Even though modern microelectronic technology has gained great achievements in high-speed information processing up to several Gbit/s order, there is still an increasingly tremendous demand for even higher speeds of more than several Tbit/s order along with the rapid development of contemporary information technology, including big data and cloud computing. An ultrahigh information processing speed of over several Tbit/s order could be expected when using photon as information carriers. Accordingly, ultrawide-band and ultrahigh-speed optical information processing chips have been one of the research fronts and focus in the overlapped fields of nanophotonics, materials, and chemistry. Nanoscale all-optical logic devices are essential and key units of optical computing systems, and ultrawide-band and ultrahigh-speed optical information processing chips. Consequently, for decades, nanoscale all-optical logic devices have been a very active and important research direction. Great effort has been put into the demonstration and experimental realization of all-optical logic functions. Nanoscale all-optical logic devices can be considered as the counterpart of photonics analogous to electronic logic devices used in the central processing unit of electric computers. Nanoscale all-optical logic devices can be divided into two categories: the first one is a simple logic gate (AND gate, OR gate, NOT gate, and XOR gate, NXOR gate, etc), and the second one is a unit logic device (adder, subtracter, multiplier, encoder, decoder, comparator, discriminator, trigger, shifter, counter, etc). The logic operation using the photon signal is a very challenging frontier research because of the fundamental requirement of very efficient light-control-light. As early as 1983, Lattes et al realized the XOR and AND gates using a LiNbo₃ Mach-Zehnder interferometer based on the third-order nonlinear optical effect, with an operating light power up to 2 W [62]. Subsequently, various schemes have been proposed to demonstrate all-optical logic devices, including using nonlinear optical crystals, optical fibers, semiconductor optical amplifiers, semiconductor nanowires, photonic crystals, silicon ring resonators, and plasmonic microstructures (shown in figure 4) [63–65]. The all-optical logic devices realized using microcavities, semiconductor optical amplifiers, nonlinear optical crystals, and optical fibers have a relatively large size of over 100 μm order, which is not suitable for practical on-chip integration applications. The all-optical logic devices realized using photonic crystals, silicon ring resonators and plasmonic microstructures have a relatively small feature size. However, up to now, the experimental reports of nanoscale all-optical logic devices are very limited, only including logic gates, adders, and so on. Moreover, no complex all-optical logic function modules, possessing the ability of performing various complicated all-optical logic operations (for example, multiplication and division arithmetic logic operation, solving equations and other mathematical operations), have been realized.

The nonlinear effects in optical waveguides play a vital role in optical signal processing towards switching, multiplexing, compression and logic functions [82]. With the enhanced scope of future possibilities in optical communication due to nonlinear effects, the use of nonlinear waveguides, like resonant cavities, are also increasing due to various advantages such as low power requirement and ultrafast switching. In signal processing operations, resonators provide an efficient way to achieve desirable transfer functions [83], logic [84] and time delay element [85] in a circuit. Along with this, optical resonant structures have the benefit of cascadability which can help in complex arrangements of circuits. Generally, the dielectric resonators of small size are preferred for integrated optics while the high index contrast micro-resonators are used for microwave applications. It is obvious that at optical frequencies, due to non-availability of good conductors, we have to keep an eye on the radiation losses while designing the layout of the structures. The structures with high index contrast enable the users to achieve a higher radiation quality factor in comparison to the overall quality factor of the system. The worries about the quality factor generally having less significance with the high bandwidth therefore resonant structure came into its own groove once high bit rates were possible. Along with this, there is a challenge associated with high bandwidth, in order to have a low quality factor with high bandwidth acute coupling between the resonator and external waveguide is essential and it is generally achieved by evanescent coupling. Coupled optical resonators can be the foundation of wavelength filters with flat-top responses and these are largely required in telecom applications. In general, there exist two kinds of resonators; first are ring/rack type resonators (shown in figure 8) that support degenerate modes of traveling waves in opposite directions. When index contrast is on the higher side, the radius of rings can be formed small enough to have lesser radiation losses and higher free spectral range (FSR). The second are Bragg reflection resonators that have standing wave modes.

Signal regeneration is primarily aimed at improving the quality of a transmitted signal. Regenerators are designed to increase system performance, reduce data degradation, extend system reach, and enhance the signal-to-noise (SNR) for higher link capacity. In general, regenerators perform three signal-processing regeneration functions (i.e. ‘3R’): (1) reamplifying, (2) reshaping, and (3) retiming the signal [92]. Conventionally, signal regeneration in an optical system is performed though optical-electrical-optical (OEO) conversion, in which a weak and distorted signal is detected, restored in electronics, and retransmitted onto an optical fiber.

Recently, there has been growing interest in fully regenerating optical signals in the optical domain. There are many potential advantages of all-optically regenerating the signal such as: high efficiency by avoiding OEO conversion, fast processing, large bandwidth, and ability to operate on the signal phase. Although ‘1R’ optical regenerators (or amplifiers, such as erbium-doped fiber amplifiers, EDFAs) have been deployed worldwide, they allow the data to become distorted in shape and time. Such distortions are much more complicated to regenerate all optically.

Advances in optical communication technologies and growth in demand for higher capacities have opened up new opportunities for all-optical regenerators. For example, the technology of optical transceivers has developed to simultaneously tailor the optical wave in multiple dimensions: amplitude, phase, time, frequency, polarization, and space. Dual-polarization and higher-order, phase-and-amplitude-based modulation formats for wavelength-division-multiplexed (WDM) channels are currently being adopted in many fiber systems [93], and space-division-multiplexing (SDM) is of great current interest [94]. Such channels impose additional transmission challenges, requiring higher SNR and lower data degradation.

Optimally, optical regenerators should operate on multiple dimensions of modulated signals, as shown in figure 9. Recent reports have investigated using optics for regenerating the following signals: (i) multiple amplitude levels [95], (ii) multiple phase levels [96], (iii) WDM channels with binary data [97], (iv) channels degraded by polarization-mode dispersion [98], (v) WDM channels with crosstalk [99], and (vi) SDM channels [100]. These demonstrations were enabled by various novel linear and nonlinear optical devices. For example, all-optical phase regeneration can be accomplished by coherently adding the signal to the conjugate of its higher harmonics using four-wave mixing to create a staircase phase function, as shown in figure 10.

