Graphene is an ideal material for optoelectronic applications. Its photonic properties give several advantages and complementarities over Si photonics. For example, graphene enables both electro-absorption and electro-refraction modulation with an electro-optical index change exceeding 10−3. It can be used for optical add–drop multiplexing with voltage control, eliminating the current dissipation used for the thermal detuning of microresonators, and for thermoelectric-based ultrafast optical detectors that generate a voltage without transimpedance amplifiers. Here, we present our vision for graphene-based integrated photonics. We review graphene-based transceivers and compare them with existing technologies. Strategies for improving power consumption, manufacturability and wafer-scale integration are addressed. We outline a roadmap of the technological requirements to meet the demands of the datacom and telecom markets. We show that graphene-based integrated photonics could enable ultrahigh spatial bandwidth density, low power consumption for board connectivity and connectivity between data centres, access networks and metropolitan, core, regional and long-haul optical communications.
In the past 25 years, data traffic has exponentially grown, and optical fibre amplifiers have greatly contributed to this expansion in traffic1. Optical fibre amplifiers enabled the Internet era, offering the faster data rates required for smart phones and social media2. The next wireless communication technology, known as 5G (fifth generation)3, requires an increase in bandwidth of three orders of magnitude (>500 Mb s−1) for each user and all objects connected to the Internet (ref.4), as the 5G evolution is driven by the growing mobile communication markets and the development of the Internet of Things (IoT)5. This growth in communication is predicted to increase the global gross domestic product to ~US$1.9 trillion6, with ~50 billion connected devices in use by 2020 (ref.7). Therefore, there is urgent demand for a technology that can meet requirements in terms of bandwidth and power consumption. Considering these growth projections, one needs to be mindful of the impact of communication technologies on global energy consumption and global warming8. At present, the information and communication technology (ICT) industry accounts for 2–2.5% of all greenhouse emissions, according to the International Telecommunications Union, and this is predicted to increase to ~4% by 2023.
Photonics is poised to play an increasingly important role in ICT (Fig. 1a), since the fixed high capacity links are largely based on photonic technologies. Photonic devices need to support ultra-large bandwidth operation, for example, 200 Tb s−1 in a single fibre9 and >10 Tb s−1 cm−2 in integrated Si photonics chips10. To achieve this, the key components of Si photonics, photodetectors and modulators, need very high performances in terms of speed (≥25 Gb s−1), footprint (<1 mm2), insertion loss (<4 dB), manufacturability (>106 pieces per year) and power consumption (<1 pJ bit−1). To date, these requirements have not been fulfilled in one system11. Furthermore, in terms of production volumes, photonics is not yet comparable to microelectronics12, even if the increase in demand for optical networks would, by 2021, lead to an average global Internet Protocol (IP) traffic of 3.3 ZB (zettabytes), corresponding to an average usage data rate of ~800 Tb s−1 (refs13,14). In the context of the IoT7, other applications, including infrared (IR) sensors, biosensors, environmental sensors, metrology, quantum communications and machine vision, will require even larger production volumes13,15.
The telecom network can be divided into three segments: access, aggregation and core (Fig. 1a). The access network is the interface between subscribers and the immediate service provider. The aggregation network aggregates all the input data streams from tributary access networks, converging towards the higher-level core network. The aggregation network includes local and metropolitan networks, which then converge to regional networks. A local area network (LAN) interconnects computers within a limited area, such as a residence, school, laboratory, university or office building. A metropolitan area network (MAN) interconnects users with computer resources or communication servers in a geographical area or region larger than that covered by even a large LAN. A regional network covers areas of radii from approximately 10 km to 500 km. The core network is the set of communication facilities that interconnect primary nodes, delivering routes to exchange information between various sub-networks.
The points of access to the access network are the land-line and wireless individual subscriber networks and the radio base stations for wireless communications. Signal routing occurs in data centres located in all segments of the communication network. In total, maintenance and evolution of the telecom network requires >1 million devices per year16,17. All equipment supporting very high bandwidths, such as the 100GE (Gigabit Ethernet)18,19 electro-optical interfaces (that is, converters from electrical to optical signals or vice versa), and with high connection capacity for access networks, aggregation networks and data centre interconnections is based on optical technologies (Box 1). The global optical high-capacity transceivers market is estimated to reach ~$6.87 billion by 2022, driven by the availability and cost-effectiveness of devices with speeds between 100 and 400 Gb s−1 (ref.20).
Data traffic at the periphery of the communication network originates from devices with IP addresses (for example, laptops, surveillance cameras and smart phones) (Fig. 1b), and it is expected to increase at a rate of ~1.6 billion connected devices per year — projected to be ~12.5 billion by 2020. Photonic technologies are increasingly playing a role in access networks. Applications range from fibre to the home (FTTH) scenarios to the backhauling of wireless nodes (e.g. access nodes or base stations)21. The emergence of IoT and, eventually, the Internet of Everything (IoE) requires intelligent management of the huge network of interconnected ‘things’, according to the International Telecommunications Union. The IoT vision is for ubiquitous ‘smart objects’ to exchange information anywhere and anytime using their individual IP addresses. ‘Smart’ is defined as sensing combined with decision-making and artificial intelligence22. Over the past 15 years, the price of sensors, processors and networking has decreased according to Moore’s Law, giving rise to new products arising from interconnected machines and devices (or ‘things’) via the network. The consequences of such technologies are far beyond the individual cases and have the potential to change our society, as the internet has done. Thanks to the widespread deployment of WiFi, it is easy to add new networked devices to the home, office or other locations. The adoption of IPv6 enables an almost unlimited number of devices to be connected to networks23. Major system vendors and market forecasts estimate ~28 billion connected devices by 2021 (ref.21) and ~100 billion by 2025 (ref.24), with a market of up to tens of trillions of dollars by 2025 (ref.25), leading to a ‘smart’ societal change.
The increase in the number of connected devices requires a large and pervasive photonic communication infrastructure (Fig. 1a), with an optical bandwidth considerably greater than 25 THz in installed optical fibre systems26,27. The 5G network will need energy-efficient cells (10–100 times more energy efficient than 4G), based on photonic28 or millimetre wave connection for fronthaul and backhaul transport29. Fronthaul transport is implemented between the centralized baseband units and the relevant remote radio units to enable a seamless connection without affecting radio performance. Backhaul transport realizes connections from the centralized baseband units to the IP core network to perform end-to-end solutions for low latency (that is, the time delay due to processing)30. Also under consideration is millimetre-wave-based ultrahigh capacity (>1 Gb s−1 per user and >10 Gb s−1 per remote antenna unit), short-range (<1 km) access31. As a result, there is a need for photonic interconnections that are cost efficient (<$10/Gb s−1 by 2020 and <$1/Gb s−1 by 2025)32,33 and have large channel bandwidths (>100 GHz). A description of the evolution of mobile communication networks with the advent of the 5G era can be found in the Supplementary Information.
At present, optical interconnections in data centres are mainly between boards that provide the platform on which the electronic components and optical or electro-optical devices are connected. In the near future, the number of optical interconnections will increase34. As a result, by 2021, the production of optical interconnections is predicted to be >10 million per year35. The photonic devices — most commonly, modulators (Box 1; Box 2) and detectors (Box 2) — needed to meet these requirements are based on LiNbO3 (refs36,37), semiconductors such as InGaAsP/InP (refs38,39) and those used in Si photonics11 (Table 1). Devices based on LiNbO3 and semiconductors are established40, whereas Si photonics is a newer and faster-growing field17,32. The parameters used to compare modulators are modulation efficiency, insertion loss and the figure of merit (FOM) for a phase shifting functionality (FOMPM) (Box 2). The modulation efficiency of interferometer-based modulators is defined as VπL, where Vπ is the voltage required to achieve a π phase shift of the optical carrier and L is the length of the phase shifter.
