Chip-Based Brillouin Processing for Microwave Photonic Phased Array Antennas

In this review paper, we provide perspectives on the implementation of high-performance, wideband, chip-based Brillouin microwave photonic processing subsystems for use in phased array antennas that require processing at every antenna element. We review recent advances in chip-scale Brillouin microwave photonic signal processing systems, including reconfigurable filters, frequency converters, tunable phase shifters, true time delays and microwave sources, which are key functionalities for phased array antennas. We discuss a roadmap for developments of Brillouin-based microwave photonic processors to reach the required performance and compact footprint for implementation in future dynamically reconfigurable phased array antenna systems.


I. INTRODUCTION
W IDEBAND and high frequency microwave systems are attracting significant interest for use with phased array antennas (PAAs) [1], [2], [3]. When these systems are combined with the ability to generate electronically steerable radiation patterns, unique functionalities, including high resolution radars [4], [5] capable of tracking multiple targets [6], and directed high data rate communications to multiple users [7] are enabled. Typically, the emitted radiation pattern of a PAA is controlled by changing the relative phase (or delay) between the radiating elements in a process that is referred to as beamforming [3], [8], [9]. To reduce the presence of side-lobes in the PAA radiation pattern, each antenna element is typically separated by half the operating wavelength [10], [11]. Consequently, the components used for beamforming and analog signal conditioning, which form the interface between subsequent frequency conversion and digitisation stages, must fit in the space between radiating elements; this becomes challenging at higher frequencies, for example above 15 GHz the element spacing is less than a centimeter. On-chip Brillouin microwave photonic processing hardware is at each element to implement the required beamforming functionality. The inset shows a silicon photonic integrated circuit, a candidate hybrid platform for integrated Brillouin processing.
As an example, RF filters, a key component within PAAs, face challenges achieving wideband operation in the required compact form factor. Conventional RF filtering approaches based on surface and bulk acoustic wave processing achieve low losses and large out-of-band rejection [12]. However, they are not suitable for millimeter wave frequencies due to large acoustic losses [12], [13]. Reconfigurable filters based on cascaded banks of high-pass and low-pass filters offer discretely reconfigurable bandwidths and central frequencies [14]. However, many parallel filters are required to enable operation over a wide frequency range, increasing both system complexity and size. Filters based on ferromagnetic and ferrimagnetic materials, including yttrium iron garnet (YIG), achieve wide frequency tunability using compact cavities [15], [16], but require an external magnetic field to control their resonant frequency, complicating integration. While the above-mentioned devices exhibit impressive performance, there is no clear RF filter technology that can operate over a wide frequency range and be implemented in a compact footprint that is compatible with element-level processing for high frequency PAAs [1].
Microwave photonics (MWP) is well positioned to overcome the challenges associated with the increasing bandwidth and carrier frequencies [17] of next-generation wireless communications [18] and radar [19], [20] systems. Systems that utilise MWP signal processing achieve wideband and high frequency operation by transducing the RF signal into the optical domain using an electro-optic (E-O) modulator. The RF signal is then This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ processed using well established photonic components, that support operation over wide frequency ranges, before conversion back to the RF domain through photodetection [21]. The bandwidth of the MWP link is limited only by the bandwidth of the E-O modulator and photodetector (PD), which convert between RF and optical domains.
On-chip MWP processors, for example RF photonic filters, have been demonstrated using integrated optical micro-ring resonators (MRRs) [32], [45], [46]. Although good performance has been achieved, MRR-based processing has limited spectral flexibility, as the shape of the photonic resonance is determined by the intrinsic waveguide losses and fixed structural properties of the MRR. Alternatively, programmable photonic mesh processors offer flexible RF responses [47], [48], but increasing the number of mesh processors to support additional RF processing functionalities requires longer waveguide lengths, which comes at the expense of increasing RF losses and size [47]. MWP transversal filters can be created using multiple optical carriers propagating through on-chip dispersive delay lines [28], [49], but require many optical carriers to achieve sub-100 MHz processing [50], consequently increasing system complexity.
Integrated MWP beamforming elements highlight the advantages of photonic integration by enabling devices with compact form factors that operate at high frequencies and support wide RF signal bandwidths [32]. Chip-based MWP beamformers have been demonstrated using switchable delay lines [33], [51], but can only produce discrete delays, limiting beam steering flexibility. Continuously tunable delay lines have been implemented using broadband dispersive effects from waveguide gratings [52], [53], by changing the coupling ratios of MRRs [32], [34], [42], or using delay line interferometers [43]. However, these approaches lack the required frequency flexibility, agility and resolution due to the fixed device structural properties and difficulties in scaling the device size, as outlined above.
Stimulated Brillouin scattering (SBS) [54] has enabled microwave photonic signal processors with unmatched frequency selectivity, tunability, and reconfigurability [17]. Inducing SBS in chip-scale platforms greatly reduces the form factor of Brillouin devices, when compared to their fibre-based predecessors [55], [56], [57], [58]. SBS facilitates desirable fine spectral resolution (MHz-level), owing to the intrinsic narrow bandwidth of the Brillouin resonance, which exhibits a Lorentzian gain and loss response with narrow bandwidth of approximately 30 MHz [54], unmatched by other optical processing methods. Because the Brillouin resonance is optically induced, its spectral response can be dynamically reconfigured by tailoring the optical pump spectrum [59], [60], [61], enabling tailored filter shapes that can be tuned over multiple octaves. The narrow bandwidth resonant response of SBS forms the basis of programmable phase and amplitude control used in MWP processors and has been demonstrated in different chip-based platforms, including chalcogenide waveguides [62], [63], suspended silicon on insulator (SOI) waveguides [64], [65], [66], and high quality factor (Q) Si 3 N 4 [67] and SiO 2 resonators [68].