Fiber-optic communication systems form the backbone of the internet. The data rates and information communication services that we enjoy today are enabled by breakthrough technologies developed over the last 40 years. The field started with the low-loss silica fiber in the late 70s; the advent of the erbium-doped fiber amplifier (EDFA) led to wavelength division multiplexing (WDM) in the 90s; and culminated in the high spectral efficiencies enabled by advanced coherent modulation and digital signal processing (DSP) and powerful coding we see today. Lightwave communications have matured to a point where commercial WDM systems can routinely transmit digital data at multi-terabit per second rates.

Notwithstanding, the spontaneous emission noise in the EDFA and the ultrafast nonlinear Kerr response of standard single-mode silica fibers place fundamental limits on the maximum data rate that can be transmitted over a fixed link, leading to concerns that the fiber capacity will soon be exhausted without further technological innovation [93]. This situation has spurred a renewed interest in fiber technologies [104] and nonlinear optical signal processing functionalities, such as parametric phase-sensitive amplification [105], optical regeneration [106] and nonlinearity compensation [107]. Arguably, the role of optical signal processing in future telecom applications is to alleviate the distortions experienced by the optical signals upon propagation in the fiber link and facilitate functionalities that would be very difficult, if not impossible, to realize by electronic DSP alone. Here, we restrict the scope to nonlinear signal processing based on the Kerr effect. Paradoxically, this very same phenomenon can be harvested to our advantage in a material platform to combat its detrimental effects in the silica fiber.

Illustrating the challenges with Kerr nonlinear waveguides: the PSA example

State of the art in Kerr nonlinear waveguides

Ideally, an integrated waveguide platform should give a nonlinear phase shift of γ  L_eff  P_cw ~ π, where γ represents the nonlinear Kerr parameter (in units of [Wm]⁻¹), L_eff is the effective length of the waveguide and P_cw the continuous-wave power. In the case of a PSA, this would roughly yield an on-chip parametric gain close to 10 dB. Since P_cw is somewhat limited to 0.1 W for state of the art on chip laser sources, the relevant figure of merit becomes the amount of nonlinear phase shift per unit power, resulting in the target γ  L_eff ~ 10 π [W]⁻¹. This discussion obviates the fact that many of these platforms display nonlinear absorption effects (such as two and three-photon absorption, free carrier effects, etc) which place an even lower limit to the amount of P_cw that can be safely delivered to the waveguide.

Table 1 illustrates the challenge. It summarizes the nonlinear and linear performance of diverse state-of-the-art nonlinear waveguide platforms. The message is clear: the waveguide platforms that feature higher nonlinear Kerr parameters display a higher amount of linear and nonlinear loss. These are all extremely interesting materials but further work is needed to achieve the ultimate goal of an integrated nonlinear waveguide platform with negligible nonlinear loss in the telecom band and sufficient nonlinearity and low loss.

Table 1.  Overview of recently developed nonlinear waveguide platforms [103, 108–111] with a large nonlinear Kerr parameter, γ; u.i. means ‘under investigation’.

Platform	γ (W.m)−1	Linear loss (dB m−1)	γ  Leff(W)−1	Nonlinear loss?
SiN	~1	~1	~1	no
Silicon rich SiN	~500	~450	~5	u.i.
Si with pin	280	~100	~10	yes
AlGaAs	660	140	~20	u.i.
Amorphous Si:H	2975	~320	~40	yes

Silicon nitride is a dielectric material with a transparency window that ranges from the visible to the mid IR. In its stoichiometric form, it has nonlinear Kerr coefficient that is about ten times higher than silica. High-confinement silicon nitride resonators display an equivalent loss down to the 1 dB m⁻¹ [108] but the material absorption itself is about an order of magnitude lower. This material platform is superb for nonlinear optics in the near infrared because of its absence of nonlinear loss. The path forward towards on-chip parametric gain with continuous-wave pumping in this platform requires further understanding of the origin of the absorption losses and the development of advanced etching techniques to be able to realize ultra-smooth sidewalls in very long (>1 m) waveguides. The device will naturally occupy a bigger area. Interestingly enough, the nonlinear Kerr coefficient can be precisely controlled with the film stoichiometry. In particular, silicon rich silicon nitride (SiRN) [109, 113] provides a dramatic enhancement of the Kerr effect, but this is accompanied by a decrease in the optical bandgap and an increase in the linear loss. Although the bandgap is large enough to avoid two-photon absorption (TPA) in the telecom band (see figure 11), it is still unclear the extent to which three-photon absorption (3PA) and associated carrier dynamics may affect the nonlinear performance in practical applications. Furthermore, the origin of the absorption losses is not well understood either.

Silicon rib waveguides can be engineered with p-i-n diodes to remove photogenerated carriers and decrease free carrier absorption losses. Recent demonstrations based on rib waveguides have allowed for efficient wavelength conversion and regeneration [110]. Nonlinear losses due to two-photon absorption remain, although the key challenge in this geometry is the control of dispersion.

Recent advances in AlGaAs bonded on an oxidized silicon wafer have allowed to achieve strong optical field confinement in nanowire waveguides, leading to huge nonlinear Kerr parameters [111]. The atomic composition of this semiconductor can be controlled with great accuracy, allowing for precise control of the optical bandgap. Recent results indicate absence of TPA [111], but similar to the case of SiRN, the effects of 3PA are not well understood yet.

Another interesting development is with hydrogenated amorphous silicon [103]. This material platform has an increased nonlinear parameter and optical bandgap than its crystalline counterpart. It still displays nonlinear losses, but the most intriguing aspect is the reported time instability (samples degrade over time), whose physical origin is also not well understood.

