1. Introduction

Recently, two-dimensional (2D) materials have been very attractive and become popular for the extraordinary properties [1–3], e.g., high carrier mobility [4] and broadband light absorption [3]. As one of the most popular 2D materials, graphene has a carrier mobility up to ∼200,000 cmV^-1s^-1 [4] and can be integrated easily onto a silicon photonic waveguide for achieving efficient light absorption [5,6]. Currently, there have been several hybrid silicon-graphene waveguide photodetectors reported [7–18]. Among them, the metal-graphene-metal (MGM) configuration [7–15] is one of the most popular options, and high-speed operation has been realized owing to high carrier mobility and short carrier lifetime in graphene. For example, the 3 dB-bandwidth for MGM hybrid silicon/graphene waveguide photodetectors can be as high as >40 GHz by utilizing the bolometric (BOL) effect [8], the photoconductive (PC) effect [12], the photovoltaic (PV) effect [14], and the photo-thermoelectric (PTE) effect [13,15].

In order to realize high performances for photodetectors, it is very important to achieve high-efficiency light absorption in the active region. When light absorption is enhanced, the size of the active region can be shrunk, which is helpful to improve the RC-limited 3 dB bandwidth, reduce the dark current, improve the signal-noise-ratio (SNR) and lower the power consumption [19]. For mono-layer graphene with single atomic thickness of 0.34 nm, the light absorption is possibly sufficient by choosing a sufficiently long absorption region (∼10² µm) when using a regular 220-nm-thick silicon photonic waveguide. In order to shrink the absorption region, one can enhance the light absorption in graphene by using photonic-crystal waveguides [20,21] as well as microring resonators (MRRs) [22], for which however the optical absorption is wavelength-selective and thus it is not appropriate for the broadband applications. In order to enhance broadband light absorption in graphene, some special waveguide structures have been reported previously, such as silicon nano-slot waveguides [23] and silicon hybrid plasmonic waveguides [8,12,14], which help improve the photodetector responsivity greatly.

In this paper, we propose and realize a compact and high-speed hybrid silicon/graphene photodetector by utilizing an ultra-thin silicon photonic waveguide integrated with a loop mirror. For example, here the thickness of the silicon core is chosen as small as 50 nm (i.e., h_co=50 nm). As it might be noticed, currently the use of ultrathin silicon waveguides is not a popular option for most foundry processes. On the other hand, there is no technical issue for the fabrication because only the silicon thickness and the etching depth changes. When using a 220-nm-thick SOI wafer for making the ultra-thin silicon photonic waveguides, an additional process of thermal oxidation or dry etching is indeed required to thin the top-silicon layer, which increases the cost and the complexity in some degree. Fortunately, such kind of process is regular and convenient for most foundry processes. More importantly, there are several unique advantages for ultra-thin silicon waveguides. When using an ultrathin silicon waveguide, the shallow strip greatly helps avoid the damage of graphene at the sidewall edges. An ultrathin silicon waveguide also has a much lower propagation loss due to less sidewall scattering than the regular 220nm-thick silicon waveguide [24,25] and provides some benefits for passive silicon photonic devices [25,26]. More importantly, the fundamental mode for an ultrathin silicon waveguide is less confined and the evanescent field is enhanced. In this way, the interaction between the evanescent field and the graphene sheet is improved significantly and thus the graphene absorption is enhanced greatly. Meanwhile, it is noted that the footprint of the passive photonic devices based on ultrathin silicon waveguides would be larger because of the weaker light confinement when there are some bent sections. Fortunately, an ultrathin silicon photonic waveguide can easily be connected with another silicon photonic waveguides with a different thickness (e.g., 220 nm) by using bi-level tapers [24], in which way compact photonic devices can be realized with thick silicon waveguides. Besides, in this paper an on-chip loop mirror is introduced at the end of the absorption region to reflect the residual light back to the graphene absorption region, which helps further enhance the graphene absorption. Note that our approach of using the ultrathin waveguide with a loop mirror can work compatibly together with some other excellent structure (e.g., nano-slot waveguides [23]), which makes it possible to improve the device further. In our previous conference paper [27], we proposed the concept of silicon/graphene photodetectors using an ultra-thin silicon photonic waveguide with a loop mirror. In this paper, we present the comprehensive results in details, including the theoretical analysis and the experimental characterizations. For the present design, the graphene absorption rate is improved by ∼5 times compared to that in the conventional hybrid silicon/graphene waveguides with h_co=220 nm. For the fabricated photodetectors, when operating at 1.55 µm, the measured 3dB-bandwidth is about 17 GHz at zero bias, while the measured responsivities are ∼1.4 mA/W under zero bias and ∼25 mA/W under 0.3 V bias.