II. PHASED ARRAY ANTENNAS
PAAs are widely used in communications [7] and radar [5], [80] systems, enabling the transmission and reception of RF energy in an electronically controllable manner. Traditionally, PAAs consist of a two-dimensional array of antenna elements spaced by half the operating wavelength, as shown in the artistic impression in Fig. 1, which depicts a chip-scale Brillouin processor that implements beamforming at each PAA element. The inset shows a compact silicon-based hybrid photonic circuit, which is a candidate technology for the Brillouin processor that we outline in Section V.
Beamforming can be implemented in PAAs by introducing a phase shift, Δφ, between elements as shown in Fig. 2. To direct the beam to an angle of θ, each successive wavefront must be delayed an additional path Δx, given by [5], [9], [80], In transmit mode, (i) a baseband signal from the DAC output is filtered, and (ii) converted to the PAA operating frequency before (iii) a delay and amplitude weighting is applied at each element before it is transmitted. In receive mode, (iv) a signal is received at each antenna element. Delay and amplitude weighting is applied at each element before (v) the spectrum is filtered to select a band of interest and converted to the baseband before (iv) the signal is filtered and undergoes analog to digital conversion. Essential microwave processing in this diagram has been demonstrated using on-chip Brillouin devices. Microwave processing and functional devices are shown in (b)-(e). (Fig. 3(b) from [72], © 2020, Optica Publishing Group. Fig. 3(c) from [71], © 2016, Optica Publishing Group. Fig. 3(d) from [69], © 2020, Optica Publishing Group. Fig. 3(e) from [77], © 2017, Optica Publishing Group.) where d is the distance between adjacent elements in the PAA. The additional path Δx is caused by the applied phase shift, and is related to the operating wavelength, λ, by, Combining (1)- (2), and expressing the angle of the PAA beam in terms of the operating frequency, f , we get where c is the speed of light. As can be seen from (3), one drawback of conventional beamforming using phase shifts is that the beam direction changes as a function of the transmitted frequency, referred to as beam squint [5]. Beam squint can be eliminated by introducing a TTD between elements to control the beam angle [81] by applying a linear phase shift to the RF signal of interest. The resultant delay, τ , is the derivative of the applied phase shift with respect to frequency [77], [78]. Substituting the relationship between the applied phase shift and frequency into (3), we find that the beam angle no longer directly depends on the PAA operating frequency [5] highlighting the advantages of using TTDs to implement beamforming in PAAs with wide signal bandwidths.
To understand the requirements for phase shifters and TTDs for beamforming, we solve for Δφ and τ in (3) and (4), respectively. Assuming that the PAA consists of N x N elements spaced by half the operating wavelength, d = λ/2, we find that the maximum phase shift and delay required is given by: Using (5) and (6) we calculate the required phase and time delay to achieve steering over, θ = ±90 • . For PAAs based on phase shifts, a full 360 • is required. For beam steering approaches based on TTDs, the required delay decreases with increasing frequency, and is less than 1 ns for 16 x 16 arrays above approximately 7.5 GHz, and only 150 ps at 50 GHz. Fig. 3 illustrates a typical signal chain of a PAA, which includes beamforming hardware at each antenna element, frequency converters and local oscillators (LO) to interface between the PAA operating frequency and baseband, filters for analog signal conditioning, and baseband digital electronics; more information is provided in Refs [9], [82]. In transmit mode, a digital-to-analog converter (DAC) converts a digital input to an analog baseband signal at point (i) in Fig. 3(a). This signal is filtered by a fixed frequency filter at the baseband to remove spurious content from the DAC. A local oscillator (LO) drives a mixer to convert the generated signal from the baseband to the PAA operating frequency, and is filtered again to remove spurious mixing terms at point (ii). The signal is then distributed to each antenna element of the PAA. Before transmission, delay and amplitude weightings are applied at each element to control the beam direction at point (iii).
The same concept applies for receive beamforming, but the signal chain is reversed, with an analog-to-digital converter (ADC) being used to digitise the received signal. In receive mode, the RF filter serves an additional purpose to select a signal of interest from the RF spectrum. Band-pass filtering may be required to select a particular band of interest, or a notch filter with fine frequency selectivity may be required to reject a particular interfering signal without attenuating adjacent signals at nearby frequencies. Fig. 3(b)-(e) highlight demonstrations of Brillouin-based MWP processing devices that perform the key functions within a beamforming network, including frequency converters [72], low phase noise RF sources [71], filters [69], phase shifters [77] and TTDs [78]; we expand on these examples and conclude with a perspective on how these distinct functionalities can be combined onto a single photonic chip to enable element-level processing for PAAs.

III. MICROWAVE PHOTONIC LINK PERFORMANCE
MWP systems enable wide bandwidth processing of RF information with optical components using the canonical architecture shown in Fig. 4(a). An RF input signal is modulated onto a continuous-wave (CW) laser, where the RF spectrum is translated to the optical sidebands of the laser. This optical signal then undergoes processing in the optical domain before a photodetector (PD) converts it back to the RF domain through square law photodetection [83]. Although MWP systems enable operation over wide frequency ranges, they must also operate with low loss, low noise, and high linearity to ensure RF signal integrity is maintained [17].
The first metric discussed is the RF link gain, defined as the ratio of RF output power to RF input power. For MWP links using intensity modulation (IM) and direct photodetection, the link gain is given by [83] where r pd is the responsivity of the PD, R L is the RF load resistance, P opt is the laser output power, L is the loss of the MWP link, φ B , and V π RF are the bias angle and half-wave voltage of the E-O modulator, respectively. The output noise power spectral density (PSD), p n , of an MWP link is determined by three major contributions and is expressed as [21], [83] p n = (1 + g MW P )kT + qr P D I P D R L 2 + 10 where q is the electronic charge, I P D is the average photocurrent, RIN is the relative intensity noise of the optical signal at the PD, k is the Boltzmann constant, and T is the device temperature. The first term in (8) corresponds to thermal noise, the second term refers to the shot noise contribution, and the final term corresponds to the RIN noise contribution to the overall link noise. The noise PSD depends on the average photocurrent, which may be expressed as [83] The key figure of merit to assess the noise degradation of an MWP link is the noise figure (NF), which quantifies the degradation of the signal-to-noise ratio (SNR) in an RF system and is given by [21], [83] where P SD is the measured noise power spectral density of the MWP link output, G MW P is the MWP link gain in dB scale (G MW P = 10 log 10 (g MW P )), and 174 represents the thermal noise power spectral density at 290 Kelvin. The third important metric of an MWP link is linearity. This must be considered to reduce the effect of unwanted harmonic and intermodulation distortions, which are introduced by nonlinearities in the MWP link. The linearity of an MWP link is quantified by the spurious-free dynamic range (SFDR), and is measured using the two tone test [83], where two tones are input to the MWP link, as shown in Fig. 4(b). While many nonlinear mixing terms are created, the third order intermodulation products are of particular interest, as their frequencies are close to the frequency of the input tones [83], as shown in Fig. 4(c). SFDR n is the power difference between the fundamental tone and the n th order modulation products when their power is equal to the noise power in a 1 Hz bandwidth, as shown in Fig. 4(d). The SFDR n corresponds to the maximum signal-to-noise ratio (SNR) that is achievable without the need to filter the n th order distortion terms [29].
A number of optical processing techniques that improve the key performance metrics of MWP links have been demonstrated. For example, the NF of MWP links can be reduced using the technique known as low biasing [21], [83], where the bias angle is moved away from quadrature (φ B = π/2), towards the null bias point (φ B = 0). This may be understood by examining (7)- (10) and noting that the noise PSD decreases at a faster rate than the link gain decreases when the bias angle is reduced. Additional MWP link optimisation techniques include optimising the optical carrier to sideband ratio, whilst maintaining constant optical power at the PD to increase gain [69], [84], [85], reduce NF [86], and reduce third order non-linearity [85]. Additionally, optical processing has been used to ensure that RF intermodulation products destructively interfere. This reduces distortion and increases the SFDR of MWP links [87], [88]. Although optical amplifiers can be used to increase the RF link gain [89], they tend to be excessively noisy. Therefore, their location in the MWP link must be carefully considered to mitigate the excess noise they generate [90], [91].
When the pump and probe have different frequencies, they create a travelling wave interference pattern. If the interference pattern travels at the acoustic velocity in the medium, and has sufficient intensity, it modifies the density of the waveguide through electrostriction and creates an acoustic wave [97], [98]. In turn, this leads to a change in refractive index through the photoelastic effect, and forms a travelling optical grating. This occurs when the pump and probe differ by the Brillouin frequency shift (BFS) in the medium, Ω B . When incident upon this grating, the pump is Bragg reflected and Doppler shifted by frequency Ω B , and constructively interferes with the probe. The increased probe power causes the intensity of the interference pattern to increase, which in turn amplifies the generated acoustic wave and resonantly enhances the SBS process. A common schematic used to generate backwards SBS in MWP processors is shown in Fig. 5. The RF input is modulated on the optical probe, and is counterpropagated against the pump in the Brillouin medium. Circulators separate the counterpropagating pump from the probe before photodetection. The magnitude, M , and phase, Φ, of the SBS response are imparted on the optical signal and mapped into the RF domain. When Brillouin gain is stimulated by a pump with frequency ω p , the magnitude and phase of the Brillouin response at frequency ω is given by [101]: Γ B is the linewidth of the Brillouin resonance and is inversely proportional to the phonon lifetime [102]. The parameter, G, is related to the intrinsic Brillouin gain coefficient, g 0 , the acoustooptic effective area, A ef f , pump power, P p , and the effective length of the Brillouin medium, L ef f , by [101] We note that it is often convenient to refer to the Brillouin gain coefficient, G SBS = g 0 /A ef f , which takes into account the intrinsic gain of the Brillouin medium and the acousto-optic effective area, which is determined by the waveguide geometry. The effective length is related to the physical waveguide length, L, and waveguide propagation loss, α, by