Other relevant material platforms include tantalum pentoxide, chalcogenides, plasmonic focusing, polymers and diverse semiconductor materials such as gallium phosphide. The tradeoffs are similar to the ones observed for the material platforms in table 1 however, namely higher nonlinear coefficients are obtained for platforms that display an increase in both linear and nonlinear losses.

From a more practical perspective, it is important to remember that these waveguide structures will require co-integration with pump lasers and interfacing with low-loss nanophotonic circuits and possibly electronics. Other potential concerns such as damage power thresholds and the nonlinear response in complex integrated circuits deserve more attention.

Nonlinear signal processing can provide significant gain benefits in telecom applications. For practical deployment, it would be necessary to realize an integrated nonlinear platform that allows to achieve a nonlinear phase shift per unit power close to 10π and has no sign of nonlinear loss. Unfortunately, most material platforms either display strong linear and nonlinear loss or insufficient Kerr parameter. If this breakthrough were achieved, it would have far reaching consequences not only in fiber optical communications, but also in quantum optics, spectroscopy and metrology.

Perhaps the wonder nonlinear material is just around the corner or perhaps we should focus on exploring different waveguide structures. For example, an interesting alternative that has not been carefully investigated in the literature to our knowledge is the possibility of a hybrid nonlinear waveguide that provides optical amplification around the pump window to compensate for the detrimental effects of the linear loss. On-chip net gain in glass waveguides doped with rare-earth elements have been reported, but little is known about their nonlinear Kerr coefficients.

Despite the excellent transmission properties of silica glass, there has long been a desire to fabricate fibres from new materials that can offer enhanced optical processing capabilities and/or extended transmission windows [114]. Although most efforts in this area have focused on non-silica glasses, such as the heavy metal fluorides and chalcogenides, over the past decade, a new class of fibre with silicon core materials has appeared [115]. These fibres are still clad in silica so that they are robust, stable, and can be handled in a similar way to the pure silica structures. However, the large nonlinear coefficients and tight mode confinement offered by the high index silicon core (as discussed in section 14) presents exciting opportunities for the development of compact and efficient all-optical processing devices that could be seamlessly integrated within existing fibre communications networks.

The first example of a silicon core fibre was produced in 2006 by a modified chemical vapour deposition (CVD) technique. In this approach, the silicon materials were incorporated inside pre-existing silica templates, such as capillaries and microstructured optical fibres. However, only two years later, a second technique was developed that allowed for the large scale production of silicon fibres via a conventional drawing tower approach, significantly enhancing their practicability. This method involves sleeving the silicon material inside a silica cladding to form a preform, before drawing it into a fibre with micron-sized dimensions. Importantly, these two fabrication techniques are complementary in terms of the materials and geometrical structures that can be accessed. For example, the CVD approach has been used to produce hydrogenated amorphous silicon (a-Si:H) fibres with the smallest core dimensions (a few microns or less) over centimetre lengths, while the drawing method has produced large core (tens of microns or larger) crystalline silicon fibres over hundreds of metres.

From a nonlinear perspective, a-Si:H has a higher Kerr nonlinear coefficient than crystalline silicon, usually reported to be around five times larger. Thus, when combined with the smaller core diameters, as of to date, the a-Si:H fibres have been the most suitable for nonlinear device demonstration. Figure 12 shows examples of two basic all-optical processing functions, (a) modulation and (b) wavelength conversion, that have been used to benchmark the performance of these fibres [115]. In both cases, the a-Si:H fibres had a core diameter of ~6 μm, so that they exhibited strong normal dispersion in the telecoms band, and a length of ~1 cm. The first demonstration made use of a cross-absorption modulation (XAM) process, whereby a high power pump pulse was used to imprint a dark pulse on a CW signal via two-photon absorption (TPA). The second scheme was based on cross-phase modulation (XPM), whereby the pump was used to impart a wavelength shift (blue then red shift as indicated by the dashed white line) on a weak probe. In both cases, the timescales for the modulation were on a sub-picosecond scale, with an extinction ratio greater than 4 dB (as high as 12 dB in the case of the XPM red shift). Although the nonlinear performance of these fibres was shown to be comparable to what has been achieved in nanoscale planar analogues, the powers required to observe these effects were higher owing to the larger cores (hundreds of watts compared to milliwatts). However, a simple way to reduce the power requirements is to use these structures in a resonator geometry to enhance the light–matter interactions, similar to what has been achieved in pure silica fibres [116]. By coupling light into the silicon core such that it is confined to circulate in a whispering gallery mode configuration, it has been shown that the power threshold for all-optical modulation can be reduced by four orders of magnitude [117].

Although much of the early nonlinear optical characterization of these fibres was focused on the telecoms band, owing to the strong nonlinear absorption of silicon in this region, interest soon shifted to longer pump wavelengths. As well as mitigating the effects of TPA, moving the pump into the mid-infrared region (i.e. λ > 2 μm) has the added advantage in that it is possible to access the anomalous dispersion regime, even in a micron-sized core. Figure 12(c) shows two examples of continuum generation obtained in a 2 μm diameter core a-Si:H fibre with a length of 4 mm; one pumped in the normal dispersion regime (λ_p ~ 1.8 μm blue curve) and the other in the anomalous regime (λ_p ~ 2.5 μm red curve) [118]. Note, the black dashed spectra are illustrative of the input pump. Clearly, the latter process generates the broadest continuum, octave-spanning over 1.6–3.4 μm, but at the price of reduced flatness and coherence. Nevertheless, these results pave the way for silicon core fibres to find application in high-speed all-optical regeneration and spectral slicing for wavelength division multiplexing within all-fibre integrated systems.

As with any new device technology, there are a number of challenges that must be overcome in order to improve the applicability of the silicon core fibres. For nonlinear processing, perhaps the most important of these is the reduction of the transmission losses. Figure 13 highlights the progress that has been made over the past decade to enhance the transmission properties in both the amorphous and crystalline materials. In terms of the a-Si:H material, improvements must be achieved through the complex deposition process, which is somewhat limiting and hence why progress is currently stalled around the 1 dB cm⁻¹ mark. In contrast, many of the advancements that have been made within the crystalline core materials can be attributed to novel post-processing techniques that have only recently been developed, and thus work in this area is still very much ongoing [119].