LiNbO3 Mach–Zehnder modulators have low (<0.4 dB mm) insertion loss and high (>50 V mm) VπL. According to Table 1 the device length needed to achieve a π phase shift with a driving voltage of 2 V is between 2.5 and 5 cm. InGaAs/InP detectors (ref.40) or Ge/Si (ref.41) have comparable performances in terms of both responsivity Rph (~1 A W−1) and bandwidth (>40 GHz) and, at present, have a higher Rph than single-layer graphene (SLG) photothermal detectors, which have so far demonstrated42 Rph ~0.36 A W−1, and up to 100 GHz bandwidth43. However, for technologies to become widespread, devices must be mass produced, cost efficient, reproducible, reliable, and compliant with existing semiconductor processes and environmental regulations. With these considerations in mind, for large-scale production, Si photonic11 devices are preferable to InGaAsP/InP ones because the technological processes are the same as those already present in Si foundries commonly used in the semiconductor industry. Thus, the Si photonics platform for single-wavelength components is a practical technology and permits co-packing of electronic functions with light sources44. Given that graphene photonics is compatible with Si photonics and other materials such as SiN and SiO2, in the following we focus on the potential for integration of graphene with Si-based technology.
A high-performance photonic device requires high-precision fabrication equipment. For example, an optical lithography node size of 65 nm in a 300 mm fab (a semiconductor fabrication plant) provides a good trade-off between performance and cost, even if optochip costs are only ~20% of today’s transceiver costs44. For Si photonics, Si-on-insulator (SOI) is used, costing >$1,000 (at 2017 prices), five times more than Si (<$200 at 2017 prices). Considering the ethernet roadmap to 2020, the important components for Si photonics — photodetectors and modulators — must be high performing in terms of speed (>50 Gb s−1), footprint (<100 µm2), insertion loss (≤1 dB) and energy consumption (~100 µW GHz−1). To date, these parameters cannot be satisfied in one system because of the trade-off between electro-optical properties and loss46.
In Si photonics, light with wavelengths between 1,300 and 1,550 nm is guided by Si and is detected by Ge p–i–n photodetectors integrated on Si (ref.41). Graphene exhibits both electro-absorption47 and electro-refraction48 and, hence, can be used for light modulation and photodetection49. There remains a need for optical transmitter and receiver modules (transceivers) integrating waveguides with photodetectors and modulators on one chip, in parallel with wafer-scale processing50.
The potential of graphene for photonics and optoelectronics has been discussed in previous reviews49,51,52,53,54. Here, we focus on the key arguments underpinning the development of graphene-based integrated photonics for high-speed datacom and telecom. Graphene-based photodetectors will remove the need for Ge epitaxy, which is currently used for Si photonics photodetectors, by replacing the Ge p–i–n photodetectors with a single or double SLG. Graphene photodetectors are not spectrally limited55, unlike Ge p–i–n photodetectors41, which operate below wavelengths of 1,600 nm. Graphene-based photodetectors can reach bandwidths of ~260 GHz (ref.56), as a consequence of the high carrier mobility, µ, of SLG (Box 3). Moreover, in voltage-detection mode, graphene-based photodetectors can function at zero dark current57.
Another key component of transceivers is the electro-optic modulator. Graphene-based modulators have advantages over Si-based modulators. They are capable of broadband 30 GHz electro-absorption operation58 based on modulating the resonance of a micro-ring resonator in and out of the critical coupling condition, they are compatible with complementary metal oxide semiconductor (CMOS) processing, and enable post-processing fabrication and the use of different substrates. SLG does not require Si or Ge doping. Hence, the waveguide can be Si, SiN, SiO2 or another transparent material. Practically, this implies a post-processing shift in manufacturing from front-end to back-end-of-line. In addition, graphene technology does not necessarily require expensive SOI wafers, or implantation for junctions, and Ge growth for detectors. Because SiN and SiO2 waveguides are wider than Si photonics ones, the lithography node can be relaxed. The waveguide size is ~0.5 µm for Si, ~1.5 µm for SiN and ~ 8µm for SiO2. All these factors will simplify the technology and reduce costs, making small and medium production volumes more affordable because the initial non-recurring engineering is less than in a SOI-based line. This means that the volume threshold to implement a product in a Si fab can be reduced, enabling the cost-effectiveness of medium-volume products (10,000–100,000 chips per year), thus opening up markets wider than those for consumer electronic products, which require higher volumes.
The basic components of communication systems are waveguides, modulators and photodetectors. Modulation of light is one of the key operations in photonic integrated circuits59 (Box 1). The properties of guided light, such as amplitude, phase and polarization, can be modulated by altering specific properties of the guiding medium. For example, electro-absorption modulators59,60 induce modulation of the amplitude of the propagating light through the modulation of the optical absorption of the waveguide. Electro-refractive modulators61,62 alter the phase of the propagating light by changing the effective index neff (the index of refraction determining the phase velocity of light in a waveguide)63. The Franz–Keldysh effect60,64,65,66 (Box 4) and the quantum-confined Stark effect67,68 (Box 4) can be exploited in electro-absorption modulators.
SLG is a broadband absorber with 2.3% absorption at any wavelength at normal incidence, as a consequence of its lack of a bandgap69. This corresponds to 10log10(1−0.023) ∼0.1 dB. By superimposing SLG on a Si waveguide, it is possible to enhance the interaction of SLG with light70. Absorption increases significantly from ~0.1 dB at normal incidence to thousands of dB cm−1 when light propagates along the waveguide71,72. If the Fermi energy, EF, of SLG is shifted, the absorption is reduced or cancelled73. When the EF is larger than the energy of the propagating photons, Eph, SLG is more transparent as a result of Pauli blocking74 and the index of refraction changes70. The electrical and optical properties of SLG depend on carrier concentration. Defects can give a background loss independent of EF and introduce losses70, thus degrading the electro-optical properties. In the Kubo model75, the presence of defects can be taken into account by introducing intraband transitions due to long-range scattering, owing to the presence of impurities, trap states and screening76, accounted for by the scattering time, τ76 (Box 3).
To illustrate the effect of τ on background loss, we consider single SLG and double SLG devices operating at 1,550 nm (corresponding to Eph ≈ 0.8 eV) (Fig. 2). For the single SLG device, one layer of SLG is placed on a Si-ridge waveguide with core dimensions of ~480 nm (width) × 220 nm (height) on top of a 60 nm-thick Si slab (Fig. 2a). A 5 nm-thick dielectric layer is placed between Si and the SLG. The slab waveguide and SLG have electrical contacts on their surface, and the result is a Si–insulator–SLG capacitor70,71. In the double SLG device, two SLGs are placed on an undoped Si waveguide (Fig. 2c). One of the SLGs is separated from the Si waveguide by a 5 nm-thick dielectric layer. Above this SLG is an additional 5 nm-thick dielectric layer, followed by a second layer of SLG. This arrangement — two SLGs and the dielectric spacer — forms a SLG–insulator–SLG capacitor70,72,77. In both cases, silica cladding surrounds the capacitors. When a voltage is applied, carriers are driven into the Si waveguide core as well as into the SLG layer (Fig. 2a) or in the two SLGs (Fig. 2c), and the accumulation of carriers causes an EF shift70,71,72,77:
where vF ~9.5 × 107 cm s−1 is the Fermi velocity and nS is the surface charge density. The voltage needed to accumulate nS is the sum of two contributions: the potential across the insulator, in both the Si–insulator–SLG and SLG–insulator–SLG capacitors, and the quantum capacitance78:
where q is the electron charge, Cox is the oxide capacitance per unit area and VDIRAC is the voltage corresponding to the charge-neutral Dirac point, with K = 1 or 2 for single and double SLG, respectively.
The computed optical absorption in single or double SLG as a function of EF is shown in Fig. 2b,d. A commercially available mode solver is used to evaluate neff and the optical absorption of the waveguide mode. SLG is modelled with an in-plane dielectric constant obtained from the optical conductivity70, and the out-of-plane dielectric constant is taken as equal to the graphite dielectric constant79. We use the closed formula for the complex optical conductivity80:
Here, σ0 = e2/4ħ is the SLG universal conductivity81,82 and kBT is the thermal energy. T = 300 K is used in the simulations. Three values for τ are considered — 10 fs, 100 fs and 300 fs — for a free-space wavelength λ of 1,550 nm and Eph of 0.8 eV. When |EF| < Eph/2, photons may induce interband transitions83. This results in light absorption at rates as large as 0.1 and 0.2 dB µm−1 for single and double SLG, respectively71,72. In this EF range, the absorption curves are almost independent of τ: interband transitions dominate over intraband ones83. When |EF| > Eph/2, interband transitions are forbidden as a result of Pauli blocking83, and SLG would ideally be transparent. However, the smaller the τ, the larger the absorption, even when |EF| > Eph/2.
