Self-configuring programmable silicon photonic filter for integrated microwave photonic processors

Programmable photonic integrated circuits (PICs)^1–4 stand out as a prominent alternative within the domain of integrated microwave photonic (IMWP) processors.^5,6 The integration of reconfigurable optical cores, alongside modulation and photodetection subsystems, empowers the development of comprehensive tunable photonic devices. With these, we can effectively address the dynamic demands of wideband spectral processing in a wide variety of emerging application scenarios, such as flexible transceivers and multiplexers in high-speed fiber communications,⁷ and also for the implementation of upgradable interfaces for future communication systems.^8,9 Furthermore, these features are key to high-speed reconfigurable interconnections in computing systems as well as the realization of quantum logic gates.^10–13 Moreover, their implementation in a silicon photonic platform offers advantages, such as compatibility with existing CMOS fabrication processes, compact footprint, and low power consumption.

Among the large gamut of applications for IMWP processors, RF photonic signal filtering stands out as one of the most prevalent and widely adopted functionalities.^14–19 IMWP filters offer a remarkable ability to achieve ultra-wideband frequency tunability, providing a practical solution to the limitations commonly encountered with conventional electronic filters. RAMZI (Ring-Assisted Mach–Zehnder Interferometer) filters are an outstanding option due to their high-resolution spectral filtering and on-chip spectral processing capabilities.^20–27 This filter structure benefits from the advantages of all-pass filters to achieve arbitrary bandpass filters,²⁸ combining both FIR (Finite Impulse Response) and IIR (Infinite Impulse Response) filters. The integration of various filter types into a single device with programmable and reconfigurable capabilities makes it ideal for an extensive range of applications in optical signal processing and RF filtering.

The use of monitors to characterize and optimize RAMZI filters is a recently reported effective strategy.^29,30 However, the implementation of this approach in a field programmable photonic gate array (FPPGA) poses challenges.³¹ Accessing each of the rings from different ports in a mesh architecture results in increased power losses and an inefficient utilization of the device, wasting a significant part of the optical ports. In addition, the final packaging of the device becomes more complex due to the requirement of analog-to-digital converters (ADCs) and transimpedance amplifiers (TIAs) at the output of the photodetectors.

Here, we propose an optimization technique for a multipurpose RAMZI filter, and we demonstrate it with a fully tunable design of such a kind of filters, which only requires a single input and a single output optical port for both characterization and tuning processes and may dispense with an electrical monitoring system. By eliminating the need for additional monitoring components, our technique also circumvents the losses induced by the couplers required for that purpose. Moreover, the proposed filter is capable of implementing different architectures. By incorporating tunable couplers at the input and at the output, the filter arms can be selectively activated, enabling the structure to operate as an independent notch filter or as a Side-Coupled Integrated Spaced Sequence of Resonators (SCISSOR)-type RF filter³² with an improved spectral selectivity. These results are experimentally assessed by means of a fourth-order RAMZI filter, fabricated using silicon-on-insulator technology. Along this line, an auto-characterization algorithm has been developed to know each ring coupling curve. Subsequently, different types of filters have been configured, demonstrating their reconfiguration capabilities in bandwidth and central frequency. Bandpass filters with bandwidths ranging from 5 to 40 GHz are implemented and measured, tuning their central frequency from 0 to 50 GHz in both optical and RF domains. Furthermore, we demonstrate its operation as a SCISSOR filter using different types of modulation to achieve bandpass and band-stop filters, as well as its reconfigurability.

In the response of a digital filter, the design parameters are defined by its poles and zeros. From these parameters, we can synthesize different spectral responses in integrated photonic structures. Using an interferometric structure, we need to decompose the response of the designed digital filter into the sum of two all-pass functions. A simplified RAMZI filter is based on the use of a cascade of rings on each arm of the Mach–Zenhder Interferometer (MZI), thus implementing the two all-pass functions on each arm.³³ The total number of rings used gives the order of the filter, as each ring implements a pole–zero pair. Since this type of filter has the versatility of using programmable unit cells throughout the structure, we can create multiple types of filters and arbitrary responses by programming them. Figure 1 shows the proposed design for a general simplified RAMZI structure of order N.

By employing the algorithm within a RAMZI-like structure utilizing all-pass filters, it becomes feasible to construct arbitrary bandpass filters by combining FIR and IIR filters. Although lattice-type structures have been suggested for this purpose, they need a greater number of tunable couplers and phase shifters compared to filters of the same order. Alternatively, by equipping the filter with tunable couplers at the input and output, each arm of the filter can be used independently. This flexibility allows us to configure the rings separately, enabling them to be used as notch filters for processing or to create SCISSOR-type RF filters with narrower bandwidths.