Silicon photonics has been extremely successful as a platform for chip-scale optics, with the technology now widely available through a number of international foundries. The key benefits of the platform include its low propagation losses at telecommunications wavelengths, high confinement of the optical field to sub-micron waveguide cross-sections, compact device footprints, and compatibility with electronic manufacturing processes. These properties have enabled the realization of a wide range of compact, on-chip linear optical components including wavelength selective filters, power couplers, interferometers and delay lines. In turn, these simple building block components have been used to construct large scale photonic integrated circuits (PICs) to perform optical signal processing functions such as temporal integration and differentiation [124]. Direct integration with electronics has allowed active feedback and control of optical signals [125], while hybrid materials integration has enabled functionality beyond that of silicon alone, for example, efficient on-chip light generation [126].

Further functionality has been enabled by the native χ⁽³⁾ non-linearity of silicon, with demonstrations of all-optical signal processing operations such as signal broadcasting, regeneration and switching [127]. Furthermore, degenerate spontaneous four-wave mixing in micro-resonators is now commonly used as an on-chip source of correlated or entangled photons for quantum optics systems [128].

Although a wide range of functions have been demonstrated on the silicon photonics platform, there remain a number of key challenges in the development of systems for all-optical processing.

1. Tuning and trimming

2. Efficient χ⁽³⁾ processes

3. Access to χ⁽²⁾effects

4. Rejection of on-chip pump light

5. Manufacturing at scale

Electrically addressable tuning of photonic devices has been enabled using both thermal and p-i-n junction technologies in silicon. These schemes are further enhanced by feedback control [125], especially where direct integration with silicon electronic drivers is possible [129]. A crucial component of such feedback systems is the ability to monitor optical power at distributed points within the PIC using compact, low-loss detectors. Stable operation of non-linear resonant devices can also be achieved by self-locking of non-linear cavities using an external resonator with gain [130], using the broadband gain of the outer loop to allow the lasing peak to follow the non-linear resonator line.

On-chip pump light filtering is commonly realized using filters based on multiple Bragg grating and/or ring resonator devices, and has been demonstrated with high on-chip rejection ratios [128], though these are currently limited by the light scattered into the silicon substrate and the strong scattering between TE and TM waveguide modes [131]. The need for such high rejection filters is driven by the low efficiency of spontaneous FWM process in silicon and can be mitigated by realizing more efficient non-linear devices on-chip.

Some of the limitations of the silicon material platform can be addressed by making use of advanced micro-fabrication techniques. For example, by introducing strain in silicon waveguides, an effective χ⁽²⁾ non-linearity can be induced [36]. Additionally, silicon waveguide technology can be interfaced on-chip with CMOS compatible materials to access performance not possible using silicon alone. Demonstrations of efficient non-linear processes have been made using materials including SiO_xN_y [131], amorphous Si and organics. Not only are the propagation losses in these materials low, they do not exhibit significant TPA at telecommunications wavelengths, making them an attractive, and easily integrated, addition to silicon photonic PICs.

The major opportunity for augmenting the capabilities of the silicon photonic platform however, lies in the heterogeneous integration of materials and devices with properties complementary to those of silicon. To satisfy the requirements of volume manufacture of silicon based PICs, the integration of such materials must either be compatible with the foundry processes, or be achievable as a fully back-end integration method. Recent work focused on the realization of on-chip light sources on silicon has led to major breakthroughs in III-V growth on silicon and wafer bonding to pre-fabricated PICs. The importance of this area has been highlighted by the instigation of the IEEE heterogeneous integration roadmap, and has clear implications well beyond the development of on-chip light sources and detectors. Recent work in thin film lithium niobate, for example, has seen demonstrations of χ⁽²⁾ modulators on silicon fabricated by bonding small pieces of material to pre-fabricated waveguides [132]; see figure 14.

In the last decade or so, it has been recognized that several single-mode fibre -based technologies are facing impending fundamental limitations in many application areas, including telecommunications, fibre lasers and amplifiers, nonlinear signal processing and sensing. For example, in modern coherent telecommunication systems, this is a consequence of the tremendous data traffic growth in all types of networks that are now said to approach their maximum capacity limit per fibre. Consequently, radically new types of fibres that can support higher capacities using multiple spatial paths of either a multi-mode fibre or a multi-core, single- or multi-mode fibre are being actively pursued. This approach to realizing potentially more cost-effective network capacity scaling, through the improved device integration and interconnectivity opportunities, is generically referred to as space-division multiplexing (SDM) [137]. Similarly, the power scaling of fibre lasers and amplifiers has necessitated the use of large mode area fibres, entailing either the management or the exploitation of their higher order modes (HOMs).

Semiconductor optical amplifiers (SOAs) have been largely used for high speed all-optical processing from the beginning of the 90s [148]. This because SOAs are compact devices which can be used for a number of applications exploiting several nonlinear effects including four-wave mixing (FWM), cross-gain modulation (XGM), cross-phase modulation (XPM) and self-phase modulation (SPM). Moreover, SOAs are active devices where signal processing can be performed at low input optical powers (0 to 10 dBm, typically) and, in some cases (XGM, XPM and SPM), processing operations are polarization independent, when polarization independent devices are employed. In the following years, plenty of all-optical processing schemes and applications, including wavelength conversion, all-optical logic gates, optical regeneration, switching, etc, have been demonstrated at ever increasing speeds, reaching hundreds of Gb/s [55]. However, despite the amount of results in research demonstrations, no actual products based on optical processing in SOAs have ever reached the maturity to enter the optical communication market or others. In the meantime, the advent of optical coherent systems in 2008 completely changed the perspective of optical signal processing. Indeed, from that moment on, only coherent polarization independent schemes were then suitable for processing optical signals carrying the novel modulation formats. Intensity dependent techniques were therefore relegated to applications where intensity modulation formats only, like multilevel pulse-amplitude-modulation (PAM), need to be processed [149, 150].

