Programmable photonic signal processor chip for radiofrequency applications

LEIMENG ZHUANG,* CHRIS G. H. ROELOFFZEN, MARCEL HOEKMAN, KLAUS-J. BOLLER, AND ARTHUR J. LOWERY Electro-Photonics Laboratory, Electrical and Computer Systems Engineering, Monash University, Clayton, VIC 3800, Australia SATRAX B.V., PO Box 456, Enschede, 7500 AL, The Netherlands LioniX BV, PO Box 456, Enschede, 7500 AL, The Netherlands Laser Physics and Nonlinear Optics group, University of Twente, PO Box 217, Enschede, 7500 AL, The Netherlands Centre for Ultrahigh-Bandwidth Devices for Optical Systems (CUDOS), Monash University, Clayton, Australia *Corresponding author: leimeng.zhuang@monash.edu


INTRODUCTION
Modern radio frequency (RF) systems, such as radio communications, radars, sensor networks, and THz-imaging demand ever-increasing bandwidth and frequency agility [1][2][3]. At the same time, they require devices that are small, lightweight and low-power, exhibiting large tunability and strong immunity to electromagnetic interference. Integrated microwave photonics [4][5][6][7], an emerging technology combining RF engineering and integrated photonics, has the potential to satisfy these needs. Harnessing the large bandwidth and tunability uniquely offered by photonic devices, it enables wideband, flexible front-end analog solutions to precede the digital signal processors that are currently limited to several gigahertz analog bandwidth [8].
Although wide-ranging in application, all such demonstrations to date are based on application-specific designs of a photonic integrated circuit. A more-flexible approach would be to design a universal photonic circuit whose topology can be reconfigured, post manufacture, similar to a fieldprogrammable electronic processor, e.g. a field-programmable gate array (FPGA) [30]. This would have variable circuit parameters [9][10][11][12][13][14][15][16][17][18][19][20][21][22], but also a flexible circuit topology to suit a wide range of signal processing functions. Recently, Pérez et al. presented the concept of software-defined processing in microwave photonic systems, addressing the anticipated flexibility requirements in future RF applications [31].
Here, we propose a suitable design for a FPGA-like photonic signal processor chip, as depicted in Fig. 1. The processor features full flexibility in circuit topology and full control of all Research Article circuit parameters in terms of both amplitude and phase. This unique combination is enabled by means of a grid of tunable Mach-Zehnder (MZ) couplers interconnected in a twodimensional mesh network topology ( Fig. 1(a)), with the MZ couplers being the inter-cell pathways. Each MZ coupler works as a photonic processing unit with freely programmable pathselecting, splitting/combining, and phase shifting capabilities. This makes it possible to define amplitude-and phasecontrolled optical routing paths in a two-dimensional plane and thereby create photonic circuits at will, such as the examples shown in Figs. 1(b) to 1(d). We anticipate this concept to be the starting point of transferring the inestimable enabling power of electrical FPGAs to photonic integrated circuits.
To give an experimental demonstration of this concept, we present here for the first time such a programmable photonic chip. For a proof of principle, the layout comprises a maximally simplified but yet fully versatile 2 × 1 mesh network with two cells. We show that the simple dual-cell circuit with a free spectral range (FSR) of 14 GHz and full parameter-programmability enables RF filters featuring continuous, overtwo-octave frequency coverage, i.e. 1.6-6 GHz, and variable passband shaping ranging from a 55 dB-extinction notch filter to a 1.6 GHz-bandwidth flat-top filter. Aiming for a new technology enabler, the results presented here pave the way for the realization of powerful photonic signal processing engines that will play key roles in the future high-bandwidth radiofrequency communication systems and networks. Figure 1(a) depicts a general waveguide mesh network comprising M × N square mesh cells, with the MZ couplers being the inter-cell pathways. Each MZ coupler has 2 × 2 connection ports, so it can be simultaneously connected to up to four other MZ couplers that constitute the mesh network. We implement the MZ couplers with phase tuning element ϕU on the upper arm and ϕL on the lower arm as shown in Fig. 1a. The transfer matrix parameters cij = Outi /Inj for such a coupler are:

DEVICE PRINCIPLE