B. Integrated Brillouin Platforms
Brillouin scattering has been induced in numerous chip-scale platforms [107] using three different strategies, as outlined in Fig. 6 and summarised in Table I, which details the key metrics used to characterise Brillouin platforms. The first approach is based on soft chalcogenide glass waveguides, which provides good confinement for both optical and acoustic modes because the speed of sound in the cladding is faster than in the core [54]. The second class of devices use silicon photonic waveguides, in which the acoustic mode is not strongly confined for traditional   [54]. The acoustic losses are reduced by suspending silicon waveguides above the substrate, so that it is essentially surrounded by air, thereby increasing mechanical isolation [64], [65], [66]. The third approach is the use of high Q factor resonant cavities, where long optical lifetimes enhance the generation of Brillouin gain, allowing for the use of materials with low Brillouin gain coefficients if they have low loss [67], [68].
The generation of backward SBS in As 2 S 3 rib waveguides [62], which was the first demonstration of on-chip SBS, shown in Fig. 6(a), significantly reduced the interaction length required to generate SBS gain compared to previous fibre-based systems, and inspired the development of additional on-chip Brillouin microwave processing platforms based on chalcogenide glasses [104], [108]. Brillouin devices based on chalcogenide glasses combine strong acoustic confinement with negligible nonlinear loss, and have achieved more than 50 dB of Brillouin gain [103] using waveguides with lengths of a few centimeters that can be routed into compact and low-loss spirals. For example Morrison et al., achieved 22.5 dB of gain in an on-chip area less than 1 mm 2 [63], as shown in Table I. Hybrid Brillouin platforms use multiple materials to enhance the functionality and performance of traditional As 2 S 3 Brillouin devices. For example, integration of As 2 S 3 waveguides with SOI platforms [63], shown in Fig. 6(b), combines the benefits of the large magnitudes of SBS gain offered by chalcogenide waveguides with the availability of passive and active components available in CMOS silicon photonics. Fig. 6(c) illustrates a typical SOI circuit that was fabricated by the IMEC foundry; this device aims to reduce the form factor of Brillouin processors by integrating As 2 S 3 waveguides with active E-O components. Additionally, Si 3 N 4 and As 2 S 3 PICs [78], [93], as shown in Fig. 6(d), have been cascaded to enable MWP notch filters with reduced pump power requirements, and TTDs using separate carrier tuning; we expand on these devices in Section IV.
Although integration of multiple materials enhances functionality, the transitions between different materials must be carefully designed to minimise reflections of the Brillouin pump [94], which increases the spurious content in the RF output [93]. For example, Lai et al. measured a 4% reflection from the interface between lensed tip fibre and monolithic As 2 S 3 chip [94], which is lower than the reflection at the interface between air and bulk As 2 S 3 (≈17%) owing to the use of lensed-fibres and the anti-reflection coating at the facet of the As 2 S 3 chip [60]. Further reducing reflections of the Brillouin pump is desirable to minimise spurious content in the output of the MWP link [93]. To reduce the reflection experienced by the Brillouin pump, a mode-matched and vertically tapered transition between As 2 S 3 and germanosilicate waveguides [94] facilitated an adiabatic transition between materials, as shown in Fig. 6(e). This device achieved minimal optical back-reflections of only ≈ 0.04% and promises to improve the performance of integrated Brillouin MWP processors [93].
Whereas SBS is typically observed in the backward direction in optical fibres and chalcogenide planar rib waveguides, it manifests in the forward direction for SOI devices. As mentioned, standard silicon waveguides lack intrinsic acoustic guidance, which causes acoustic modes to leak rapidly from the waveguide core, inhibiting the generation of Brillouin gain [109]. However, the phonon mode can be confined if the waveguide is suspended, thereby reducing acoustic losses and supporting the propagation of transverse acoustic waves, leading to the generation of forward Brillouin scattering [109]. As seen in Table I, silicon-based Brillouin platforms have large Brillouin gain coefficients and shorter interaction lengths than chalcogenide devices. However, as mentioned, they suffer from nonlinear loss at high pump powers, limiting the magnitude of Brillouin gain that can be generated [54]. Despite the aforementioned limited intrinsic acoustic guidance and nonlinear losses, SBS has been demonstrated in a number of SOI platforms [64], [65], [105].
Acoustic confinement has been achieved in an SOI platform using an underetched waveguide [64], as shown in Fig. 6(f). However, the magnitude of the generated Brillouin gain saturated at approximately 1 dB, limited by optical losses and freecarrier absorption at higher pump powers [64]. Subsequently, 30 dB of on-chip Brillouin gain was generated in an SOI platform using a Brillouin active membrane waveguide structure [106], shown in Fig. 6(g). However, the Brillouin-active waveguide was placed in an optical resonator to increase the magnitude of the generated Brillouin gain, limiting the functional bandwidth of the device.
One challenge faced by forward Brillouin devices is that the strong co-propagating pump must be filtered to avoid the generation of spurious terms caused by beating between the pump and probe at the PD [75]. The photonic-phononic emitter-receiver (PPER) structure demonstrated by Shin et al. [66] shown in Fig. 6(h), overcomes this challenge by placing the Brillouin pump and probe in spatially separated waveguides. In this approach, the pump is injected into the emitter waveguide and is modulated by an RF signal with a frequency that coincides with the Brillouin frequency shift (BFS). This generates an acoustic wave through forward Brillouin scattering [66], which modulates the phase of the optical signal in the adjacent, receiver waveguide through photoelastic coupling [66]. The signatures of acoustic waves with narrow linewidths as low as 2.4 MHz [110] have been observed, facilitating MWP filters with fine frequency resolution, as we discuss later in this section.
High Q optical resonators enable increased interaction times between the pump and probe, allowing for materials with low Brillouin gain coefficients to be used, and facilitates narrow linewidth optical tones that are suitable for generating stable RF signals. For example, silica wedge resonators are attractive for SBS-based processing due to their high Q and compact form  Table I, Si 3 N 4 MRRs are able to generate Brillouin gain, owing to their extremely low losses.