Among the all-optical functions, the wavelength conversion is one of the most relevant, as it is a key tool for proper traffic management in mesh networks. For this reason, several all-optical schemes have been studied in recent years for this application, exploiting a number of different techniques. However, no solutions have showed enough maturity to be reliably used in installed transmission systems. Indeed, a number of practical drawbacks are common to most of the proposed schemes, namely, a low conversion efficiency, wavelength dependence, large operating optical power, modulation format dependence, stability issues and others. On the other hand, coherent systems make the development of reliable and transparent wavelength converters even more desirable. Indeed, as sketched in figure 17, opaque wavelength conversion nodes require the use of a pair of complex coherent receiver and transmitter, which are made by several optical components, including an I-Q modulator, four balanced receivers, a local oscillator and an optical hybrid, and power hungry and expensive high speed electronics. All of these components could be replaced, if using a transparent optical wavelength converter, by a local tunable laser and a proper nonlinear element only, thus reducing, at the same time, the optoelectronic component count and overall power consumption of the network nodes. This advantage, moreover, becomes even larger if super-channels with even more complex transmitters and receiver hardware are considered. To make this possible, wavelength conversion schemes must properly address a number of strict requirements like:

• Operation at low optical powers
• Large conversion efficiency with large output OSNR
• Wavelength independence with flat conversion efficiency
• Polarization independence
• Low conversion penalty
• Format transparency

Microwave photonic (MWP) technologies are expected to play key roles in modern communication systems where the wireless and optical segments of the system seamlessly converge. The technology promises wide processing bandwidth, low loss over long transmission distances, and wide frequency tuning unlike anything achievable by traditional RF systems. These unique features allow microwave photonic technologies to grow beyond their rather simple initial purpose, namely as an alternative to the lossy, rigid, and heavy coaxial cables for transmission of high frequency RF signals.

Nowadays, microwave photonics systems routinely deliver the promise of generation, transport, processing, and measurement of RF signals with enhanced performance. Landmark demonstrations, among others, include the generation of ultra-broadband signals with tens of GHz bandwidth [160], distribution and transport of RF over fiber [161], wideband tunable filters [162], and photonics-enhanced radar systems [163]. This progress subsequently positioned MWP as a prime technology solution to the impending challenges in communications including the bandwidth bottleneck in communication systems and the internet of things.

While impressive, these landmark results were demonstrated in bulky systems composed of relatively expensive discrete fiber-optic components, which are sensitive to external perturbations, such as vibrations and temperature gradients. Hence, a major technological development for microwave photonics was the adoption of photonic integration technology. Leveraging photonic integration allowed dramatic footprint reduction of MWP systems with fairly high complexity, making it more comparable to RF circuits (figure 19).

The rise of the cloud has created the need for warehouse-scale data centres (DCs) that host applications (e.g. cloud computing, data storage, search) serving millions of users. These data centres comprise tens to hundreds of thousands of servers connected via optical links, generating massive server-to-server data traffic that doubles every 12 to 15 months [170]. Data centre to user traffic only takes about 16% of the total data traffic. About 77% of the generated traffic remains within a data centre, known as an intra-DC interconnection with a maximum reach of about 2 km. The communications between data centres are known as inter-DC interconnection, currently accounting for 7% of the total traffic and with a maximum reach of about 80 km due to latency constraints [171]. It is expected that the inter-DC connection will have a strong growth in the coming years to form virtual DCs that consolidate multiple data centres in the same region.

Currently, server I/O is dominated by 25 Gb/s pluggable optical transceivers based on on-off keying modulation and direct detection. A 100 Gb/s link is formed either by employing four parallel fibre or four wavelengths by coarse wavelength division multiplexing (WDM). With the adoption of four-level pulse amplitude modulation (PAM4) format by the IEEE P802.3bs task force, it is expected that the intra-DC I/O will soon increase to 50 or 100 Gbit s⁻¹-per-wavelength, forming 200 or 400 Gbit s⁻¹ per fibre using WDM [172].

For inter-DC interconnection, both direct detection and coherent detection are employed to facilitate the 400 Gb s⁻¹ interconnections. Commercial direct detection solutions rely on dispersion compensation modules and erbium-doped fibre amplifiers to compensate for the chromatic dispersion (CD) and loss, respectively, to enable 100 Gb/s per wavelength PAM4 transmission at the telecom C band [173]. Recently proposed direct detection techniques exploit digital signal processing (DSP) in conjunction with analogue-to-digital (AD) or/and digital-to-analogue (DA) converters to compensate for CD, obtaining >100 Gb/s data rate per wavelength for over 100 km transmission distance [174, 175]. Current coherent transceivers are designed for long-haul systems that emphasise the spectral efficiency and the tolerance to low optical signal-to-noise-ratio (OSNR). It requires power-hungry ADCs and DSP for the key functions including polarisation demultiplexing, equalisation, carrier recovery, clock recovery and decoding. This inevitably adds power consumption and latency that are at odds with the needs of DC interconnection. Nevertheless, coherent transceiver technologies, which offer full-field modulation on both polarisations and high receiver sensitivity, are considered as the only scalable option to accommodate the need for data rates in future data centres. Therefore, future research should focus on the reduction of power consumption, latency and cost for integrated coherent transceivers.

The general challenge for data centre interconnection is to increase the throughput under thermal, latency, space and cost constraints. This requires reducing transceiver power consumption and end-to-end communication latency, and accommodating WDM with an increased loss budget in a mega-size data centre. Future data centre interconnection will scale up to >400 Gb s⁻¹ per wavelength data rate, requiring a >50 GBd symbol rate and high order modulation formats (16QAM and beyond). It will be very difficult for direct detection transceivers to keep up with this trend as the required sensitivity will be too low to meet the loss budget.