where ϕA = (ϕU + ϕL)/2 and ϕD = (ϕU -ϕL)/2 respectively govern the phase and coupling coefficient of its output ports, and exp(j2πfΔτ) represents the frequency-dependent phase shift caused by the propagation delay Δτ of the coupler. By controlling ϕD, an MZ coupler is able to function as an arbitraryratio coupler ( 0 < sin(ϕD), cos(ϕD) < 1), or function simply as a length of 2-port waveguide with the coupler either in bar-status (sin(ϕD) = 1 and cos(ϕD) = 0) or cross-status (sin(ϕD) = 0 and cos(ϕD) = 1) as illustrated in Fig. 1(a). In the latter function, the cross-bar status of the MZ couplers determines the routing direction of the light from one such waveguide to the next in the mesh network, and the total length of a routing path can be defined by allowing the light to travel through a corresponding number of such waveguides. Figure 1(b) illustrates the basic circuit building blocks and their implementations in such a mesh network by programming the MZ couplers accordingly, including a 2 × 2 coupler, a length of 2-port waveguide, and a circular waveguide loop. Based on this programming mechanism, one can synthesize various circuit topologies in the mesh network. Figure 1(c) illustrates the implementations of the two general types of waveguide filters that are commonly used to perform signal processing, namely finite impulse response filters based on tapped-delay-lines and infinite impulse response filters based on ring resonators [32]. As far as  the mesh network dimension allows, one can reach circuit topologies with arbitrarily extendable functionality, an example of which is depicted in Fig. 1(d). It is important to mention that next to the freedom in circuit topologies, ϕD in the couplers and ϕA in the constituent waveguides provide the defined circuits with full control capabilities of the amplitude and phase of the light, which facilitates the complete function-programmability of the device. Figure 2(a) presents a first-demonstrator chip with two mesh cells, fabricated in a commercial Si3N4 waveguide technology (TriPleX TM [33], see supplement 1). To simplify the fabrication, we use dedicated phase shifters and MZ couplers with a single phase tuning element to perform respectively the effect of ϕA and ϕD of MZ couplers with two phase tuning elements. The phase tuning elements are implemented using resistor-based heaters which cause waveguide refractive index change by locally varying the waveguide temperature [10]. On this chip, the phase shifters are found with a full tuning range of 0 to 2π; the power coupling coefficient of the MZ couplers can be tuned very close to the ideal case, i.e., tunable between 0 and 0.99. By programming the values of the phase tuning elements, we demonstrate four distinctively different circuit configurations, including a single-ring notch filter [34], a single-ring Hilbert transformer [35], a dual-ring bandpass filter [36], and a dualring delay line [37]. The corresponding settings of the chip and the measurements of the frequency response shapes that verify the circuit functionalities are depicted in Fig. 2

RF filter implementation
Using the demonstrator chip as a programmable photonic signal processor, we implement a new microwave photonic approach of generating RF filters. A schematic of the system and an illustration of the working principle are presented in Fig. 3. Here, an electro-optic modulator is used to create a double-sideband modulation spectrum from an input RF signal under small signal condition [38,39] The chip is programmed into a circuit comprising a cascade of two ring resonators (as the delay line in Fig. 2(b)), whose resonance frequency and resonance strength are controllable via phase shifters ϕn and couplers κn, respectively [32]. We program the two ring resonators such that Ring 1 and Ring 2 have their resonance frequencies in the upper and lower modulation sidebands respectively, and both feature a sharp phase transition and a significant amplitude notch around the resonance frequency and nearly flat phase and amplitude there outside (for simplicity, we consider here only resonance effect for normal dispersions [32]). The equivalent RF responses of the two sidebands after direct detection are depicted alongside, assuming a high-speed photodetector providing sufficient RF bandwidth. The highlighted area exhibits a frequency region where the two RF responses have nearly equal amplitudes and a phase difference of π, in contrast to the equal-phase areas on its two sides. Eventually, these two RF responses add up vectorially at the photodetector output, resulting in a RF filter as illustrated in Fig. 3: a band-stop filter or a band-pass filter, depending on the phase relation between the optical carrier and the sidebands at the modulator output. In practice, a dualparallel Mach-Zehnder modulator can be used to provide the desired optical spectrum with either in-phase or complementary-phase sidebands (Fig. 3) by controlling the modulator biases [40,41]. Moreover, the programmability of the chip also allows us to implement a RF filter using a conventional microwave photonic approach based on singlesideband modulation [42], where the chip is programmed into an optical filter (e.g. a notch filter as in Fig. 2(b)). Although also easy to implement, this conventional approach requires an additional processing step to remove one modulation sideband, which increases the system complexity and leads to an extra 3-dB loss in the system gain.