C. Brillouin Microwave Photonic Filters
Microwave filters are essential components used to attenuate unwanted spurious or interfering signals in the RF spectrum [29]. MWP filters are required to exhibit wideband frequency tunability and fine frequency selectivity to attenuate unwanted signals whilst preserving nearby signals of interest. Therefore, Brillouin processing is particularly well suited for use in MWP filters as its fine spectral resolution enables narrow passbands or notches [17], [29].
The modulation scheme used in MWP filters has a significant impact on filter functionality and performance [29]. By altering the optical modulation scheme, both bandpass and notch filter functionalities can be realised [111]. Initial demonstrations of chip-based Brillouin MWP filters used single sideband modulation (SSB), and consequently were limited by a one-toone spectral mapping of Brillouin gain to RF filter rejection ratio [111], [112], [113]. This is undesirable as it demands a considerable amount of pump power to generate large RF suppression [114]. For the remainder of this section, we outline approaches that combine advanced modulation formats [29], [111] with amplitude and phase control offered by SBS to enable high-performance MWP filters that do not require excessive Brillouin pump powers.
Bandpass filters enable transmission of a particular frequency range and are used to suppress out-of-band signals over a broad frequency range. In Brillouin-based processing systems, bandpass filters are realised using a modulation scheme that results in broadband destructive interference of output RF photocurrents in the absence of any Brillouin resonance. However, when SBS is applied to one sideband, it breaks the condition for destructive interference and forms a narrow passband [69], [111]. Fig. 7(a) depicts one modulation scheme that provides this functionality [69], [85]. Fig. 7(b) highlights how the parametric generation of Brillouin gain enables reconfigurable MWP filters, demonstrating passband bandwidth broadening from 30 MHz to 440 MHz [61], by broadening the Brillouin pump spectrum using an E-O modulator. In this demonstration, the probe was phase modulated, which led to destructive interference of RF photocurrents outside the Brillouin resonance [111] and increased the MWP filter RF rejection ratio compared to MWP filters based on SSB modulation. Filters with widely tunable passbands from 1-20 GHz have been demonstrated by changing the relative frequency between the pump and probe, shown in Fig. 7(c). Rejection ratios exceeding 40 dB are achieved using only 14 dB of on-chip SBS gain [69]. Multiple passbands are generated by using additional Brillouin pump tones, shown in Fig. 7(d).
Notch filters are used to attenuate unwanted signals at a specific frequency. RF notches are formed by Brillouin-based devices, which use the SBS resonance to ensure RF photocurrents destructively interfere [115], [116]. One way to achieve this is shown in Fig. 8(a), where the upper and lower sidebands of the probe are π out of phase and mismatched in amplitude by the magnitude of applied Brillouin gain [115], [116]. At the Brillouin resonant frequency, the upper and lower sideband are equal in amplitude, but π out of phase, and therefore their RF photocurrent contributions destructively interfere upon photodetection. Outside the Brillouin resonance, the upper and lower sidebands are not equal and therefore do not cancel. Fig. 8(b) shows the modulation scheme used by Marpaung  shown in Fig. 8(c). Similarly, Casas-Bedoya et al. required only 1 dB of forward Brillouin gain to demonstrate a widely tunable filter response using a SOI platform, as shown in Fig. 8(d) [116]. One advantage of the modulation scheme in Fig. 8(a) is that the pump power required to form a notch is greatly reduced when compared with non-interferometric approaches. However, the broadband π phase difference between upper and lower sidebands creates an RF gain penalty when the upper and lower photocurrent contributions are added.
Combining on-chip linear optical devices, for example MRRs, with Brillouin processing [93], [117] is an attractive method to implement an MWP notch filter without a gain penalty in the passband. In this notch filter scheme, an over-coupled MRR imparts a spectrally localised π phase shift to the upper sideband at its resonant frequency. Brillouin gain is then applied at the MRR central frequency to overcome the reduction in signal amplitude caused by the MRR resonance. The upper sideband therefore experiences a localised π phase shift with an unchanged amplitude, while the lower sideband is equal in amplitude and has unchanged phase, as shown in Fig. 9(a). An RF notch is formed upon photodetection as the photocurrent contributions from upper and lower sidebands are equal in magnitude at the overlapped MRR and SBS resonant frequency, but π out of phase.
The first demonstration of this concept used optical fibre as the SBS medium [118] and was subsequently demonstrated in a monolithically integrated As 2 S 3 platform [117]. The measured magnitude and phase response of the As 2 S 3 MRR is shown in Fig. 9(b), where a π phase shift is observed at the MRR resonance. The SBS gain response in As 2 S 3 waveguides is shown in  Fig. 9(c), which does not impart any phase shift at its resonant frequency. The RF filter response using the monolithically integrated As 2 S 3 platform exhibits tunable notch frequencies by changing the probe frequency relative to the MRR resonance, as shown in Fig. 9(d). Propagation losses in the As 2 S 3 MRR limits the 3-dB bandwidth of the RF notch to 3 GHz [117]. The key to the passband link gain improvement offered by this scheme is attributed to the fact that the SBS and MRR responses do not impart a phase or amplitude change outside of their resonance, as seen in Fig 9(b)-(c). This results in constructive interference of RF photocurrents in the passband from the in-phase IM sidebands, and is not the case for the Brillouin-based notch filters discussed above, where the sidebands are out of phase with each other and do not sum constructively in the passband which increases RF losses [115], [116].
More recently, an MWP filter scheme was demonstrated that combined tunable and narrow linewidth Si 3 N 4 MRRs cascaded with As 2 S 3 waveguides to enhance functionality and reduce the 3-dB RF notch bandwidth from 3 GHz to 500 MHz [93]. Up to three low-loss Si 3 N 4 MRRs were overlapped with three Brillouin gain resonances to enable an MWP filter with three independently controllable notches, as shown in Fig. 9(e). The filter was demonstrated to suppress undesired signals by up to 49.4 dB, which were 500 MHz from a signal of interest, shown in Fig. 9(f)-(g).
PPER waveguides, as outlined in Section IV, offer an alternate approach for implementing MWP bandpass [120], [121] and notch filters [110], which circumvent the need to filter the high-power Brillouin pump, as demonstrated by Gertler et al. An MWP filter with a passband centred at the BFS, Ω B , is realised by encoding the RF input signal onto the pump laser through intensity modulation [120], as shown in Fig. 10(a). If the RF input contains a signal with frequency Ω B , an acoustic wave is generated, which modulates the phase of the CW laser in the adjacent receiver waveguide. An optical bandpass filter removes one sideband of the phase modulated signal so that when this filtered optical signal is converted to the electrical domain, the RF signal is present in the device output. However, input RF signals that do not coincide with the BFS do not induce an acoustic wave, and are therefore suppressed by the MWP filter. Fig. 10(b) shows the bandpass RF filter response of the fabricated device with a BFS of 3.86 GHz and 3-dB bandwidth of 3.5 MHz [122]. The passband frequency is changed by using phase modulation to encode the RF information onto the pump, and introducing a second laser to the pump spectrum, which has a frequency offset from the original pump laser [120]. Tuning of the passband frequency is shown in Fig. 10(c). MWP notch filters have been demonstrated using PPER waveguides [110], as shown in 10(d). In this scheme, the PPER waveguides select a narrow region of the RF spectrum, which is subsequently subtracted from the original input signal, forming a notch filter through destructive interference. Using this technique, an RF notch filter, as shown in Fig. 10(e), with rejection of up to 57 dB has been observed with bandwidth as narrow as 2.7 MHz [110].