Optimising coherent transceivers for data centre interconnection should consider its short-reach features, including small CD, negligible polarisation mode dispersion and nonlinearity. Current industrial developments primarily focus on reducing the DSP complexity in coherent transceivers, aiming for 400 Gb s⁻¹ coherent WDM pluggable optics with less than 15 W power consumption. A further increase of the data rate is limited by the achievable bandwidth and the resolution of the DACs and ADCs. The ultimate performance and power consumption of this DSP-based technological pathway will be constrained by CMOS technology that is approaching its fundamental physical limit [176]. A scalable technological pathway that allows a further reduction of power consumption by at least one order of magnitude (less than 1.5 W per 100 Gb/s) will be needed.

Another critical challenge is the end-to-end communication latency. The transceiver signal processing (including carrier and phase recovery, pulse shaping and impairment compensation), clock and data recovery (CDR), and data switching introduce a latency up to a few hundred microseconds that severely limits the performance of an intra-DC network. Tackling the latency constraints involves research on transceiver design as well as on the network and switching architecture. Current intra-DC network switching architecture is based on a folded Clos interconnect topology that has four levels of ethernet packet switches [177]. Optical switching can potentially achieve high radix (>1000 port) and fast switching speed (less than 2 ns). It offers a perspective of a flat switching architecture that can best adapt to the distinct traffic pattern in data centre networks (DCNs). In such DCNs, the transceivers and switches should be designed to handle short (20–300 ns) burst mode data traffic. For inter-DC communication, the round-trip time must be less than 2.0 ms [178], which indicates that the signal processing induced delay should be less than 1.2 ms, considering a light propagation delay of 0.8 ms through optical fibres.

Compared to the conventional DSP-based heterodyne coherent transceivers, homodyne coherent transceivers based on analogue signal processing may offer ample opportunities for a further reduction of power consumption and latency. Recent research predicted that it could potentially achieve power consumption of 2 W per 100 Gb/s under the assumption of 90 nm CMOS technology [179]. Analogue signal processing inherently has ultra low latency, and the homodyne coherent detection offers a 3-dB higher SNR compared to its heterodyne counterpart, allowing for an increased loss budget for FEC-free transmission. The intra-DC interconnection may largely benefit from these advantages. In the inter-DC interconnection, the flexible impairment compensation capability offered by DSP remains desirable. However, using analogue signal processing to assist or simplify DSP may reduce the power consumption and latency that are crucial in inter-DC interconnections.

Figure 21 compares the architectures of the current DSP-based coherent transceiver and the envisaged analogue-based homodyne coherent transceiver. At the transmitter side, the challenge of insufficient DAC resolution for high bandwidth signal generation could potentially be alleviated by using the optical domain signal synthesis, which in principle offers higher optical signal-to-noise ratio (OSNR) than the conventional laser plus modulator paradigm. The potential of using segmented modulator [180], phase locked directly modulated lasers [181] or parallel electroabsorption modulators [182] to generate high-order PAM or complex signal modulation, promises a reduction of transmitter power consumption, photonic chip size (thus high density) and cost.

Heterogeneous integration, due to its potential for large-scale processing for photonics, has been intensively studied over the last decade [189]. In silicon photonics, the integration of III–V materials by this approach enables increased functionality with on-chip efficient lasers, amplifiers, modulators and photodetectors. An example is an optical synthesizer that uses integration of pump lasers, nonlinear resonators and second harmonic generation (SHG) to stabilize lasers and enable 1 Hz stability and tuning across 50 nm [190]. It also leads to a reduced cost because of the use of automated CMOS processing and large wafer sizes.

In the last few years, motivated by the success of integrated nonlinear photonics, there has been tremendous interest in constructing nonlinear devices by heterogeneous integration [191–194]. The key advantage of this approach is the ability to introduce high-quality, single-crystalline materials for on-chip nonlinear processes, with much higher efficiency compared to traditional nonlinear devices due to very small modes and high nonlinear coefficients. The significantly improved efficiency can relieve the requirements of high pump power and large device size in previous optical signal processing [195]. Furthermore, heterogeneous bonding enables including nonlinear components into a fully integrated photonic system, which can potentially lead to scalable photonics integrated circuits (PICs) with optical signal processing functionality, whose cost can be 1000 times lower and power efficiency can be 1000 times higher compared to previous bench-top systems [190].

The key technology in heterogeneous integration is the thin film bonding method [196], either by direct wafer bonding or benzocyclobutene (BCB) bonding. Usually the former method is preferred because it enables more reliable bonding and device performance without the problems of polymer degradation. During fabrication, as shown in figure 22, a single-crystalline thin film is bonded onto low index insulator layer (usually SiO₂), followed by substrate removal and waveguide etch. The waveguides are usually fully cladded with insulators, which provides high confinement and compactness for nonlinear applications.

As the network demand maintains its impressive growing rate, comprehensive network systems need to be developed and implemented. Optical super-channels are an obvious response towards high capacity point-to-point links because of the inherent increase of spectral efficiency. Modern modulation formats are bringing the tools to prolong the transmission distance and at the same time increase the amount of the information exchanged [204]. However, optical networks are more complex than a point-to-point link because of the need for a capability to redirect and transfer optical channels independently. Hence, advanced optical multiplexing and demultiplexing systems require development to allow the handling of very high bandwidth throughput super-channels [205].

Recently, multiple types of all-optical super-channel multiplexing have been demonstrated based on frequency multiplexing (AO-OFDM, Nyquist WDM) or time multiplexing (OTDM, Nyquist-OTDM) [206–208]. The all-optical construction of these super-channels can unlock the large bandwidth of the installed optical network. Indeed, because of the inherent bandwidth limitation of the RF signal sources, electrically multiplexed channel such as in OFDM can only be aggregated to a sub-hundred GHz total bandwidth. Even with recent larger than 100 GBd single carrier channels, the optical network capacity growth depends upon a denser than WDM usage of the available spectrum.