For Si-based waveguides composed of SLG (Fig. 2a), an EF shift is obtained by applying a voltage through the Si–insulator–SLG capacitor75,77. Carriers accumulate in the SLG and the underlying Si waveguide and, as a result of plasma dispersion84 (Box 4), the carriers in the Si waveguide cause absorption. Thus, although carriers make SLG transparent for high ns (|EF| > 0.5 eV), they also make Si opaque. Instead, for small ns (|EF| < 0.4 eV), the Si losses (green curve, Fig. 2b) are negligible, and absorption is mainly a consequence of interband transitions in SLG. When |EF| > 0.4 eV, interband transitions in SLG are forbidden (Fig. 2). The black, blue and red curves, which represent the contributions to losses arising from SLG for different τ, have a similar behaviour in double SLG. Losses reach a minimum in the range 10−2–10−3 dB μm−1 for |EF| ≈ 0.5–0.6 eV and then increase because of intraband transitions. However, losses in Si increase monotonically with ns. The net result is a waveguide that is never as transparent. For Si-based waveguides covered with double SLG, losses decrease to minimum values of ~10−2, 10−3 and 5 × 10−4 dB μm−1 for τ = 10 fs (black curve), 100 fs (blue curve) and 300 fs (red curve), when |EF| ≈ 0.5−0.6 eV (Fig. 2d). If ns is increased further, so that |EF| > 0.6 eV, losses increase again as a consequence of intraband transitions. Both single and double SLG electro-absorption modulators can be obtained by varying the absorption between |EF| < Eph/2 and |EF| > Eph/2 (refs71,72).
Equation 3 implies that a change in EF affects both the real and the imaginary part of the conductivity of SLG, more specifically the absorption, α, and refractive index, n. Therefore, SLG may be used to realize phase modulation70,77. In a MZI configuration, phase modulation allows for binary modulation formats, such as non-return-to-zero85. We demonstrated a SLG-based MZI phase modulator, transmitting over a standard fibre link86. This paves the way for a number of complex modulation formats for increased spectral efficiency in transmission systems, for example, phase-shift keying87, differential phase-shift keying88, quadrature phase-shift keying89 and quadrature amplitude modulation89. All these are likely to increase the spectral efficiency of a communication link90,91,92.
The computed Δneff as a function of bias in the Si–insulator–SLG and SLG–insulator–SLG capacitors are shown in Fig. 3a. We compute the voltage according to equation 2 by using a SiO2 dielectric layer and compare the result with that of typical Si modulators93,94,95,96,97,98,99,100,101,102,103 based on p–n junctions or capacitors in the same range of voltages. By biasing SLG to |EF| > Eph/2, where absorption is small, a neff modulation ~2 × 10−3 can be obtained79. At an operating wavelength of 1,550 nm, this accounts for a phase modulation ~8.1 rad per millimetre of propagation along the waveguide, over ten times larger than in typical Si phase modulation based on reverse-biased p–n junctions59.
For comparison, we evaluate ∆neff as a function of bias for a Si reverse-biased p–n junction modulator with p and n doping ~5 × 1017 cm−3 (Fig. 3a). The modulation achieved by using two SLGs gives the largest ∆neff, as well as the steepest Δneff. In the 2.5–4 V range, ∆neff of double SLG as function of voltage is higher than that of other technologies, even though SLG is absorbing. Above EF ∼0.45 eV (5 V), SLG becomes transparent, and ∆neff has the steepest variation with applied voltage, and this allows for a short-phase modulation section.
In Fig. 3b, we compare the FOM of graphene modulators with those of p–n junction Si modulators, Si–insulator–Si modulators104 (black curve) and InGaAsP membranes on Si (refs105,106). The points are experimental data taken from literature, whereas the lines are theoretical estimations for optimized modulator parameters. The Si–insulator–Si modulator comprises a vertical stack of a Si thin layer and a 5 nm-thick layer of SiO2 followed by a layer of poly-Si. This Si–insulator–Si forms the waveguide core. Electrical contacts on the Si and poly-Si slabs form a capacitor across the insulating SiO2 layer. By applying a voltage to the contacts, charge accumulates at the Si/SiO2 and poly-Si/SiO2 interfaces. Through plasma dispersion, these charges lead to phase modulation. ∆neff is significant, because charges are accumulated in the middle of the waveguide, maximizing overlap of the optical mode.
Single or double SLG electro-absorption modulators can also be used as electro-refractive modulators, for example, in a MZI modulator with SLG phase shifters on its two arms70,77. The FOMPM for SLG-based electro-refractive modulators is about ten times higher than that for Si-based phase modulators when µ of SLG is high (τ > 100 fs), comparable to that of InGaAsP (refs105,106). This higher FOMPM is a result of the combination of the large electro-refractive effect, VπL, and low αloss. A larger ∆neff requires either a smaller voltage or a shorter L to obtain a π phase variation. For SLG modulators70, VπL < 2.8 V mm (single SLG) and <1.6 V mm (double SLG) or <0.7 V mm for a double SLG embedded in the core of the waveguide77. In a typical Si p–n junction modulator, VπL ∼20 V mm (ref.101), and in Si–insulator–Si capacitor modulators, VπL ∼2 V mm (ref.107).
In p–n junction modulators or Si–insulator–Si modulators, ∆neff is due to the depletion or accumulation of charges (Box 4). However, charges absorb light108, resulting in a large insertion loss. In the Si–insulator–Si modulator, poly-Si, which typically has a larger αloss than crystalline Si, is part of the waveguide core. As a result, the losses in Si–insulator–Si108 modulators and in depleted p–n junction modulators109 are ~5 dB mm−1 and 0.55 dB mm−1, respectively. These limitations of Si–insulator–Si modulators — the large insertion loss and the presence of lossy poly-Si — are circumvented in SLG modulators. Therefore, in single SLG modulators, αloss is comparable to depleted p–n junction modulators (Fig. 3b). In a phase shifter composed of SLG, a large ns (EF > 0.45 eV at an operating wavelength of 1,550 nm) must accumulate in both SLG and Si to achieve carrier accumulation in the capacitor and, therefore, a considerable electro-refraction or electro-absorption effect. The doping of Si induces αloss (see equation 12). Conversely, in the double SLG modulator, αloss can be small (Fig. 2d) because Si doping is not required (Fig. 2b). If the modulator is biased to an EF where interband transitions are inhibited, αloss arises only from intraband transitions. For τ > 100 fs, αloss as low as 2 dB mm−1 can be obtained70. In terms of the overall FOMPM, double SLG modulators with τ > 100 fs are at least 3–5 times better than InGaAsP/InP modulators38,39, 5–10 times better than Si modulators and 10–20 times better than LiNbO3 modulators36,37.
For a double SLG device in the SLG transparency region, the theoretical estimate is VπL ∼1.6 V mm for an undoped waveguide with double SLG70, with an overall FOMPM < 2 V dB for τ = 100 fs and FOMPM < 1 V dB for τ = 300 fs (Fig. 3b). These theoretical values, if achieved, would result in devices that outperform Si photonic devices. However, hybrid technologies relying on III–V semiconductor membranes bonded on doped Si photonics waveguides can achieve performances similar to SLG photonics. For example, static phase shifters based on a 150 nm-thick InGaAsP membrane on 5 nm Al2O3 deposited on a Si waveguide showed VπL = 0.47 V mm and an insertion loss of ~1.9 dB mm−1 at an operating wavelength of 1,550 nm (ref.106). This modulation efficiency corresponds to FOMPM ≈ 0.91 V dB (ref.106). Similarly, an InGaAsP membrane bonded on 10 nm SiO2 on an n-doped Si waveguide core has both high modulation efficiency and high speed105. With this configuration, a 3 dB modulation bandwidth of 2 GHz and a data rate of 32 Gb s−1 with VπL = 0.9 V mm were reported105. The InGaAsP–thin-oxide–Si stack exploits the capacitor concept first used in the Si–insulator–Si capacitor (SISCAP)93, with the advantage of replacing the lossy poly-Si top layer with a highly efficient electro-refractive n-doped InGaAsP membrane105,106. The SLG–thin-oxide–SLG capacitor exploits the same effect, with good electro-refractive efficiency, but with the advantage that the capacitor is placed on a passive waveguide made of, for example, either Si, SiN or SiO2 (refs70,77).