D. Brillouin Phase Shifters
Phase shifters and TTDs are essential components responsible for beamforming in PAAs. The ideal RF phase shifter provides a 360 • tunable phase shift with flat amplitude response over a wide RF frequency range. Brillouin-based MWP phase shifters typically use SSB modulation and apply a phase shift to the optical carrier from the phase shift associated with the Brillouin resonance, given by (12). This phase shift translates to a broadband phase shift in the RF output [90].
Pagani et al. demonstrated a MWP phase shifter using on-chip SBS [74] by applying equal amounts of Brillouin gain and loss to impart a phase shift to the optical carrier with minimal amplitude fluctuations, as shown in Fig. 11(a). The generated RF phase shift was tuned by changing the Brillouin pump powers, as shown in Fig. 11(b). A continuous RF phase shift of 240 • was observed, shown in Fig. 11(c), while keeping RF power fluctuations within only ± 1.5 dB. Although good performance was demonstrated, achieving 360 • phase tunability, which is required for beam steering in PAAs, demands more than 50 dB of Brillouin gain, leading to excessively large optical pump power [75].
RF interferometric techniques can be harnessed to provide a full 360 • phase shift, without requiring high Brillouin pump powers; this approach uses RF vector addition to amplify the RF phase shift generated from an optical phase shift [123]. This scheme, summarised in Fig. 12(a), consists of two RF photocurrent contributions. The first contribution, RF 1 , is assigned a phase of 0. The second, RF 2 , has a relative phase shift of π and is used for interferometric phase amplification. The output of this system, RF net , is given by the vector summation of these two signals. When RF 1 and RF 2 are added, the overall phase shift, Θ 2 , is larger than the optical phase shift, Θ 1 , applied to RF 1 , at the expense of a reduced RF link gain [78]. The phase enhancement factor is defined as the ratio of the enhanced phase shift to the raw phase shift [75]. Fig. 12(b) shows the un-enhanced ±7.3 • phase shift over the RF frequency range 5-20 GHz generated from 2 dB of Brillouin gain using suspended silicon waveguides [75]. When the phase is amplified by the interferometric phase enhancement technique, a 360 • phase shift is achieved using an enhancement factor of 25, as shown in Fig. 12(c). Subsequently, this scheme was demonstrated using As 2 S 3 waveguides and a 360 • phase shift was generated up to 20 GHz using only 19 dBm of coupled on-chip pump power [76], compared to a total of 27 dBm on-chip pump power required to generate the 240 • phase shift [74] in a similar As 2 S 3 platform. The demonstrations of interferometric phase enhancement highlight the primary benefit of this interferometric technique, enabling 360 • RF phase shifts while using minimal amounts of Brillouin gain.

E. Brillouin True Time Delays
While RF phase shifters may be used to implement analog beamforming for narrowband signals, beam squint occurs as the signal bandwidth increases [124]. TTDs are therefore preferred for wideband signals as they eliminate beam squint. MWP systems generate TTDs by applying a linear dispersive phase slope over a frequency range that includes the optical carrier and modulated sideband. However, this requires a delay to be applied from the optical carrier to the highest frequency component on the RF sideband, which is greater than the bandwidth of most photonic components [125]. Alternatively, the separate carrier tuning (SCT) [126] method is well suited to generate TTDs for MWP applications and is implemented by applying a linear phase slope, such as SBS slow-light effects [102], [127], [128], [129], to an optical sideband and separately applying the appropriate phase shift to the optical carrier, as shown in Fig. 13(a).
The SCT method has been demonstrated using chip-based Brillouin devices by Aryanfar et al. [77], shown in Fig. 13(b). The TTD was verified using a two-tap filter, which is an unbalanced interferometer with one arm containing the TTD and the other maintaining constant delay. The change of free spectral range (FSR) of the interferometer in Fig. 13(c) corresponds to a  [77]. The phase shift applied to the optical carrier is indicated in Fig. 13(d) by the change of the null frequency in the interferometer pattern.
The delay generated by an MWP system can be enhanced through interferometric techniques [78] using a similar principle as the phase enhancement technique for MWP phase shifters. Liu et al. [90] demonstrated dispersive phase slope enhancement which increased Brillouin slow light delay, but lacked carrier phase tunability to meet the SCT condition for generating a TTD. McKay et al. demonstrated TTD enhancement using the SCT method [78], shown in Fig. 14(a). In this approach, two widely separated optical carriers, C 1 and C 2 , are used and both undergo single sideband modulation. However, the sidebands of each carrier are π out of phase relative to each other. Broadband Brillouin gain is applied to the sideband of C 1 , which causes a dispersive phase slope. A Si 3 N 4 MRR applies the appropriate phase shift to C 1 to implement SCT, and a second MRR applies the same phase shift to C 2 to ensure that C 1 and C 2 maintain a π phase offset. The overall phase shift is increased through the summation of the output photocurrents RF SBS and RF int , which originate from the two optical carriers. Fig. 14(b) shows the raw phase slope from the Brillouin resonance over a 1 GHz bandwidth, between 11 and 12 GHz. Fig. 14(c) shows the phase slope created when using enhancement factors of up to 30. The corresponding delay for each phase slope in Fig. 14(c) is shown in Fig. 14(d). We note that fluctuations in the phase and delay responses are amplified due to the phase enhancement scheme, and can in principle be reduced by finely tailoring the Brillouin gain spectrum to achieve the desired response [78]. Fig. 14(e) shows the two tap-filter response with and without the TTD applied. The change to the FSR corresponds to a delay of 330 ps [78], which is suitable for beam-steering applications for a 16x16 array operating in the K-band and above, as outlined in Section II.
If even larger delays are required for PAAs with more elements Brillouin-based light storage [130], quasi-light storage [131], and Brillouin dynamic grating based approaches [132], [133], [134] have been shown to achieve larger delay-bandwidth products compared to the Brillouin slow-light based approaches outlined above. However, so far those techniques have only been shown for optical signals and they still have to be demonstrated in an MWP link and only the former has been shown in an integrated platform [130]. Besides the large fractional delay achieved using Brillouin light storage it has been shown that the strict phase-matching condition between optical and acoustic waves in the Brillouin process can be used to induce non-reciprocal delays [135] which might open possibilities for novel processing methods in PAAs.
Alternatively, recent approaches utilise the polarisation dependence of Brillouin gain in both spun birefringent [136], and standard single mode [137] fibres to introduce narrow sub-MHz spectral dips within the Brillouin gain resonance. The spectral dip gives rise to a steep phase slope, which could form the basis of an MWP phase shifter or TTD with the potential to reduce the link gain penalty compared to the phase and delay enhancement techniques outlined above. Initial demonstrations of polarisation pulling using SBS in As 2 S 3 waveguides by Athanasios et al. [138] highlights possibilities of implementing such optical processing techniques on PICs in a compact form factor.