In the case of Nyquist WDM, before multiplexing, each optical sub-channel is shaped in a rectangular spectrum and the corresponding sinc-shaped pulse. As such, each sub-channel occupies its Nyquist bandwidth, and consequently multiple channels can be grouped with channel spacing equal to the symbol rate, using a simple optical combiner. Consequently, the spectrum previously dedicated to guard-bands to maintain WDM channels separated can increase the link bandwidth [207]. Under certain circumstances channel multiplexing can go well beyond the density of standard WDM. When orthogonality can be achieved between data carried by the sub-channel, spectral overlap does not result in inter-symbol interference, as illustrated by the AO-OFDM super-channel multiplexing [208]. Using an optical Nyquist pulse source as well as appropriate delay between channels, super-channels can be build based on OTDM. A near perfect Nyquist pulse enables a single optical source to feed multiple low speed modulators in parallel [209]. By aligning the sub-channels sinc-shaped pulse peaks and nulls, an ISI-free operation point can be obtained.

At the difference of the Nyquist WDM and Nyquist OTDM multiplexing techniques, AO-OFDM does not require pre-shaping of the sub-channels. The simple optical addition of channels spectrally separated by a strict integer of their baud rate is sufficient, as long as the transmitter bandwidth is sufficiently large to avoid distortion of the data pattern. The one optical tool that allows the implementation of the different super-channel types is the optical comb source [210], as it provides a simple frequency spacing and phase locking between independently modulated sub-channels and hence reducing penalties.

Up until now, guard-bands have allowed for advanced modulation formats with high spectral efficiency to be used on single carrier channels; furthermore, they relax the optical ROADM requirements. Standard filtering techniques have now been improved to an impressive level, such as wavelength selective switches (WSS) with sub-GHz resolution or free-space filter with very large edge roll-off [206]. Current implementation of optical nodes is orthogonal with the operation of granular bandwidth sized super-channels with fractional internal guard-band or even more difficult over-lapping spectra sub-channels [205]. Furthermore, multiplexing smaller optical channels makes the use of currently available DSP possible, at the cost of high-power consumption and limited progression margin [207]. Nowadays, optical networks are required to be agile in bandwidth in function of the customer need, driven by the large peak-to-average aspect of video consumption, while the cost per bit, both energetic and monetary, must be lowered.

With the maturation of photonic integration, high quality optical filters can be designed in large quantity. In the case of tightly packed sub-channels, optical filtering has been shown using enhanced delay line interferometers to produce a compact Nyquist WDM super-channel [211]. Also, symmetrically large super-channels can be demultiplexed and received in minimum volume and cost, at a spectral efficiency greater than standard WDM.

Of course, an accurate matching filter is of particular importance in the case of demultiplexing. Indeed, the manipulation of sub-channels cannot impact the adjacent ones; nor affect the channel over few nodes crossing. Here lie the difficulties of multiplexing and demultiplexing channels into super-channels. Modulation format optimization can be applied to the sub-channel, constituting the different flavor of super-channel with cyclic prefix or offset-QAM, applied in order to maintain optimum transmission performance. Furthermore, such advanced filters can only be applied to super-channels without spectral overlap and hence with lesser than 100% channel bandwidth occupation. For an optimal spectral use, super-channels based on Nyquist-OTDM or AO-OFDM need to be considered and recently active demultiplexing systems have been demonstrated.

In the case of Nyquist-OTDM super-channels, the all-optical demultiplexing is achieved through the use of an optical sampling function [212]. Such a sampling function can be achieved through a sampling optical pulse interacting in a non-linear medium. Similarly, high speed electro-absorption modulators and Mach-Zehnder modulators can be used to temporally filter the sub-channel of interest using narrower than multiplexed pulses or applying the corresponding match filter. At the cost of some complexity, the channels multiplexed in time can be demultiplexed in wavelength through the use of optical Fourier transform.

In the case of the data carrying signal of the sub-channels are overlapping in time as well as spectrally, the demultiplexing requires a matching filter. For example, in the case AO-OFDM the matching filter is spectrally shaped as an a Sinc function, or its equivalent FFT approximation. The demultiplexing of AO-OFDM channels has been proved based on an all-optical implementation of the FFT of various orders, followed by an optical sampling function. The demultiplexing of such super-channels will invariably degrade adjacent channels in the case of an optical network node for example.

A solution to this limitation has been reported recently, capable of super-channel demultiplexing and sub-channel deletion compatible with all types of super-channels and overlapping spectra ones in particular: terabit interferometric add and drop, and extract (TIDE) [213]. The demultiplexing of a sub-channel either for deletion, detection, or routing, can be obtained through an interferometric design. TIDE consists of an all-optical node in which sub-channels can be multiplexed and/or demultiplexed into AO-OFDM super-channel. As has been argued before, optical filtering is unsuitable for this case. As illustrated in figure 25, the incoming super-channel is split in two copies and then inserted in two arms of a fiber Mach-Zehnder interferometer. The idea is to select the sub-channel in one of the two arms with a matched filter, to sample it and reshape it. Then the sub-channel is suppressed interferometrically at the output coupler when the obtained copy beats with the unmodified super-channel transmitted in the second arm. The interferometric nature of the sub-channel extracted ensures the clean suppression and that the adjacent channels remain unaffected.

The bizarre and non-intuitive working of nature at the smallest scale described by quantum mechanics has massive potential in revolutionizing modern information technology. Quantum computing, which is based on the principles of superposition and entanglement of quantum states, allows parallelization of computations and hence tackles problems which are impossible to be solved classically. A suitable platform and a precise and reproducible nanofabrication method for the networking and scaling of quantum systems is required to achieve integrated quantum information devices.