Plots of FOMEA, defined as ER/IL, as a function of nS and EF for single and double SLG devices at different voltages are shown in Fig. 3c,d. The peak-to-peak voltage, VPP, is the voltage needed to switch the modulator from a high to a low transmission level, or vice versa, and more specifically, to emit a ‘0’ bit (smaller transmission) or a ‘1’ bit (larger transmission). The resulting ER is the ratio between the high and low level transmitted power. FOMEA depends significantly on τ: in a double SLG device for VPP = 2 V, FOMEA ∼5 for good-quality graphene (τ > ∼100 fs), while FOMEA ∼3 for poor-quality graphene (τ > ∼10 fs).
To understand how these performances compare with the specifications of 100GE standards, we recall that the modulator for a 100GE CLR4 system is required to provide ER~4.5 dB, an IL~3 dB, therefore FOMEA ~1.5 (Box 2). SLG-based electro-absorption modulators can achieve these requirements in both single and double SLG configurations, even when VPP = 1 V (Fig. 3c,d).
For comparison with SiGe electro-absorption modulators, let us consider a double SLG modulator with VPP = 2 V and FOMEA ~5. This outperforms the Ge waveguide electro-absorption modulator in ref.110 that operates at VPP = 2 V, with ER~4.6 dB and IL~4.1 dB, corresponding to FOMEA = 1.12. In another Ge-based electro-absorption modulator111, FOMEA ~2–3.3 was reported for VPP = 4 V and λ = 1,610–1,640 nm. Similar results were reported for GeSi electro-absorption modulators112, revealing that double SLG electro-absorption modulators are more promising than GeSi modulators.
We note that SLG-based electro-absorption modulators operate across the |EF| = Eph/2 threshold, where interband transitions are excluded as a result of Pauli blocking71,72. In this condition, as a consequence of the simultaneous modulation of α and n, both amplitude modulation and phase modulation occur113. This results in an instantaneous variation in optical frequency, which is termed chirped modulation85. A positive chirp modulation is useful for transmission in anomalous dispersion fibre links to compensate for the high-frequency components of light, which travel faster than the lower-frequency ones85. The difference in velocity of the high-frequency and low-frequency components is caused by the negative chirp produced during propagation through the fibre link. This effect has been reported113 to enable 10 Gb s−1 transmission of light at λ = 1,550 nm along 100 km of standard single-mode fibre even with an input ER as low as 3.5 dB.
The capacitance of the modulator is the main parameter affecting the modulation speed of Si–insulator–SLG and SLG–insulator–SLG capacitors (Fig. 2a,c). Increasing the distance between the SLG layers improves the cut-off frequency of the capacitor, but increases the signal voltage and worsens both modulation efficiency and power consumption. These factors can be improved by embedding the SLG layers in the waveguide core77, as well as by optimizing the length of the electrical leads (that is, the metal connections between contacts and SLG) and ensuring that the SLG length exceeds the width of the waveguide core.
Graphene integration on SiN
The platform of Si photonics can be a base upon which processes or technologies are developed for integrated photonics — most importantly, transceivers. Modulation and detection in Si photonics are key elements of transceivers and require a layer of crystalline Si sandwiched between a buried SiO2 insulator and a top cladding of SiO2 (ref.61). This allows propagation of single optical modes in the waveguide (SOI wafer technology)114. The Si guiding layer can be p-doped or n-doped to enable modulation61, and an epitaxial Ge layer can be grown on it to enable detection through carrier photogeneration115. A more convenient configuration is that of a pair of SLG layers, that is, a capacitor consisting of a SLG–insulator–SLG stack on a passive waveguide72.
SLG–insulator–SLG capacitors have potential advantages over Si photonic modulators. First, the fabrication is not dependent on the waveguide material, electro-absorption modulation or electro-refractive modulation. Photogeneration is provided only by the SLG structures fabricated after the waveguide in post-processing116. As a result, the optical guiding circuit does not require doping or Ge epitaxy for detection. Instead, it is fully passive, and only core material with no doping is required, which can be exploited to simplify the fabrication technology. Given this flexibility, SiN can be selected as the core material. SiN is amorphous, with a refractive index larger than that of silica (nSiN = 1.98 at 1,550 nm), and it is transparent in the visible and IR regions117. The large difference in the refractive index between the core (SiN) and the cladding (SiO2) ensures mode confinement down to ≤1µm2 and therefore waveguide miniaturization118. To fabricate the waveguide, SiN is deposited on buried SiO2 between SiN and Si. The SiN platform, compared with the Si photonics one, is low cost because it requires standard Si wafers rather than speciality wafers such as SOI. SiN single-mode waveguides are typically 1 µm wide at 1,550 nm (ref.119). This means that the lithography resolution is relaxed compared with Si single-mode waveguides, which are 0.5 µm wide11. The SiN core can be defined with either a low-resolution 400 nm node optical lithography stepper or with a mask aligner with a resolution <1 µm (ref.120). The cost of the low-resolution mask, compared with the high resolution one required for Si photonics, could be at least fivefold less121. The consequence of this reduction in cost is that the production volumes to amortize the investments are smaller, opening medium-volume (for example, telecom) and small-volume (for example, ultra-long-haul optical systems) markets. Thus, whereas manufacturing in a standard CMOS line leads to mass-market products, the graphene photonics approach permits the use of already amortized fabs and lowers the cost of fabrication, therefore opening medium-volume markets.
The subassembly, which integrates the SLG photonics chip, laser and fibre array, comprises the largest fraction of the total manufacturing cost122. The high cost is related to laser integration, fibre array coupling and pigtailing. In a SLG photonics circuit, by exploiting SiN as a passive waveguide platform, laser integration and fibre coupling are more fabrication tolerant than in Si photonics. Because the difference in the refractive index between core and cladding is smaller in SiN waveguides than in Si photonics ones, the numerical aperture of the SiN waveguide is closer to that of the laser or fibre to be matched, and the impact of the packaging is reduced. As a further example, a double SLG stack can be assembled on any other core material, such as silica123 or other materials.
Si photonics transmitters are mainly based on MZI modulators124. Depending on the ratio between the bit time, TBIT (the time duration of a single bit), and the transit time of the optical wave through the modulator, TT, modulators can be classified as lumped70,77,93,94,105 or travelling wave96,97,98,99,125. TBIT is the reciprocal of the bit rate, defined as the number of bits per second. For example, in a transmission with a bit rate of 10 Gb s−1, each bit has duration TBIT = 1/bit rate = 100 ps. For a lumped modulator, TT ≤TBIT, and for a travelling wave modulator, TT ≥TBIT (ref.126). A MZI modulator is characterized by two electrical parameters: the total capacitance of the two optical phase shifters through which the voltage is applied and VPP, which sets the extinction ratio of the optically encoded signal and, consequently, the required energy. The energy performance efficiency used in electronics, also called the power-delay product127, here is called the energy cost per bit128, corresponding to the energy associated to one bit, in pJ bit−1 or mW GHz−1. In a lumped configuration, an electronic circuit (driver) charges and discharges the total capacitance at VPP for each data transition129. The maximum power consumption of the two phase shifters is given by the energy consumption in a bit time130:
In a travelling wave configuration, the MZI modulator is driven by a terminated electrical transmission line placed on the MZI phase shifting arms131. The total capacitance is split into N capacitances along the line. The power required to generate an on–off modulation is100,130:
where ZL is the characteristic impedance of the transmission line100. From equations 4 and 5, the energy cost (or energy per bit), expressed as Pout/bit rate, scales inversely with bit rate in the travelling wave configuration, whereas it is constant in the lumped configuration. The energy cost reduction for the travelling wave compared with lumped wave at the same VPP is 2CTZL multiplied by the bit rate. For very large bit rates, for example, the 56 Gb s−1 case discussed at the Optical Internetworking Forum OIF CEI-56G (Common Electrical Interface at 56 Gb s−1)132, this reduction can be relevant. If we consider the capacitance of a lumped Si modulator, CT ~ 1 pF, and the load impedance of a travelling wave modulator131, ZL = 50 Ω at a bit rate of 56 Gb s−1, then an approximately sixfold energy reduction is obtained. For this to be achieved using travelling wave operation, long-length (compared with radio frequency wavelength) modulators are required109. In this respect, SLG photonics outperform Si photonics. Phase modulation in Si photonics MZIs requires doping of the Si waveguides. Therefore, it is inherently lossy, with losses ~0.55 dB mm−1 (ref.109) or 5 dB mm−1 (ref.107). The losses in double SLG phase modulators in sections of MZI can be < 5 dB mm−1 for τ > 100 fs (ref.70) at EF~0.5 eV (Fig. 2d).