F. Brillouin Frequency Converters
RF frequency converters, often referred to as mixers, are used to convert between RF and a lower intermediate frequency (IF) [72], which is a critical function in PAAs. A mixer is a three port device, as shown in Fig. 15(a), where the frequency of the input RF signal is translated according to its difference in frequency from a second CW input tone, referred to as a local oscillator (LO). When downconverting a spectrum that contains a desired signal at frequency, f RF , and a signal equally spaced but on the other side of the LO at frequency, f IM , they are both converted to the same IF frequency, as shown in Fig. 15(b)-(c). The undesired signal in the IF spectrum is referred to as an image and causes distortion. The process of suppressing the image signal in the IF spectrum is referred to as image rejection [139], and is depicted in Fig. 15(d).  Fig. 16(a, d) from [72], © 2020, Optica Publishing Group. Fig. 16(e) from [73], © 2021, IEEE.) Brillouin-based MWP mixers inherently perform image rejection by ensuring that the output photocurrents destructively interfere in the absence of a Brillouin resonance. Zhu et al. demonstrated a chip-based SBS image-rejection mixer [72], which utilised a dual-parallel Mach-Zehnder modulator (DP-MZM) by placing the RF input on the top sub-MZM and the LO input on the lower sub-MZM, shown in Fig. 16(a). The DPMZM was biased to ensure destructive interference outside the Brillouin resonance, as shown in Fig. 16(b). Frequency conversion occurs when Brillouin gain is applied to the signal of interest on the upper sideband, and SBS loss is applied to the signal of interest on the lower sideband, as shown in Fig. 16(c). When SBS gain and loss are applied to the signals of interest, the photocurrent amplitudes are unequal and cause an output to form. An image rejection of 45.3 dB [72] was achieved using 15.1 dB of SBS gain and 13.3 dB of SBS loss, shown in Fig. 16(d). This system exhibited a conversion gain of −9 dB over the 3.2-13.2 GHz frequency range [72]. Although this demonstration provides low-loss conversion and deep image rejection, it only supported frequency conversion of narrowband signals. However, broadband image-rejection frequency conversion is required for compatibility with modern PAAs. McKay et al. demonstrated broadband image rejection downconversion using on-chip SBS by modulating the Brillouin pump over a broad bandwidth. In doing so, broadband image rejection of 28 dB and 16 dB was achieved over 100 MHz and 400 MHz, respectively [73], as shown in 16(e).

G. Brillouin Radio Frequency Sources
RF sources are essential components in the PAA, and when used in conjunction with an MWP frequency converter, enable PAAs with widely tunable central frequencies. RF sources with low phase noise are critical in wireless communication [140] and radar [71], [141] systems. However, RF frequency synthesizers typically generate an output that is up-converted from a stable, low frequency electronic source. The up-conversion process increases the phase noise by a factor 20 log 10 (N ) [70] when the frequency is multiplied by a factor of N . In practice, the phase noise is increased by additional noise from RF amplification stages [142]. Conversely, MWP-based RF sources directly generate outputs at radio frequencies by heterodyning two, or more optical tones. Brillouin laser cavities have been used to enable significant reductions to the linewidth, in addition to close-in frequency noise and RIN noise [67] of optical tones. Brillouin lasers have formed the basis of RF sources that achieve phase noise comparable to [70], or lower than the phase noise of high-performance RF microwave sources [143]. Additionally, photonic-based approaches have the distinct advantage of being able to generate RF tones over wide frequency ranges [71], [143].
Li et al. used Brillouin lasing in a high Q silica disk resonator [68] as the basis of a low phase noise RF source [70]. The main advantage of this scheme is that common mode jitter is reduced because the Brillouin laser lines share the same optical path [70], reducing the phase noise of the RF output. The operating principle of this device is shown in Fig. 17(a), where cascaded Stokes waves are generated in the silica disk. The first and third order tones co-propagate and are separated in frequency by twice the BFS, 21.7 GHz [70]. An RF signal with a frequency of 21.7 GHz is formed when both tones mix at a PD. Although the close-in phase noise (< 10 kHz) of this device is relatively large when free-running, it is reduced using closed-loop control to form a phase-locked loop, as shown in Fig. 17(b). The feedback loop allows for the use of a stable low-frequency quartz oscillator as a frequency reference [70], which improves the output phase noise, shown in Fig. 17(c). A low white-phase-noise floor of −160 dBc/Hz was obtained [70]. When the output frequency is divided, the phase noise improves, as shown in Fig. 17(d). We note that high Q Si 3 N 4 MRRs, demonstrated by Gundavarapu et al., have been used to form an RF source through cascaded Brillouin generation. This approach offers narrow Stokes linewidths of only 0.7 Hz [67], and is an attractive method to generate low phase noise RF signals. It is desirable to generate microwaves with continuous frequency tunability for use in PAAs, however, the frequency of the RF output is limited to discrete frequencies determined by the BFS [67], [70].
Li et al. extended on the work presented in [70] by demonstrating a continuously tunable RF source using two lasers locked to distinct modes of a common Brillouin laser cavity [143]. A reduction in phase noise was achieved through optical frequency division.
Optoelectronic oscillators (OEOs) [144] are an alternate method to generate low phase noise RF signals. However, most OEOs rely on a narrowband filter to select one cavity mode, which are not widely tunable if implemented using traditional electronic methods [71]. On-chip SBS enables widely tunable filtering functionality, and has been used as a widely tunable MWP filter to select the desired OEO loop mode and enable wideband RF signal generation, as shown in Fig. 18(a) [71]. In this approach, Brillouin gain amplifies one cavity mode to create phase-intensity modulation conversion and sustain an oscillation. Using this technique, an RF tone was synthesised with frequency up to 40 GHz, as shown in Fig. 18(b). The advantage of optically generating microwave tones over a wide frequency range is highlighted in Fig. 18(c), where signals at 10.9 GHz and 40 GHz have similar phase noise.