To verify the approach illustrated in Fig. 3, measurements of RF filter responses were performed for both band-stop and bandpass cases (see supplement 1). In Fig. 4(a) and 4(b), the measurements show clearly that a band-stop and a band-pass filter can be generated, both having passband-stopband extinctions > 17 dB and passband dispersions < 2 ps/MHz. The fitted curves show that the measured filter shapes are consistent with the theoretical filter transfer function (see Supplement 1). In Figure 4(c) and 4(d), we demonstrate continuous tuning of the filter center frequency without changes in filter shape. This is performed by controlling the phase shifters (ϕ1, ϕ2) of the two ring resonators such that Δf1 and Δf2 (as referred to in Fig. 3) are shifted simultaneously with a constant ΔfRF = |Δf1-Δf2|. Subject to the frequency periodicity of ring resonators, the maximum frequency coverage of the RF filter equals half of the ring resonator FSR that is 14 GHz in this case. Here, we demonstrate the frequency tuning from 1.6 GHz to 6 GHz (31% of the ring resonator FSR), showing a frequency coverage greater than two octaves. It is worth mentioning that a two-octave frequency coverage in combination with continuous frequency tuning is difficult to achieve with electronic RF filters [43][44][45][46]. Besides, our RF filter employs only two tuning elements (ϕ1, ϕ2) to perform frequency tuning, implying easy control.  Next to the continuous frequency tuning, the full control capabilities of the phase shifters (ϕ1, ϕ2) and couplers (κ1, κ2) of the two ring resonators also allow for variable passband shaping. Figure 5 presents the measurements of many different filter responses ranging from a 55-dB extinction notch filter to a 1.6-GHz-bandwidth flat-top filter. In Figure 5(a) and 5(b), we demonstrate variable passband shaping by controlling the phase shifters (ϕ1, ϕ2). Unlike the operation for the filter center frequency tuning, Δf1 and Δf2 (as referred to in Fig. 3) are shifted independently in this case and the frequency difference between them ΔfRF = |Δf1-Δf2| determines the width and ripple of the filter shape. This effect applies to both band-stop and band-pass type of filters. From a practical perspective, however, such as in flat-top filters, it is undesirable to increase the passband ripple when increasing the filter bandwidth. This issue can be addressed by an appropriate setting of the couplers (κ1, κ2), which is shown in the measurements in Figure  5(c). We have achieved wideband flat-top filters with 1-dBbandwidth of up to 1.6 GHz (11% of the ring resonator FSR or equally 36% of the filter frequency coverage). In addition, when the two couplers (κ1, κ2) are set with identical coupling coefficients, a passband-stopband extinction of 25 dB can be reached, a measurement of which is shown in Fig. 5(d). Such RF filters with frequency agility and adjustable bandwidth have great application potential for high-spectrum-efficiency RF technologies such as cognitive radios [47].

DISCUSSION
The proper operation of such function-programmable photonic chips relies on the tunability of the MZ couplers, which translates to stringent design requirements for the coupler phase and coupling coefficient tuning range. Our demonstrator chip features good tunability (ϕ = [0, 2π], κ = [0, 0.99]), but the complexity of the programmable circuit topologies (functions) are limited by the small dimension of the mesh network (two mesh cells). However, by scaling up the network dimension, a myriad of functions that are based on more complex circuit topologies are expected to be implementable, such as tappeddelay-line filters, multi-channel (de)multiplexers and crossconnects, high-order coupled resonators, and various combinations of such circuits [32]. With sufficient space on the chip, it is also possible to implement multiple independent functionalities simultaneously, enabling a 'multi-task photonic processor'.