To showcase the capabilities of the filter, we have successfully fabricated a fourth-order RAMZI filter using silicon-on-insulator (SOI) technology. The fabrication process involved 193-nm deep ultraviolet (DUV) lithography at the AMF SiP foundry. Figure 2 shows a detailed photograph of the fabricated device, highlighting a tunable ring. The overall structure of the filter, which includes pad arrays for electrical signals, measures 2980 × 1170 m. Strip waveguides with a height of 220 nm and a silicon core width of 500 nm are employed, whose propagation losses are 1 dB/cm. For the input and output ports of the filter, we have used 1D grating couplers, with insertion losses of around 5 dB/facet. The measured effective index value of this platform is around 2.38–2.41.

The tunable rings are designed with a tunable MZI consisting of 2 × 2 multimode interferometers (MMIs) and phase shifters to enable precise tuning of the ring’s coupling and resonance wavelength. MMI-based tunable couplers have higher insertion losses (0.5–0.7 dB) compared to a plain dual-drive directional coupler (0.2 dB), but the former enjoy broader bandwidth operation (>40 nm) compared to DC-based ones (around 10 nm).³⁴ The chip contains a total of 12 MMIs and 18 phase shifters, of which only 11 are driven, one of each tunable MZI and one of the two front phase shifters. The power consumption per π phase shift for each phase shifter is ∼1.3 mW. Hence, in the worst case, the device would consume at most 14.3 mW, which means a very low power consumption. The round trip length of the ring is set to 1310 µm, resulting in a 55 GHz FSR and a Q-factor of 2.28 ⋅ 10⁵. In place of simple directional couplers, tunable MZIs are also used at the input and output of the RAMZI structure, allowing for switching between the upper and lower arms. This configuration facilitates independent characterization and tuning of each ring without the need for monitoring, greatly simplifying the overall device. Moreover, it provides versatility in achieving different types of filters. For instance, by putting the input and output MZIs in the bar state, we can create a second-order side-coupled integrated spaced sequence of resonators (SCISSOR) filter. Alternatively, employing only one ring enables the creation of a single notch filter.

Throughout the entire structure, we incorporate thermo-optical phase shifters, which work in conjunction with trenches to avoid thermal crosstalk and achieve lower power consumption.

Tuning of the photonic filters is accomplished by precisely adjusting the resonance of each ring independently. To achieve this, it is necessary to characterize the coupling factor of each ring individually. In addition, since we have tunable MZIs at both the input and output of the filter, it is essential to characterize them as well in order to facilitate switching between the upper and lower arms to access the individual response of each ring. The block diagram in Fig. 3(a) shows the algorithm used for the characterization of these two components, separating it into two parts: first, the characterization of the input and output MZIs, and second, the characterization of the rings.

To characterize the input and output MZIs, a fixed-wavelength laser is utilized at the input of the filter. For each MZI, specific current values are set for one of the phase shifters. The optical output power of the filter at the bar port is then measured for all the different combinations of these current values, as depicted in Fig. 3(b). The minimum received power values correspond to totally opposite states (bar or cross) at the input and output MZIs, e.g., κ_i = 0 and κ_o = 1. This indicates that all the power is directed to the cross port of the filter.

By extracting the power curves for the remaining current values at these specific points, we obtain the output power curve as a function of the current set for one of the MZIs. In this configuration, the other MZIs are programmed to operate as a waveguide rather than an interferometer, enabling independent measurement. Normalizing the output power measurements, we can derive the transmission factor at the bar port, denoted as t_bar = sin²(θ/2).

Figure 3(e) shows the coupling factor and phase obtained for four different rings.

After fully characterizing the device, tuning is performed for the desired filter response. Figure 4(a) represents the followed procedure to set a filter with the obtained parameters from (5). The rings of the upper arm are set to a κ = 0 coupling to achieve a flat frequency response. Each ring is then separately tuned to the desired θ value using the characterization curves of Fig. 4(c). The extinction ratio of the resonance is optimized to match the simulated response of the ring, that changes the magnitude of the pole position. The ring phase is tuned by increasing the current in the corresponding phase shifter heater, causing a resonance shift in wavelength, which changes the angle of the pole position. Using a laser fixed at the desired resonance wavelength, the output optical power of the filter is measured until finding a power minimum. The current values for θ and ϕ of each ring are optimized in the same way for the remaining rings. Finally, the input and output couplings are set to κ_i = κ_o = 0.5, and one of the front phase shifters is tuned by increasing its current to find the final response of the filter by comparing the measured response with the simulated response. The minimum MSE provides the final response of the optimized filter. An optimization algorithm with progressive refinement is used to optimize each of the currents, both in the coupling and phase of the rings and in the front phase shifter. In this way, we can effectively find the optimal current value with sufficient resolution. This tuning method remains effective when operating within a band around the wavelength used for device characterization, since the MZI coupling factor is wavelength dependent.