This low loss in SLG can be exploited to increase the modulator length (Fig. 2b,d), allowing the travelling wave configuration. VπL for SLG is better than that for Si photonics MZIs. Thus, the signal voltage at a given device length is smaller, and this contributes to further reduce the energy cost per bit. A modulator for a 100 GBE interconnect with ER=8 dB (Box 2) can be achieved with VPP ~0.37Vπ (refs124,133). Assuming a double SLG lumped modulator, VπL = 1.2 V mm (ref.70) and L = 400 µm (ref.134), VPP ~ 1.1 V. With CT ~1 pF (ref.132), we get 0.6 pJ per bit from equation 4, independent of bit rate and insertion loss. For comparison, a Si photonics modulator132 has a consumption of 2–3 pJ per bit and IL~2.6 dB at 1,310 nm. If we consider a graphene-based travelling wave configuration with a device length of 1.2 mm and a bit rate of 56 Gb s−1, TBIT ~TT = 17 ps and VPP ~0.35 V, the energy consumption is ~25 fJ per bit for ZL = 50 Ω and IL~1 dB.
Double-SLG-based modulators have a better FOMPM (~1 V dB) than optimized Si photonics travelling wave MZI (FOMPM = 12 V dB)125 as well as lumped Si photonics MZI modulators (3 V dB)134. The FOMPM of double SLG-based modulators is comparable to that of InGaAsP membranes on Si (ref.105). The performance of state-of-the-art Si photonics modulators and SLG-based modulators are compared in Table 2.
Ethernet-type interconnections in optical networks are based on IP packet switching to route data streams135. Packet switching can be performed in both data and telecom networks135. In telecom networks, this is achieved mainly in parts of the network where IP transfer is supported between access networks. The aggregated traffic of the different data streams is becoming sufficiently large (100 GBE and higher)18 so that new functionalities can be performed by optoelectronic switches. At each switching node, opto-electrical and opto-electro-optical conversions are required to support electrical disaggregation, switching and re-aggregation, and aggregated packet routing136,137. In the overall aggregation and opto-electro-optical conversions, latency and energy consumption138 are major issues. Si photonics technology can enable the switching of data streams, mitigating the bottleneck of latency and power consumption. Switching can be performed directly in optics, without resorting to opto-electro-optical conversion and without data packet handling139. This can be an advantage when large data streams are aggregated and switched. For example, power saving and latency reduction can be achieved when longer persistent data streams, like elephant flows, can occur in a datacom network140. An elephant flow is an aggregated data stream transmitted through a fixed path for a sufficiently long time (>10 s)140. In this case, the data stream can be routed along different optical paths (or different fibre links) with dedicated hardware (for example, an optical circuit switch)141. Other applications include periodic data centre backups to prevent and minimize data losses142; reconfigurable add–drop multiplexing of large data streams143; protection of a connection path through restoration and an alternative pre-provisioned path in both data and telecom networks144; implementation of network resource provisioning through software-defined networking145, initially done for datacom, but now also suitable for telecom146; and reconfigurable fibre connections in a meshed photonic network, the so-called optical cross connection144.
Another important field in telecommunications is wireless. Highly directive antennas will be used in 5G networks to steer the beam of a radio signal in the direction of several final users with high available throughput (up to 10 Gb s−1 in the microwave or millimetre wave spectrum), also known as massive beam forming147. Beam-forming antennas may lead to new types of network that rely on more efficient resource allocation and optimized power consumption and require a more extended use of the frequency spectrum148. The interconnection of several radio base stations equipped with beam-forming antennas can be achieved in line with the 5G fronthaul and backhaul network evolution by using optical circuit switches28. As a consequence of the increased radio link bandwidth and the need for fast network responsiveness (<1 ms latency, depending on the service) with respect to previous generations, switching is set to become a pervasive optical function25. Using dedicated software that considers each single hardware network element as ‘virtualized’ (that is, representable as a controllable network entity characterized and/or summarized by some relevant tuneable parameters), optical switches will be able to handle aggregated 100 GBE and beyond data streams19, being operated remotely on demand through all computational resources supplied by one or more interconnected data centres. This approach defines a smart photonic cloud network149, capable of delivering an undetermined number of services directly to the network system and/or the final user. A high virtualization in optical switch operations can be performed by leveraging the potential of photonic integrated circuits, which can be efficiently monitored and controlled by software, further lowering the equipment cost per bit, the footprint and the power consumption146.
Over the years, many approaches have been proposed for realizing optical switches139,150, such as electro-optic (mainly in Ti:LiNbO3)151,152,153,154,155, acousto-optic156,157,158, thermo-optic159,160,161,162,163, liquid crystals164,165,166,167, microelectromechanical systems (MEMS)168,169,170,171,172,173 and semiconductor optical amplifiers174,175,176,177,178,179. These devices require expensive equipment and could be replaced by Si photonics integrated circuits to comply with the requirements of miniaturization and cost-effectiveness180, through mass manufacturability181.
Si and SiN optical add–drop multiplexers (OADM)182,183,184 have been realized to demonstrate switching185,186,187,188,189,190,191. OADMs are four-port filters in which a micro-ring resonator is placed between two waveguides and is coupled to the waveguides183,184. This component is designed for wavelength division multiplexed (WDM) optical networks, that is, systems that make use of multiple wavelengths (or colours) to transmit different data streams and increase the capacity of the system192,193,194. In the multiplexer, the waveguide that carries the incoming signal is usually referred to as the bus183. In a typical dense wavelength division multiplexed (DWDM) optical link195, the bus carries 72 wavelengths between 1,520.25 nm and 1,577.03 nm (the so-called C-band)196, spaced 100 GHz apart. The second waveguide of the multiplexer is referred to as the drop183. When the incoming light has a wavelength that coincides with the resonance of the micro-ring resonator, light is transferred from the bus to the drop. This operation mode is called DROP184. When the incoming light is out of resonance, it proceeds unswitched, and the operation mode is called THROUGH184.
DROP–THROUGH operation is determined by coherent interference inside the ring183. The light circulating in the micro-ring resonator interferes at each round trip with the incoming light from the THROUGH channel. At each round trip, a fraction of the field is extracted from the ring in the DROP channel, while the remaining part continues in the micro-ring resonator183. This operation implies a latency that, as in any resonant circuit, grows with the quality factor of the resonance. However, the quality factor cannot be too high or the spectrum of the signal travelling inside the ring becomes heavily distorted. Off-resonance, the periodical interference cancels out, and the field from the incoming bus waveguide continues in the THROUGH channel with no DROP. Switching configuration (that is, the possibility of enabling /disabling the DROP of each single wavelength in a DWDM link) is increasingly important in optical networks193. In a micro-ring resonator, this corresponds to tuning and/or detuning the resonances with respect to the wavelength of the signal that has to be dropped. For a switch to drop a single wavelength, the micro-ring resonator free spectral range (FSR) needs to be greater than the WDM spectrum to avoid spurious drops197. Detuning the resonance of the micro-ring resonator can be achieved by metal heaters198,199,200,201,202 in the interlayer dielectric (waveguide cladding oxide) on top of the waveguide. Electrical energy is supplied to the heaters, which diffuse heat to the waveguide202. The thermo-optic effect203,204 results in an increase in the index of refraction of Si with temperature205, leading to the detuning of the micro-ring resonator (that is, a shift in resonant frequency)202.