2. Structure and design

Figure 1(a) shows the schematic configuration of the proposed hybrid silicon/graphene waveguide photodetector, which is based on an ultra-thin silicon waveguide. Here the loop mirror is introduced at the end of the absorption region to reflect the residual light back to the graphene absorption region. Figure 1(b) shows the cross-section of the active region of the present graphene photodetector, and Fig. 1(c) shows the mode profile. Here a 5 nm-thick SiO₂ thin layer was thermally formed to cover the 50 nm-thick silicon-core region. Then, a mono-layer graphene was placed on the top by using the wet-transferring process. Since the silicon core is very thin, it is important to introduce an upper-cladding to make the waveguide quasi-symmetric in the vertical direction. Otherwise, the fundamental mode in an ultrathin silicon waveguide might be cut-off due to the vertical asymmetry introduced by the air upper-cladding according to the guided-mode theory. Therefore, we used a PMMA upper-cladding for the fabrication convenience, which can definitely be replaced by the regular silica. On the other hand, high-speed photodetector cannot be achieved if the PMMA upper-cladding contacts with the graphene sheet directly because of the photogating effect [28]. In order to insulate the PMMA layer and the graphene sheet, a 10nm-thick CVD hBN layer was introduced between them. This hBN layer does not introduce any notable distortion of the mode field in the hybrid silicon-graphene waveguide. In order to achieve reduced graphene-metal contact resistances, which helps improve the RC-limited 3 dB-bandwidth, here we introduce the metal-graphene-metal sandwiched structure for the ground-electrodes. With such a sandwiched ground-electrode, the graphene-metal contact resistance can be reduced at least 40% because of the enhancement of the effective graphene-metal coupling, as well as the improvement of the graphene doping in the sandwiched structure, as described in [29]. The distance between the metal electrode and the waveguide edge is L_d = 0.9 µm to avoid the metal absorption loss.

 

Fig. 1. (a) Three-dimensional schematic configuration of the proposed hybrid silicon-graphene photodetector based on an ultra-thin silicon waveguide with a loop mirror reflector; (b) Cross-section of the hybrid silicon-graphene waveguide; (c) Mode profile of the present ultrathin silicon waveguide.

Download Full Size | PDF

Figure 2(a) shows the simulated graphene absorption coefficient α_G in the present hybrid silicon-graphene waveguide with different silicon-core width w_co (varying from 0.6 µm to 1.4 µm) as the silicon-core thickness h_co varies from 50 nm to 220 nm. Here we assume that all the materials except graphene are lossless, which is reasonable. As a result, the graphene absorption rate is given by α_G=8.68 n_{eff_im} (2π/λ₀), where n_{eff_im} is the imaginary part of the effective index for the guided mode. From this figure, it can be seen that there is an optimal thickness h_{co_opt} for the silicon core to achieve the maximal graphene absorption coefficient for any given silicon-core width.

In order to understand why there is maximal absorption, we show the normalized modal-field profiles for different core-thicknesses (i.e., h_co=50, 70, 150, 220 nm) when the waveguide width is fixed as w_co=1 µm, as shown in Fig. 2(b). It can be seen that their peak powers are different. In order to see more clearly, Fig. 2(c) shows the calculated absorption intensity along the graphene sheet [see the red line in Fig. 2(b)]. The graphene absorption intensity A_g(l) is given by ${A_\textrm{g}}(l )= \frac{1}{2}\textrm{Real}({{\sigma_\textrm{g}}} ){|{{{\vec{E}}_\textrm{t}}(l )} |^2}$(W/m²), where Real(σ_g) is the real part of the graphene conductivity, $\overrightarrow {{E_\textrm{t}}} $ is the transverse component of the electric fields along the graphene surface of the launched waveguide mode [12]. When the silicon-core thickness h_co decreases from e.g. 220 nm, the confinement becomes weak and the evanescent field in graphene becomes strong. As shown in Fig. 2(c), the graphene absorption intensity is increased dramatically at the center of the graphene sheet and at edges of the waveguide. Consequently, the graphene absorption coefficient increases greatly. For example, the graphene absorption coefficient α_G for the case of w_co=1.0 µm increases from 0.07 dB/µm to ∼0.19 dB/µm when the silicon-core thickness h_co decreases from 220 nm to 70 nm. On the other hand, when the thickness h_co decreases, the mode spot size increases. Therefore, the optical mode confinement becomes very weak and thus the mode size becomes very large when the thickness is reduced to be very thin (e.g., several tens of nanometers). In this case, it can be seen that the field intensity at the center of the graphene layer becomes weak when the thickness h_co further decreases from 70 nm to 50 nm. Meanwhile, the absorption intensity at the edge of the waveguide is increased a little. As a result, the graphene absorption decreases slightly as the silicon-core thickness further decreases to be h_co<h_{co_opt}. