunable range is much broader thanks to the intrinsic advantages of microwave photonics (MWP) technology. Second, for all the integrated MPFs with high integration degree ($\ge 50\%$), our work shows the most advanced filtering performances in combination of a broad frequency tunable range (from S-band to K-band), a narrow 3 dB bandwidth (360 to 470 MHz), and a high rejection ratio ($>40\mathrm{dB}$), which are comparable to or even better than other integrated MPFs with low integration degree ($\le 25\%$) and discrete MPFs [10–13]. Meanwhile, the filter type can be flexibly and rapidly switched ($\sim 48\mathrm{\mu s}$ response time) between bandpass and band-stop, which holds exciting benefits for the current dynamically changing RF environment. More importantly, compared with the InP monolithic solution to realize 100% photonic integration [23], our hybrid integration approach combines the foundry-enabled low-loss silicon photonics platform and readily processed InP chip-scale light sources and, therefore, is more cost-effective, provides a higher level of scalability, and offers new opportunities to overcome the limitations of individual material platforms [38].Table 1.

Performance Comparison of State-of-the-Art Integrated MPFs and Electronic Microwave Filters

Because the achieved filtering performances are capable of coping with the practical tasks, the remaining hurdle is the RF link performance. The lowest RF-to-RF insertion loss of our proposed integrated MPF is 28.2 dB, which is much larger than that of electronic filters. There are two major factors responsible for the RF insertion loss and they are the high optical link loss and the limited E/O and O/E conversion efficiencies. Hence, the RF insertion loss is possible to be greatly decreased if these two issues could be effectively solved. For reducing the optical link loss, the hybrid integration approach by micro-optics adopted in this work can be replaced by a more efficient photonic wire-bonding technique to achieve an ultralow coupling loss of 0.4 dB [35]. The main on-chip insertion loss derived from the silicon MZ modulator can also be reduced through optimization of the doped region in the waveguide to avoid unnecessary free-carrier absorption [50]. Alternatively, chip-based III-V/Si optical amplifiers could be considered for compensating the propagation loss [51]. For improving the efficiencies of E/O and O/E conversion, a highly efficient silicon MZ modulator with a U-shaped PN junction has recently been demonstrated to have a low half-wave voltage of only 1.7 V [52], and a Ge-Si PD based on whispering gallery structure has reached a high responsivity of 1.04 A/W [53]. Furthermore, utilizing the novel cascaded MRR scheme reported in Ref. [54] will also yield an improvement of link gain. In addition to the optimization of optical devices, the RF loss can also be compensated in the electrical domain by placing a low-noise RF amplifier at the input to the modulator.

Additionally, the SFDR of this integrated MPF is limited by the intrinsic nonlinearity of the silicon modulator, which may also need further promotion. Prospects to improve the SFDR of the silicon modulator have been illustrated by linearization approaches [55–57]. For example, via manipulating power distribution of a dual-parallel MZ silicon modulator, the SFDR for IMD3 has been improved to an ultrahigh value of $120\mathrm{dB}·{\mathrm{Hz}}^{6/7}$ [57]. The reduction of the RF insertion loss discussed above is beneficial to improve the SFDR as well.

Looking forward, with recent developments on large-bandwidth ($>60\mathrm{GHz}$) silicon modulators [58] and PDs [59], assisted by optimized silicon MRRs with ultrahigh $Q$-factor (loaded $Q>1.3\times {10}^6$) and large FSR ($>100\mathrm{GHz}$) [60], a hybrid integrated MPF featuring an operation frequency extending to the mm wave region (30 to 300 GHz) with a spectral resolution of $\sim 100\mathrm{MHz}$ is within reach. In parallel, the serious internal RF crosstalk arising from the packaging processes can also be eliminated by proper redesign of chip layout and PCB. Therefore, we can envision that our proposed integrated MPF will find important and immediate applications in various RF systems. Moreover, our hybrid design can be easily expanded to other MWP functional systems, such as optoelectronic oscillators [61] and instantaneous frequency measurements [62].

In summary, we have demonstrated an all-integrated high-performance MPF with bandpass/band-stop switchable filtering response, through hybrid InP and Si integration. This integrated MPF achieves high filtering performances in combination of a broad frequency tunable range, a narrow 3 dB bandwidth, and a large rejection ratio as well as agile ($\sim 48\mathrm{\mu s}$) switchable filtering function, which is superior to any highly integrated MPF reported before. For the band-stop filtering, a frequency tunable range of 3 to 25 GHz and a 3 dB bandwidth of 380 to 450 MHz are obtained, along with a high rejection ratio of $>40\mathrm{dB}$. For the bandpass filtering, the frequency tunable range and 3 dB bandwidth are 3 to 21 GHz and 360 to 470 MHz, respectively. With respect to RF link performances, an RF gain of $-28.2\mathrm{dB}$, a noise figure of 51.2 dB, and a spurious-free dynamic range of $99.7\mathrm{dB}·{\mathrm{Hz}}^{2/3}$ are achieved. To validate the capacity of this integrated MPF toward real-world application scenes, high-resolution RF filtering of realistic signals for both interference suppression and channel selection is carried out. It is anticipated that the hybrid integrated MPF demonstrated here could greatly advance the development of various RF application fields ranging from radar to wireless communication in near-future, for accessing to larger frequency-band, smaller footprint, and higher efficiency.

Acknowledgment. The authors thank Advanced Micro Foundry (AMF) for silicon device fabrication and Shenzhen PhotonX Technology Co., Ltd. for the support on packages.