Figure 4(b) shows the pole and zero diagram of a Chebyshev type II filter designed with a bandwidth of 20 GHz (36% of the FSR) and a stopband attenuation of 30 dB. Below, the table of parameters obtained from the decomposition algorithm is provided. In Fig. 4(c), we can see the simulated response of the filter in frequency together with the corresponding response that each ring has separately.

Both the characterization and tuning algorithms have been implemented in Python, together with the control of the instrumentation used, as well as the complete polynomial decomposition algorithm with which the filter parameters are obtained. Consequently, the script only requires the input parameters specifying the desired characteristics of the digital filter. The set of algorithms within the script takes care of the rest, streamlining the implementation process.

The characterization of the device, as well as the tuning of the different types of filters, was carried out by measuring the spectrum in wavelength. The PIC is placed on a thermoelectric cooler (TEC) with thermal grease paste, which is responsible for keeping the temperature of the device stable, together with a heat sink to dissipate the heat. Keeping the temperature stable during the execution of the calibration and tuning algorithms is crucial to preventing the ring resonances from drifting. Cleaved fibers have been used to enter and exit the chip through grating couplers.

To measure the optical spectrum, we have used a tunable laser together with a component tester (EXFO T100S-HP and CT440), both controlled through a Python script. To feed the electrical signals, we have used multi-channel voltage sources (Qontrol Ltd.), also controlled by the same script, which are connected to multi-contact probes (MPI TITAN™) with which we make contact directly with the chip pads.

Due to the inherent reconfigurability of the filter, a diverse range of designs can be implemented. This flexibility allows for the construction of IIR filters, such as Butterworth, elliptic, and Chebyshev of up to fourth order, contingent upon the number of rings employed. By leveraging the addition or subtraction of two all-pass functions, these filter types can be realized effectively. Notably, when both output ports of the filter are utilized, complementary responses are obtained, thereby enabling the device to function as an interleaver.

Figure 5 presents the experimental measurements of four distinct types of IIR filters, with different bandwidths. For all filter types, the optimization process employed the tuning algorithm discussed in Sec. IV. Notably, the measured responses fit perfectly with their corresponding simulated responses, considering the losses induced by the rings. Note that the frequency input to the Chebyshev type II design function sets the beginning of the stopband rather than the end of the passband as in the other designs.

The configuration parameters acquired during the calibration process for each filter can be preserved and utilized subsequently, as long as the temperature remains the same. It is important to note that alterations in temperature do not impact the shape of the filter; however, they do affect its center frequency.

The filters exhibit complete reconfigurability, allowing for adjustments in both the center frequency and bandwidth to cater to diverse applications. Figure 6 demonstrates the reconfiguration of the center frequency across the entire FSR. In an ideal scenario, this reconfiguration could be achieved by uniformly shifting the phase of all four rings. However, due to variations in the group index with wavelength and certain thermal crosstalk between rings, large frequency displacements are not feasible. Consequently, a partial tuning algorithm is necessary to move the resonances to the corresponding wavelengths at the four rings.

For bandwidth reconfiguration, the entire algorithm must be rerun, as both the amplitude and phase of the poles and zeros change completely compared to other bandwidths. Figure 7 shows the remarkable reconfigurability capacity of four distinct filter types, spanning bandwidths ranging from 5 to 40 GHz while maintaining a stopband attenuation of 30 dB. It is important to note that the passband losses increase as the bandwidth decreases due to ring round trip losses. Among the various filters, Butterworth and Chebyshev type II filters exhibit wider transition bands, while Chebyshev type I and elliptic filters demonstrate a faster roll-off. For example, the slope rate obtained for the elliptical and Chebyshev type I filters is 7.66 and 6 dB/GHz, respectively, in the filters measured with 40 GHz bandwidth. Meanwhile, for the Chebyshev type II and Butterworth filters, we obtained a slope rate of 4.95 and 4.42 dB/GHz, respectively.