H. Non-Reciprocal Devices
Non-reciprocal components are key to on-chip photonic processors, as they protect optical components from back-reflected light [145], which is particularly important for backward Brillouin photonic devices in which counterpropagating pump and probe signals are required [54]. Optical devices traditionally achieve non-reciprocity using magneto-optical materials which induce a non-reciprocal Faraday rotation [146]. While chip-scale isolators and circulators using magneto-optic garnets [147] have demonstrated large 30 dB isolation over a wide 16 nm 10-dB bandwidth [148], these devices are challenging to integrate into PICs [149].
Non-reciprocity requires the breaking of time reversal symmetry, which can be implemented using SBS [150]. Acoustooptic non-reciprocal devices rely on phase matched optical mode conversions that occur for optical waves travelling in one direction but not the other [150], [151]. Chip-based demonstrations of optical non-reciprocity have utilised resonant structures [149], [152], which limited the usable bandwidth, or required suspended waveguides to increase acoustic guidance [95], [153], which increased fabrication complexity. On-chip non-reciprocity has been achieved without the use of resonant structures [38], [154]. For example, Kittlaus et al. recently demonstrated an electrically induced acoustic wave using standard silicon waveguides and on-chip mode multiplexers to facilitate non-reciprocal intermodal acousto-optic scattering [155]. This device achieved over 16 dB of optical isolation over a 100 GHz bandwidth without the use of an optical cavity [155].
An alternate approach introduced by Liu et al. [156] avoids the requirement for a non-reciprocal element by exploiting backward inter-modal Brillouin scattering (BIBS) in combination with passive mode-selective filters. However, less than 1 dB of BIBS gain was observed [156]. Therefore, further developments to this platform are required to increase the magnitude of BIBS gain generated to make this platform suitable for MWP processing.

V. OUTLOOK FOR INTEGRATED BRILLOUIN PHASED ARRAY ANTENNA SYSTEMS
The long-term goal of MWP is a fully-integrated processor with enhanced functionality that supports signals with wide bandwidths over wide frequency ranges. This system must also meet the demanding performance specifications of real-world RF systems in terms of noise figure, linearity and insertion loss. This section outlines technological advances that are required to realise a high-performance and fully-integrated Brillouin-based MWP processor that incorporates multiple passive and active photonic components to enable key PAA processing functionalities, including beamforming elements based on Brillouin TTD, and image rejection frequency converters to interface between the RF operating frequency and the baseband.

A. Architecture of Brillouin-Based Phased Array
A potential architecture of a PAA using fully-integrated MWP processors is shown in Fig. 19(a). To simplify the discussion without loss of generality, we discuss the PAA operating in receive mode. The proposed architecture consists of a Brillouin beamforming network with a PIC at each row of the PAA, where integrated Brillouin processors implement TTD beamforming at each antenna element. Chip-based pump and probe lasers are routed to each integrated Brillouin processor and the spectrum of each pump is broadened by a DAC which drives MZM1. The signal from each integrated Brillouin processor is combined and converted to the baseband using a Brillouin-based image rejection mixer driven by a low-noise Brillouin-based oscillator. This ensures signals outside of the band interest are attenuated through destructive interference [72], [73], essentially acting as a MWP filter cascaded with a frequency converter.
The core components of the integrated Brillouin processor are highlighted in Fig. 19(b), which works as follows; an input RF signal is modulated onto the optical carrier by MZM2 at (i), a modulation transformer [157], [158] generates the required modulation format at (ii), on-chip optical amplifiers set the pump and probe power as required before the probe undergoes Brillouin processing at (iii), an optical notch filter suppresses reflections of the Brillouin pump at (iv). Reflections are primarily caused by index changes between different materials and can be greatly reduced in an appropriately designed integrated platform, compared to systems that rely on a combination of chip-scale and fibre components [94]. Finally, the signal is converted back to the RF domain. This architecture can be extended to enable phase or delay amplification, as outlined in Section IV, by adding additional core components outlined in Fig. 19(b).