SLG may represent a breakthrough in switching thanks to a new approach that exploits high-µ SLG electro-absorption rather than tuning. By placing two SLGs on the micro-ring resonator waveguide and by changing the voltage applied to those SLGs, losses in the micro-ring resonator can be varied from very large (~1,000 dB cm−1) to small (<10 dB cm−1)79. When SLG losses are large, the field circulating in the micro-ring resonator is entirely absorbed in a single round trip, the interference with the incoming signal is suppressed, and DROP is disabled. When the loss is negligible, light can resonate and DROP is enabled. The main difference between this SLG-based switch and the Si photonics counterpart is that enabling and/or disabling of the DROP is obtained by suppressing, rather than detuning the micro-ring resonator resonances206. As a consequence, SLG-based switching has a major advantage over its Si photonic counterpart. In Si photonics, DROP disabling is obtained by placing the resonance between two adjacent channels in the ITU-T (Telecommunication Standardization Sector of the International Telecommunications Union) DWDM grid. In this way, the crosstalk between channels, defined as the ratio between the power of spurious and main signals, becomes critical, and the system tolerance becomes tight (that is, all the wavelengths in the transmission system should be locked with high accuracy to the ITU-T frequency grid)207,208,209,210,211,212. Another advantage of a SLG switch is the power consumption. The usual scheme based on thermal tuning of the Si micro-ring resonator leads to a continuous power consumption ~0.11 nm mW−1 (ref.213). Assuming a micro-ring resonator resonance trimming due to fabrication errors ~1 nm, this corresponds to ~9 mW for each ring213. In the SLG switch, the variation in losses for resonance suppression is obtained by capacitive charging71,72. Thus, there is no power consumption in static operation, which is the normal state for a switch. Energy is consumed only during capacitance charging206. A further advantage is that the functionality that depends on SLG is part of post-processing and is therefore independent of the waveguide platform.
The performance of a SLG-based switch is shown in Fig. 4a. This device, which is composed of the waveguides depicted in Fig. 2c, is schematically represented in Fig. 4b. The transmission spectra in the C-band for a four-port micro-ring resonator (radius of 10 µm) with a FSR of 1.2 THz and a bandwidth of 20 GHz (at 1 dB) are shown (Fig. 4a). The device has an ADD–DROP filter suitable for WDM switching79. The isolation from adjacent channels is achieved with τ ~300 fs (Fig. 4c). The performance in terms of suppression of resonance and, hence, DROP disabling, is determined by the maximum absorption achievable in a round trip along the micro-ring resonator. As a result, the double SLG configuration is preferable to the single SLG configuration71,72. DROP requires the coherent interference of signals travelling within the micro-ring resonator183. This is the ON state, and the losses in a single round trip must be as small as possible. SLG needs to be transparent for DROP to be performed. Similar to the phase modulation case, this transparency is related to τ. DROP spectra in the ON state can give a quantitative estimate of the influence of τ on the performance of the switch (Fig. 4d). For τ ~10 fs, the loss of SLG is large (>10−2 dB µm−1) at any value of EF (Fig. 3d). In a 10 µm ring, this accounts for a loss ~0.6 dB per round trip. This limits the coherent superposition of signal replicas inside the micro-ring resonator to ~20 waves. The result is a ~4.8 dB loss in the DROP channel (Fig. 4c). A larger τ gives rise to a smaller αloss (Fig. 2d) ~1 × 10−2 for 100 fs and ~5 × 10−3 dB µm−1 for 300 fs. Therefore, a coherent superposition of a greater number of replicas inside the micro-ring resonator (~200 waves and ~1,000 waves for ~100 fs and 300 fs, respectively) gives a much smaller loss in the DROP channel ~0.7 dB and 0.15 dB for ~100 fs and 300 fs, respectively. For the THROUGH channel, an insertion < 1 dB is obtained when operating the ring in the OFF state, when setting |EF| < Eph/2, with Eph ~ 0.8 eV in the C-band (Fig. 4d). Losses as large as 0.1 dB µm−1 may arise because of allowed interband transitions (Fig. 2d) with an almost negligible dependence on τ. Incoming light from the bus is partially coupled to the ring, the light is rapidly absorbed in the micro-ring resonator at a rate ~6 dB per round trip, and no coherent interference is formed. As a result, the micro-ring resonator DROP channel is disabled, and the light travels to the THROUGH port.
Graphene photodetectors have been extensively reviewed49,51,53. and integration of graphene-based detector arrays with CMOS electronics was demonstrated214. Here, we focus on waveguide-integrated and speed-optimized photodetectors relevant to Si photonics. Rph should be comparable to that of Ge devices in Si photonics, that is, ~0.85–1.15 A W−1 at an operating wavelength of 1,550 nm (ref.41).
The key principle of photodetectors is the conversion of absorbed photons into an electrical signal. Several detection mechanisms have been identified for graphene photodetectors, including photovoltaic215,216, photo-thermoelectric216,217, bolometric218, photogating219 and plasma-wave-assisted220. Each of these may become dominant in different photodetector configurations, such as SLG p–n, metal–SLG and single-double–SLG junctions216,221. SLG has unique properties and advantages for photodetectors49,53. For example, SLG has a zero-gap band, which implies frequency-independent absorption69, allowing for light detection from the UV to far-IR region with a single material222 and in a single technological step. Another advantage is the potential for less than picosecond photovoltage generation221 and fast photo-switching rates, thus far up to ~270 GHz (ref.56). SLG also has high internal quantum efficiency with a ratio of electrons produced per photons absorbed >80%42. In SLG, the photo-thermoelectric effect can be optimized for both high detection speed (25–100 GHz) and efficiency223, potentially above 100% as a result of hot carrier multiplication224. The photo-thermoelectric effect involves the following steps: first, electron–hole pairs are generated by the absorption of photons, and second, ultrafast carrier scattering generates hot electrons and holes on a <50 fs timescale224,225. These electrons and holes generate a local photovoltage via the Seebeck effect226, which drives a current from source to drain. When the light is switched off, electrons cool back to equilibrium in 2−4 ps (refs225,227,228).
This cooling time yields the intrinsic limit of the photo-switching rate. The photo-thermoelectric effect can be highly efficient, as a large fraction of the photon energy is captured as electron heat owing to the ultrafast carrier scattering and weak coupling to phonons222,225. This results in a voltage source, because the Seebeck effect is an electromotive force that generates a voltage rather than a current226, which can be used in transceivers needing a voltage to drive the receiver electronics. In contrast to a photovoltaic detector, in which the photogenerated current is typically amplified and converted into a voltage by a transimpedance amplifier228, a photo-thermoelectric-based graphene photodetector generates voltage directly225, and the transimpedance amplifier can be replaced by a simpler voltage amplifier. This property has advantages in terms of cost reduction and power consumption. In addition, direct voltage detection may overcome issues with the dark current of photovoltaic schemes with bias.
The photo-thermoelectric effect has been exploited extensively for graphene-based photodetectors55,57,229,230,231,232,233,234. On-chip integrated photodetectors with Si photonics have been reported232,233,234,235,236, typically based on metal–SLG–metal structures evanescently coupled to Si waveguides. In these photodetectors, the guided mode enables longer interaction between SLG and the optical waveguide compared with free-space illumination. This longer interaction raises the optical absorption above 2.3% and, by increasing the interaction length, up to almost 100% of the light is absorbed and can contribute to photovoltage. Because of the evanescent coupling, the typical length needed to achieve nearly complete absorption in metal–SLG–metal photodetectors is ~40–100 µm. A speed-optimized graphene photodetector with a rate of ~50 Gb s−1 was reported237. The device consisted of a chemical vapour deposition (CVD)-grown SLG on a Si waveguide operating at 1,550 nm. The evanescent field of the mode propagating in the Si waveguide overlaps with a p–n junction as a consequence of the EF shift at the metal interface of one of the contacts. One limitation is the contact metal in the evanescent field of the optical mode, which leads to a reduction in Rph. The small size gave a capacitance of ~20 fF and a resistance of ~185 Ω, with a large bandwidth237.