To demonstrate the tunable microwave photonic (MWP) filter, single-sideband (SSB) modulation is employed. In this setup, we utilize a dual-drive intensity modulator (DD-MZM Sumitomo T.DEH1.5-40-ADC) in conjunction with a wave-shaper incorporating a stopband filter to eliminate one of the RF sidebands. A schematic diagram illustrating the microwave photonic system experiment is presented in Fig. 8.

Initially, an optical carrier generated by a laser source is split into two paths using a 90/10 optical splitter (OS). The 10% fraction is directed to the external part of the photonic integrated circuit (PIC), while the remaining 90% is subjected to modulation. The dual-drive MZM is modulated using a performance network analyzer (PNA), operating at the optimum quadrature bias (QB) point. The SSB-modulated signal is then amplified by an erbium-doped fiber amplifier (EDFA) before being coupled into the optical integrated filter via a cleaved fiber. Subsequently, the filtered signal is collected at the filter output using another cleaved fiber. Once outside the chip, the carrier is recombined with the signal using a 50/50 optical coupler (OC). Finally, the complete signal is further amplified by an additional EDFA before being sent to a photodetector and the network analyzer. Figure 9 displays the measurements obtained from an elliptical filter, demonstrating its tunability at five different frequencies. In this case, the frequency tuning was achieved by shifting the phase of the rings as shown in Fig. 6.

The circuit’s total reconfigurability enables us to achieve a versatile setup. By configuring the input and output MZIs in the bar state, we can effectively establish two distinct SCISSOR filters on each output port. These specialized filters offer significant advantages, such as narrower bandwidths, making them highly valuable in communication applications.

For its demonstration, the above structure has been configured to implement bandpass and band-stop filters in the RF domain. A modulated optical signal passes through the upper or lower arm of the filter, which is composed of two cascade ring resonators. The RF signal is modulated to generate a double-sideband (DSB) spectrum. One of the ring resonances needs to be in the frequency range of the lower sideband, and the other needs to be in the frequency range of the upper sideband. The two rings must be in critical coupling, as that will result in a sharp slope phase response. By means of direct detection, the two processed sidebands are down-converted back into two RF components and interfere with each other. When the resonances of the two resonators are aligned at different frequencies of the sidebands, the equivalent RF responses have a region with a phase difference of π.

Using an intensity modulator (such as a quadrature-biased MZM) produces two sidebands that are in phase with the carrier, while a phase modulator generates sidebands with a complementary phase relation. This particular frequency region will be translated into a rejection band in the case of intensity modulation or into a transmission band in the case of a phase modulation. Figure 10 shows the experimental results using the two types of modulation and demonstrates the bandwidth and center frequency tuning capability, being able to obtain filters from 1.5 GHz of bandwidth centered at 20 GHz or more.

In this work, we demonstrate a high-performance photonic filter that can be reconfigured automatically. The algorithm can adaptively tune the filter rings to achieve reconfiguration of bandwidth and central frequency. The programmability and scalability of these filters make them promising for a variety of applications, such as fully integrated RF photonic front-ends, beamforming, and image processing, due to their need for a rapid change in real-world scenarios.

Moreover, the proposed design brings the possibility of its incorporation into photonic processors since it is not necessary to use monitors for the calibration and tuning procedures. Embedding alternatives are used as a high-performance building block (HPBB) or for programming its structure within a mesh. The latter would require very short programmable unit cells (PUCs) to achieve rings with a short round trip length and a large FSR. Higher-order  IIR filters can be implemented for better performance on lower bandwidth filters, adding more rings to the design. Higher Q-factor in the rings, which reduces their losses as well as their length, is an improvement to consider for future designs, achieving higher order filters with low losses and a higher FSR.

We acknowledge the financial support by Huawei through Contract No. YB20200065124 and the Generalitat Valenciana Project No. PROMETEO/2021/015.

The authors have no conflicts to disclose.

C. Catalá-Lahoz: Conceptualization (equal); Data curation (equal); Formal analysis (equal); Investigation (equal); Methodology (equal); Software (equal); Validation (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). D. Pérez-López: Conceptualization (equal); Data curation (equal); Formal analysis (equal); Supervision (equal); Writing – review & editing (equal). T. Huy-Ho: Conceptualization (equal); Project administration (equal); Supervision (equal). J. Capmany: Conceptualization (equal); Data curation (equal); Funding acquisition (equal); Project administration (equal); Supervision (equal); Validation (equal); Writing – review & editing (equal).

The data that support the findings of this study are available from the corresponding author upon reasonable request.