B. Integrated Platforms and RF Link Performance
Integration of multiple materials is key to realising the integrated Brillouin processor in Fig. 19, as no single material platform provides all the necessary functionalities [159]. The integrated Brillouin processor requires integration of Brillouin waveguides, E-O modulators, optical amplifiers, circulators, optical filters, and PDs. A number of hybrid and heterogeneously integrated platforms exist that combine many of these components. For example, butt-coupled TriPleX Si 3 N 4 and InP chips [26] combine the advantages of low-loss Si 3 N 4 waveguides with optical sources, amplifiers, modulators, and PDs offered by InP, and have formed the basis of fully-integrated MWP processing devices [32]. Initial studies have been undertaken to optimise the magnitude of Brillouin gain generated in Si 3 N 4 waveguides [160], [161], however, the Brillouin gain coefficient in Si 3 N 4 is orders of magnitude lower than As 2 S 3 and SOI platforms. Although it is highly desirable to generate Brillouin gain in Si 3 N 4 waveguides due to their widespread use and lack of non-linear losses, As 2 S 3 and SOI platforms are currently more suitable for the SBS medium, as they have demonstrated much higher magnitudes of Brillouin gain. SOI platforms leverage advances in CMOS fabrication methods, offering on-chip E-O modulators and PDs [34], [157]. Fully-integrated MWP processors have been achieved by pairing SOI chips with free-space coupled laser diodes [162]. Although these devices offer compact form factors, silicon lacks intrinsic second order optical nonlinearity, and is therefore doped to enable E-O modulation [163]. This introduces a trade-off between device bandwidth, losses, and modulation efficiency, especially at high RF frequencies [164].
Thin-film LiNbO 3 modulators are positioned to greatly improve the link performance of integrated MWP processors as they offer record-low V π RF [164], [165], [166]. The integration of etched LiNbO 3 modulators with InP lasers [167] and micro transfer printed [168] amplifiers [169] offer an attractive path forward for integrated platforms used in high-performance MWP processors.
Care must be taken to minimise reflections of the Brillouin pump, which are caused by refractive index changes between materials. As discussed in Section IV, reflections must be minimised in systems that use backwards Brillouin processing to avoid degradation to the RF link performance [93]. The vertically tapered interface between Ge:SiO 2 and As 2 S 3 that Lai et al. demonstrated [94] is an attractive method to achieve lowloss and low reflection interfaces in heterogeneously integrated MWP systems. Demonstrations of similar low-reflection transitions in future integrated Brillouin MWP processing platforms are essential to improve device performance.
Brillouin microwave processors require integration of circulators with isolation in the order of 30 dB to separate the Brillouin pump and probe. As discussed in Section IV, magneto-optical materials are difficult to implement in chip-scale devices, which leads to fabrication complexities [147]. Hence, acousto-optic non-reciprocal devices based on the same materials used in the MWP processor are an attractive alternative. Demonstrations of acousto-optic non-reciprocal devices using silicon [155], Si 3 N 4 [149], and LiNbO 3 [170] waveguides highlight that non-reciprocal elements can be paired with many existing integrated MWP processing platforms. However, the efficiency and isolation exhibited by acousto-optic non-reciprocal devices must be improved to meet the requirements for Brillouin-based processors. For example, this can be achieved by increasing the efficiency of acoustic wave transducers [155], or reducing acoustic losses [170].
For MWP processors to be adopted in RF systems, their noise figure, linearity, and insertion loss must be competitive with their traditional RF electronic counterparts [17]. Generally, it is desirable to achieve low-loss transmission with noise figure below 10 dB and high linearity with SFDR exceeding 120 dB.Hz 2/3 [17], [79]. However, most integrated microwave photonic systems do not meet these metrics, or do not report them. Recent demonstrations of MWP filters with optimised link performance have come close to the target metrics, using a combination of fibre-based and chip-based components [45], [46], [158]. For example, Daulay et al. reported an MWP filter using on-chip processing that achieved a link gain of 10 dB, and noise figure of 15 dB, and SFDR of 116 dB.Hz 2/3 [158]. This performance was enabled using low-biasing, high power lasers and optical amplifiers [158]. When the same MWP filter was optimised to maximise linearity, an SFDR of 123 dB.Hz 2/3 was observed [158]. Although the noise figure and linearity were not all optimised simultaneously, this device recorded highest linearity, and lowest noise figure in an MWP link that uses on-chip processing to-date [158]. However, this level of performance is yet to be demonstrated in fully-integrated MWP processors. To bridge this performance gap, low loss waveguides [171], [172], [173], low V π RF modulators [165], on-chip lasers with high power and low RIN noise [174], [175], low noise optical amplifiers [175], [176], and high-power PDs [168], [177], [178] must be integrated into a single platform.
To reduce the loss associated with converting between RF and optical domains multiple times, it is desirable to implement all the required functionalities in a single MWP link. Although many MWP processors have been used to demonstrate multiple RF processing functionalities [48], [158], [179], they only demonstrate a single RF processing function at any given time. This motivates future work to implement MWP processors that simultaneously achieve multiple RF processing functionalities in one optimised MWP link. Reza et al. performed a theoretical analysis of a fully-integrated MWP beamformer that combines frequency conversion and phase shift beamforming [44]. Through simulations they showed that the overall RF link gain of the multi-functional MWP device was comparable to single function MWP devices, essentially halving the link gain penalty by distributing it between two RF processing functionalities. This highlights the advantages of combining multiple RF processing functionalities into a single MWP link.

C. Size and Power Considerations
It is desirable to reduce the pump power required to generate SBS gain in Brillouin waveguides. Although RF interferometric processing techniques reduce the pump power required by Brillouin-based MWP processors [73], [93], [115], relatively large pump powers are still required to generate Brillouin gain over a wide bandwidth. For example, approximately 350 mW of coupled pump power was used by McKay et al. to generate 4 dB of Brillouin gain over a 400 MHz bandwidth [73] in a Ge:SiO 2 -As 2 S 3 PIC [94]. By analysing (13), the magnitude of Brillouin gain generated for a fixed pump power is related to the figure of merit (FOM), G SBS L ef f , and determines how effective a material platform is at generating Brillouin gain [63]. Increases to both the Brillouin gain coefficient and the effective waveguide length are needed to reduce the required pump power. The former is improved by developing new Brillouin processing platforms and materials, such as fully-etched As 2 S 3 waveguides [63], which have larger Brillouin gain coefficients than traditional As 2 S 3 rib waveguides. The latter is increased by reducing waveguide losses of the Brillouin medium, which is currently dominated by sidewall scattering [180].
The form factor of each component must be considered to ensure the device is contained in the half-wavelength spacing between PAA elements, which is only 3 mm at 50 GHz. Many components in the design presented in Fig. 19 are sufficiently compact and introduce no significant form factor limitations. These include MRRs, which have a bend radius of approximately 100 μm [171], [181], and PDs, which typically have maximum linear dimensions of only tens of microns [168], [177], [178]. Optical gain sections used for SOAs and laser sources are slightly longer with lengths close to 1 mm [168], [182]. While Brillouin As 2 S 3 waveguides can require interaction lengths exceeding 6 cm [63], they have been incorporated in compact spirals with dimensions of less than 1 mm 2 [63], as discussed in Section IV. On-chip PPER waveguides typically require shorter interaction lengths, with a recent demonstration using Brillouin waveguides with only 17 mm length [110], which can easily be integrated into a footprint less than 1 mm 2 .
Most work on high-performance modulators to-date has focused on lowering the V π RF , however, considerations must be made to ensure these devices maintain compact linear dimensions for deployment in space-constrained applications, such as PAAs. For example, the lowest V π RF of an E-O modulator is 2.12 V at 50 GHz [165] and has a length of 2 cm, which can be challenging to integrate at each element of a PAA where space is highly constrained. However, folded [183], [184], [185], or spiral [186] E-O modulators have smaller maximum linear dimension when compared to their straight counterparts, and are attractive solutions for space constrained applications, such as at every element of a PAA. Similarly, acousto-optic non-reciprocal devices have been folded to reduce their on-chip footprint, as demonstrated by Kittlaus et al. [155].

VI. CONCLUSION
In conclusion, we reviewed recent advances to microwave photonic processors that use on-chip SBS to enable key RF processing functionalities for possible implementation in future PAAs, including filters, phase shifters, time delays, mixers, and frequency sources. The performance and functionality of these Brillouin-based devices has been enhanced by recent demonstrations of novel platforms that facilitate high-performance integrated Brillouin microwave processing. These integrated Brillouin platforms, along with recent advances in on-chip photonic components, stand to greatly reduce the form factor of microwave photonic processors when compared to current RF solutions with equivalent frequency operating range, bringing these devices closer towards real-world applications. We discussed challenges that must be overcome to implement fully-integrated SBS microwave processors for element-level beamforming in PAAs, whilst meeting demanding RF link performance requirements. In doing so, we have highlighted how on-chip SBS provides a path to enable compact microwave processors with wideband capability unmatched by traditional RF electronic solutions.

ACKNOWLEDGMENT
We acknowledge the ARC Grants DP190100992, DP200101893, and DP220101431 and fruitful on-going collaboration with our collaborators, Prof Michael Steel, Prof Christopher Poulton, Prof Steve Madden and Dr Bill Corcoran. We would like to thank Dr Stu Milner, Keith Kelly, Dr James Coyte, Nicholas Athanasios, and Ziqian Zhang for valuable feedback during the preparation of the manuscript. We would also like to thank Dr Yang Liu and Dr Alvaro Casas-Bedoya for their contributions to the design of the photonic chip shown in Figs. 1 and 6.