The nitrogen vacancy (NV) color centers in diamond have emerged as favorable candidates for quantum information systems. The electronic ground state of the NV center forms a spin triplet which can be polarized by excitation with 532 nm wavelength. One of the spin states fluoresces brighter than the other with a zero phonon line (ZPL) at 637 nm. This spin dependent fluorescence emission is utilized for optical readout of the spin state with coherence times comparable to that of trapped ions, even at room temperatures [214]. Thus, NVs can form the heart of a quantum information system, the quantum bit (qubit), as schematized in figure 26. In addition, the magnetic coupling of the NV spin state with the atomic nuclei in the diamond matrix, such as N or ¹³C impurities, allows the control of the nuclear spins with impressively long coherence times on the order of seconds. Creation of quantum networks in diamond is possible by spin state entanglement of nearby NVs via photonic excitation through the optical guiding structures connecting them [215]. Demonstrations illustrating the NV-NV entanglement over large distances have opened the door for few qubit protocols in diamond [216].

Encoding and processing information in quantum bits (qubits) open new opportunities within computing and communication. Photons are the natural carrier of quantum information over long distances enabling quantum communication. To this end, highly efficient and thus scalable sources of coherent photonic qubits are required, which may be obtained by the use of single-photon emitters. A quantum dot (QD) is a solid-state single-photon emitter, which can naturally be embedded in photonic nanostructures; figure 28 illustrates a QD in a nanophotonic waveguide. In this case, a deterministic photon-emitter interface can be obtained enabling a deterministic source of single photons or a spin-photon interface. A coupling efficiency exceeding 98% has been reported [230], which corresponds to a single-photon cooperativity of 50 if all broadening mechanisms are suppressed. Remarkably, transform-limited single-photon emission has been demonstrated with QDs even at a moderate cryogenic temperature of around 4 K [231] paving the way for such a coherent photon-emitter interface. Novel opportunities and functionalities immediately follow including: (1) an on-demand source of coherent single photons, (2) nonlinear operation at the single-photon level, or (3) the deterministic generation of multi-photon entangled states.

• (1) A QD in a waveguide can be operated as a highly efficient source of single photons, e.g. by optically exciting the QD that emits photons into the well-defined mode of the waveguide. Subsequently, the photons may be coupled off-chip into an optical fiber by, for example, engineered evanescent coupling. Multiple sources of single photons may be constructed from a single on-demand source by implementing optical switching and subsequent delays [232]. Here, the rapid recombination rate of the QD source and the ability to strongly suppress even slow decoherence processes means that the sources could readily be scaled up to N ~ 10 simultaneous independent photons, and further improvement is within experimental reach [233]. Section 34 describes an alternative approach to an integrated single-photon source based on nonlinear optics.
• (2) The efficient and coherent photon-emitter coupling may be used to mediate photon-photon nonlinear interaction. In this case, every photon incident on the emitter will interact and since a quantum emitter can only scatter a single photon at a time, two incident photons are entangled by the interaction with the emitter. Such a nonlinear response can be the basis for single-photon switches and sorters enabling, e.g. quantum non-demolition photon detection or deterministic Bell-state analyzers [234].
• (3) The introduction of a coherent single spin in a QD opens new avenues for realizing more advanced photonic resources. One promising approach involves multi-photon entanglement generation by repeated excitation and emission from a single QD where the spin is coherently rotated in-between each emission and excitation cycle [235]. With such an approach potentially a long string of entangled photons can be generated, which after generalization to 2D cluster states constitutes a universal resource for one-way quantum computing [236] or all-photonic quantum repeaters [237].

Quantum processing that utilizes the counterintuitive properties of quantum mechanics, such as superposition and entanglement, has opened the possibility for revolutionary technologies that will have a profound impact on our society, Examples of such disruptive technologies are quantum computing, secure communication system and quantum enhanced sensing [241, 242]. Although, different physical systems, such as superconducting circuits, trapped ion and electron spin, have achieved major progress in recent years and are benefiting from significant industrial investments, photonic quantum technologies are going to play an essential role in the quantum revolution. These schemes are based on photons—quanta of light—and have some intrinsic advantages over other physical systems: (i) they provide a natural interface between quantum computing systems and low-loss communication links; (ii) their coherence time is long enough for the logic gate control; and (iii) they are nearly free of environmental noise, namely thermal and electromagnetic noise.

Over the last two decades, theoretical analysis and preliminary demonstrations have established the supremacy of photonic quantum processing [243]. However, the major challenge for its practical application lies in the development of a highly reliable and large-scale photonic quantum system. Building such a system based on free-space optical components seems unviable due to its lack of robustness, stability and scalability. Integrated optics, on the other hand, allows for the harnessing of CMOS technology and hence achieves multi-scale integration of passive and active components in a mature silicon-based platform that can be scaled up for manufacturing.

The ultimate goal of integrated quantum optics is to utilize the existing CMOS infrastructure to realize a fully integrated quantum photonic chip that contains high efficiency photon sources, low-loss and reconfigurable linear waveguides and near-optimal single photon detectors. Many of the core components of such a chip have already been demonstrated in integration platforms, including low-loss circuits, tunable phase shifters and near-optimal detectors. An on-chip photonic quantum state source that generates single and entangled photons on-demand, however, is the ‘Achilles heel’ of this scheme. The on-chip photon source is the key to move photonic quantum processing into the real world.

Two technical approaches are being pursued for building an on-chip single photon source. One is collecting single photons from a ‘single-emitter’ embedded in the circuits (so called quantum dots). Although impressive results have been obtained, the remaining barriers for this scheme are indistinguishability of the single photons from distinct sources and the lack of spectral flexibility. The second approach is to directly generate single photons within the photonic circuits via nonlinear optical effects, such as spontaneous parametric down conversion (SPDC) or spontaneous four wave mixing (SFWM). Figures 29(a) and (b) show the schematic of SPDC and SFWM in LiNO₃ and silicon waveguides, respectively, and their underlying principles are illustrated in figures 29(c) and (d). In this nonlinear optics approach, the signal and idler photon are time-correlated such that the detection of one photon heralds the existence of the other, hence it is referred to as a heralded single photon source. Compared with the quantum dots, the heralded quantum light sources can generate highly indistinguishable heralded single photons at room temperature and their frequency is tunable within a broadband spectrum. An intrinsic limitation of this scheme, however, is the low efficiency of the stochastic photon generation.