In view of such significant promise, it is required as well to discuss what challenges would be raised in fabrication and operation when increasing the network dimensions. For one thing, increasing network dimension means enlarging chip area. In our case, where Si3N4-waveguides are employed for realizing a demonstrator, the chip carries MZ couplers with a length of 3450 µm at a group index of 1.71 (including a heater section with a length of 2100 µm and two 3-dB directional couplers, each having a length of 675 µm). This means an area of about 0.35 × 0.35 cm 2 for one mesh cell and will for instance scale up by 100 times when aiming for a 10 × 10 mesh network. This bears higher risks of waveguide non-uniformity across the chip due to fabrication tolerance and may cause some degradation in device performance, such as via a limited tuning range of part of the couplers [32][33][34][35][36]. Regarding losses, the total device insertion loss includes the losses in the optical paths and coupling losses, in our case about 9 dB in total, dominated by two times fiber-chip coupling loss of about 4 dB/facet (which is expected to decrease effectively when particular waveguide designs or interposers are used to minimize the mode-profile mismatch at the coupling). However, a large network dimension may incur increased losses due to longer optical paths that are provided and due to possible power leakage at each coupler therein. Therefore, a low-loss waveguide technology is of great importance for the system performance, particularly for RF applications where some processing schemes have the system loss in quadratic relation with the optical loss [38]. Moreover, large network dimensions also mean a large number of tuning elements, the calibration and control of which requires a considerable engineering effort due to the possible initial offsets and interelement crosstalk. For device characterization and proper operation, dedicated power monitoring ports can be implemented as part of a circuit topology alongside the targeted functions. In this work, a commercial 12-bit control subsystem (SATRAX B.V.) is used, which is sufficient to implement designed algorithms for circuit parameter configuration and crosstalk compensation (yielding the results in Fig. 4 and 5). Further, when using thermo-optical tuning as is done here, a powerful chip temperature control setup may be required as the total heat dissipation of the chip scales with the number of tuning elements, causing possible increase in device size and weight. For the waveguide heaters in this work, an average power consumption of 0.25 W/heater is measured during operation, which is expected to be reduced by a factor of 5 when using optimized designs of the waveguide and heaters [24,33]. Many orders in power reduction might become available when implementing piezo-tuning [48], or, using other platforms, implementing electro-optic tuning [11,19,23]. From a broad perspective, a general solution to address the above concerns is a low-loss waveguide technology that enables further device miniaturization through higher index contrast and provides index tuning with high power efficiency and on shorter length scales. With this regard, silicon-on-insulator waveguide technology has shown interesting results [15,16,25,[34][35][36][37], enabling tunable MZ couplers with lengths of tens of micrometers. This offers to investigate realizing of the proposed waveguide mesh networks with more than an orderof-magnitude decrease in size (two orders in area). In addition, further device miniaturization also means that the FSRs of the circuits can be enlarged [32]. Our demonstrator chip is able to synthesize circuits with FSRs in the order of tens of GHz. Such FSRs are suitable for RF applications. A transition toward larger FSRs will considerably expand the application potential of such photonic signal processor chips, e. g., via tapped-delay-line equalizers, wavelength-division-(de)multiplexing, and reconfigurable add-drop multiplexers for optical communications [49].
Regarding RF filter implementation, we showed RF filter passbands with a frequency resolution in the order of GHz. This can be further scaled down by means of an according increase of the FSR of the processor chip. However, in the case of a sharper frequency resolution, for instance in the order of tens of MHz as required by mobile communication channels and satellite transponders [1,2], the quality of the CW laser is critical to the filter performance as the filter frequency stability relies largely on the laser linewidth and frequency jitter which may be significant compared to the filter bandwidth. Promisingly, kHz-linewidth lasers have been demonstrated on chip [50], which could be a low-cost solution to address this concern. Other drift-like fluctuations in the system control could be overcome by means of high-resolution tuning and adaptive control algorithms. The chip in this work employs thermo-optical tuning, so the programming speed is limited to the range of milliseconds. However, this could be significantly improved by the advancing of the modulator technologies, where the state-of-the-art devices have demonstrated modulation speed in the order of hundreds of picoseconds [51]. Moreover, our RF filter implementation is subject to the principle of a microwave photonic link [38]. This means that the same challenges with respect to system gain, noise figure, and dynamic range also exist. The key to overcome these challenges resides to a large extent in the advancing of optoelectronic components for the conversion between electrical and optical signals. With regards to these properties, promising results have been achieved in the last decade: system gain values larger than 10 dB have been demonstrated, and noise figure values below 6 dB at frequencies beyond 10 GHz, using special, highly-sensitive (low Vπ) modulators and with novel detectors that can handle high currents [52].

FUNDING INFORMATION
This research work is enabled by the funding provided from Dutch Agentschap NL IOP project PROMISE2DAY with no. IPD12009 and Australian Research Committee Laureate fellowship with grant no. FL13010041.
See Supplement 1 for supporting content.