The highest Rph photodetectors to date are ~40 µm in length and integrated on a 520 nm-wide Si waveguide42. The drain and source were placed asymmetrically with respect to the waveguide core to create a p–n junction near the optical mode and obtain a net voltage drop. A gate was also used to maximize the Seebeck coefficient of SLG. SLG was encapsulated in hexagonal boron nitride (hBN) to improve µ to ~40,000 cm2 (V−1 s−1) and the Seebeck coefficient, resulting in a larger Rph~0.36 A W−1. The side contacts provided a resistance as low as 77 Ω. The 3 dB bandwidth response was 40 GHz, very close to the target value for an optical communication receiver238.
These results are promising for optical communication links, but there is a drawback: upon application of a bias, a continuous current flows. However, if two SLGs are arranged as shown in Fig. 5, the lower SLG provides thermoelectric generation and voltage detection, while the upper SLG acts as a split-gate, tuning the lower SLG to the optimized Seebeck coefficient. Calculations based on realistic parameters, for example, a Seebeck coefficient ~0.2 mV K−1 (feasible for high-quality SLG with µ > 10,000 cm2 V−1 s−1) with very low (EF ≤ 40 meV) residual charge density and a cooling thermal conductance ~70 nW K−1 m2 (ref.226) predict Rph > 0.8 A W−1 or, in case of voltage detection, that is, the measurement of the electrical output in volts per optical input (in Watts), Rph > 100 V W−1.
Another important performance metric of photodetectors is the normalized photo- to dark-current ratio, NPDR = Rph/Idark. The higher the NPDR, the larger the photodetector noise rejection and the ability to perform when interference (noise) is present. To achieve higher NPDR, Idark must be reduced, and Rph must be maximized. A promising route to increase Rph while minimizing Idark is to create a Schottky junction with rectifying characteristics (that is, a diode) at the SLG–Si interface239. By operating a Schottky diode in reverse bias (photoconductive mode), Idark is suppressed compared with Rph, and the entire Schottky contact area contributes to photodetection. A compact (5 µm in length), waveguide-integrated, plasmonic-enhanced metal–SLG–Si Schottky photodetector was reported to have Rph ~0.25 mA W−1 at 1.55 µm (ref.239). When the same detector is reverse biased with 1 V, Rph becomes ~85 mA W−1, and Idark becomes ~20 nA (ref.239). This detector configuration shows a one order of magnitude increase in Rph over that of the standard metal–Si configuration without SLG. By taking advantage of the Schottky diode operation in the reverse bias, Rph could be further increased239 up to ~0.37 A W−1 at 3 V, comparable to that of state-of-the-art SiGe devices41.
Most devices reported to date are on the laboratory scale and have contacts fabricated using metal lift-off 240. This is not suitable for the very large-scale integration required for modern chip manufacturing, because lift-off has limitations, such as redeposition of metal, formation of ears at the metal edges and partial retention of the metal240,241.
A standardized SLG–CMOS-compatible contacting scheme is yet to be developed. Studies reporting full integration of the wafer and SLG242,243,244,245 are limited to examples in which the SLG is integrated at the last level of integration242 or combined with metal contacts through lift-off243,244,245. This limitation hinders the adoption of SLG technology by the semiconductor industry. A SLG–CMOS-compatible integration module consisting of a sequence of processing steps in conventional CMOS tools, which guarantees compatibility with the reliability standards of the semiconductor industry, is needed to persuade industry to adopt SLG as a viable and reliable alternative to conventional materials. To allow the integration of SLG-based devices heterogeneously packaged with Si technology (Fig. 6a), SLG should be integrated through modules similar to those used to integrate semiconductor devices246.
At present, the connections between devices in semiconductor applications are typically achieved by Cu damascene modules247. In this process, developed by IBM248 and Motorola249, the dielectric is patterned by dry etching with trenches or vias in which the conductor metal is later deposited (Fig. 6b). On the dielectric, a metal that overfills the trenches is deposited. Then, chemical mechanical polishing is used to remove the overburden metal on top of the dielectric. This process is known as ‘damascene’, by analogy to the art of incrusting wires of Au (and sometimes Ag or Cu) on the surface of Fe, steel or bronze247. The integration of a SLG device on a waveguide contacted through a damascene module is described in the sequence shown in Fig. 6b. The initial six steps describe SLG integration on a waveguide, and the remaining steps are related to the damascene contact module248,249.
The fabrication does not require complex implantation sequences to activate the semiconductor locally. In addition, there is no need for seeding layers to epitaxially grow crystalline materials on the waveguides for light detection or modulation, in contrast to the fabrication of Ge photodetectors41,250 and SiGe or Ge electro-absorption modulators60,251. SLG integration reduces the complexity and the number of steps compared with 3D device integration60,251, as well as the total temperature budget to integrate the active detector or modulator on the Si waveguides. A reduced temperature budget decreases the threat to the integrity of the devices processed earlier on the exposed wafer252,253,254 and allows for flexible hetero-integration. The ease of integration of SLG devices with Si, as well as the reduced footprint of the SLG devices, should afford photonics technology increased cost-effectiveness compared with the competing semiconductor solutions.
To assess the costs of SLG–CMOS integration, progress in the integration maturity is needed. The fabrication steps in Fig. 6b require engineering to control performance and reproducibility and to achieve compatibility with the standards of the semiconductor industry. SLG can now be grown with similar quality to exfoliated SLG on metal or metal-coated substrates255. Growth of 5 cm × 50 cm SLG with >99% oriented grains is possible256 for 300 mm wafers. Direct growth of SLG on Si wafers (hence, avoiding a transfer step) has been realized257. However, this method will not be adopted by the semiconductor industry because it requires wafers with processed devices. The highly diffusive catalytic metal (in this case, Cu)255 is brought into contact with the dielectric of the target wafer and will diffuse into it when exposed at the temperature at which SLG growth occurs (at least 300 °C)258. This temperature will destroy the integrity and performance of the dielectric and/or lifetime of the devices259,260. The present understanding is that the growth of high-quality SLG requires a metal catalyst and high temperature, as well as an efficient separation of SLG from the metal catalyst after growth255 followed by transfer to the target wafer261,262 (Fig. 6b). The manipulation of SLG is one of the most critical steps for SLG integration on Si and should be done in a controlled environment during transfer. SLG is an impermeable layer for molecules263. An uncontrolled environment results in contamination at the transfer interface, leading to uncontrolled traps and random strain fluctuations, one of the dominant sources of disorder in SLG devices264. Transfer has the advantage of allowing the interface between SLG and the target wafer to be engineered. Charge traps at the interface between the oxide and SLG result in a distribution of positive and negative doping puddles265, which affect the local EF, resulting in non-uniform electro-absorption75 and µ reduction266, with decreased device performance. The transfer of SLG to the target surface should be engineered to secure reliable adhesion. SLG has no dangling bonds on the surface to chemically interact with the surrounding dielectric, and the adhesion to the dielectric substrate is secured through van der Waals forces267.
SLG encapsulated between hBN flakes enables a high µ at room temperature251,268,269. hBN has an atomically flat surface that significantly reduces electron–hole puddles compared with SiO2 (refs270,271). This suggests that a fundamental step towards control of the performance of large-area CVD-grown SLG can be achieved through the integration of SLG sandwiched between hBN. A single layer of wrinkle-free hBN can be grown on sapphire272, paving the way for the integration of engineered hetero-stacks on the wafer scale273. The growth of a dielectric on SLG (Fig. 6b) without affecting the SLG opto-electric performance and avoiding defect formation and/or chemical interactions is another challenge. Plasma-assisted dielectric deposition technologies tend to induce defects in SLG274,275. The atomic layer deposition (ALD) of high-k dielectrics has been studied276. The role of the nucleation density in order to achieve rapid layer closure has been extensively investigated277,278. It is imperative for the starting surface to provide enough reactive sites for reactions with the ALD precursors277. On self-passivated materials, the nucleation of the dielectric typically occurs at the reactive defect sites. Therefore, it is possible to correlate the efficiency of the closure of the ALD layer to the quality of the SLG, where a less effective closure is observed for high-quality material, as the nucleation density is linked to the number of defects279,280. The state of the art is non-robust for future material improvement. Therefore, alternative seeding approaches need to be developed to realize the growth of the dielectric independent of SLG quality.
A further required development relates to the metals incorporated as SLG contacts (Fig. 6b). Some metals are not compatible with the CMOS production environment (as detailed in the International Technology Roadmap for Semiconductors) because their use affects device reliability and yield. Typically, only Al, W, Cu, Ni, Co, Mo, Ti and Ta are compatible with CMOS fabrication281. The contact architecture to SLG also has an impact on the complexity of the integration scheme, and edge contacts have proved to be the optimal architecture282 and the easiest to achieve in a damascene module.
The performance gains and cost savings should push the industry to invest in the integration of graphene and related materials in the Si production line. Once answers are developed for these challenges, integration will comply with the standards of the semiconductor industry and will pave the way for the adoption of technologies based on graphene and related materials as a standard in the portfolio of the semiconductor industry. An alternative approach is based on single-crystal transfer in a predetermined position283. This involves growing single-crystal SLG at nucleation points predefined on a Cu support. These single crystals are set to overlap with the device on the destination wafer after transfer. The advantages are as follows: first, the growth of multiple individual single crystals is less challenging than growing a single wafer-size crystal; second, the predetermined position ensures the transfer of SLG crystals only where needed by design (for example, onto the waveguide modulators and detectors); third, transfer printing can be used to populate an entire wafer284.
Another consideration is the integration of the electronics (electrical driver and transimpedance amplifier, TIA) and photonics on the SLG-based photonic circuit. The electronic integrated circuit (EIC) wafer and the optical integrated circuit (OIC) wafer (Fig. 6c) can be integrated by thermally bonding contacts or by using Cu pillars. This is important because the SLG photonic layer (OIC) consists of a post-processed SLG stack on passive guiding structures. SLG post-processing is therefore the final stage in the fabrication of the photonic layer. Contacting and metallization are achieved in the back-end-of-line, with no perturbation in the SLG quality or optical or electrical characteristics. This ensures full compatibility of SLG photonics with the electronic circuitry in the integration process.
Conclusions and outlook
The telecom and datacom industries are driven by the continuous increase in requirements for the bandwidth of communications. The advent of the 5G communication era will boost the bandwidth requirements as a result of the introduction of communications with the world of high-definition virtual reality, augmented reality, gaming, as well as connected objects, more specifically, the IoT. Managing the increased bandwidth demand requires significant advances in photonics hardware, well beyond incremental improvements.
We consider graphene-integrated photonics for telecom and datacom systems as an evolutionary step in integrated photonics and, in particular, Si photonics. The main advance enabled by the adoption of graphene is post-processing on a passive waveguiding structure, dedicated only to passive optical circuitry. All active functionalities, such as modulation, detection and switching, are graphene-based and applied on the passive underlying optical circuit. The separation of the guiding circuit from active functionalities leads to a technology that does not require a full integration approach, unlike Si photonics, which is strongly tied to CMOS processing.
We analysed the main functions (modulators, detectors and switches) of graphene photonics and compared them with established technologies. The graphene photonics modulator exploits carrier accumulation in a SLG–insulator–SLG capacitor and can provide the necessary bandwidth to match the telecom roadmap evolution, combined with low energy consumption (0.1 pJ bit–1), size miniaturization (phase shifter length <0.5 mm or electro-absorption length <0.1 mm) and, most importantly, a FOMPM ~0.1 V dB for the phase shifter, which represents a significant improvement with respect to existing technologies, combined with a FOMEA of ~3 dB for the electro-absorption case. Graphene phase shifters can be driven with a voltage ≤1 V. This is relevant to the cost of the transmitter and its power budget, as the signal to the modulator is usually supplied by an electronic driver. If the voltage is ≤1 V, it is possible to design simplified electronics with no need for a specific electronic driver.
At present, graphene bolometric detectors provide responsivity of at least ~0.5 A W−1 (ref.285), while those based on the photothermal effect have at least ~0.4 A W−1 or 10 V W−1 responsivity, which is bound to increase with mobility optimization, combined with negligible dark current. Such detectors can be used in either current or voltage mode. The conventional configuration is based on current detection and requires electronics, such as a transimpedance amplifier, for current-to-voltage conversion and amplification. However, the photothermal effect in graphene generates a voltage; therefore, the voltage configuration is more natural for graphene. The receiver design is thus simplified and is less costly to produce.
Graphene-based modulation and detection are key elements in a point-to-point transmission system. A telecom or datacom system also requires optical switching to route the signals through the communication network. The graphene-based switch is a building block that, analogous to a modulator, can be based on a SLG–insulator–SLG capacitor, which can switch one input signal from one output port to another. The main feature of the SLG–insulator–SLG capacitor switch is that it can enable or disable output ports by leveraging the charge accumulation in the capacitor by means of a voltage. This is an improvement over conventional current-driven thermo-optic switches in Si photonics, which are operated under a continuous electrical current flow.
Thus, graphene photonics offers a combination of advantages in terms of both performance and manufacturing. The main material parameter to be optimized to achieve the best operating conditions is the carrier mobility. At mobilities >10,000 cm2 V−1 s−1, a carrier concentration ~1012 cm−2 would ensure competitive modulation, detection and switching performances.
High mobility can be reached by using single crystals, an optimized transfer process and/or encapsulation. These aspects need to be combined with an optimized process to minimize the contact resistance. The remaining steps in the fabrication of graphene photonics coincide with those used for Si photonics integrated circuits. Given this, graphene photonics will offer a unique evolutionary pathway for photonics integration, with no technological discontinuity with respect to the existing and well-developed technologies.
We note that graphene is only one of thousands of possible layered materials49. In particular, transition metal dichalcogenides have a strong light–matter interaction and nonlinear optical effects. They could also be exploited for graphene encapsulation, instead of BN, to further increase graphene’s mobility. These materials have been used as detectors53 at non-telecom wavelengths and could also be used to make phase and electro-absorption modulators, as well as switches. However, their development is not yet at the stage of graphene-based devices.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Ethernet Alliance: http://www.ethernetalliance.org/
Information and Communication Technologies, Environmental Sustainability and Climate Change: http://www.itu.int/en/action/climate/Pages/default.aspx
International Technology Roadmap for Semiconductors 2.0, 2013 Edition, Design: http://www.itrs2.net/2013-itrs.html
International Telecommunications Union: Interfaces for the optical transport network: https://www.itu.int/rec/T-REC-G.709/en
International Telecommunications Union: Support of IP-based services using IP transfer capabilities: https://www.itu.int/rec/T-REC-Y.1241/en
Internet of Everything: https://newsroom.cisco.com/ioe
New 2018 Ethernet roadmap looks to future speeds of 1.6 terabits/s. inside HPC: https://insidehpc.com/2018/03/new-2018-ethernet-roadmap-looks-future-speeds-1-6-terabits-s/
Photonics Component & Circuit Design Software: https://www.lumerical.com/tcad-products/
Si-on-insulator (SOI): https://order.universitywafer.com/default.aspx?cat=Silicon%20on%20Insulator%20(SOI)%20wafers
The ethernet roadmap: https://newsroom.cisco.com/ioe
This work was conceived within the Graphene Flagship project. The authors acknowledge funding from the European Union H2020 Graphene Project, European Research Council (ERC) Grant Hetero2D and Engineering and Physical Sciences Research Council (EPSRC) grant nos. EP/509 K01711X/1, EP/K017144/1, EP/ N010345/1, EP/M507799/5101 and EP/L016087/1.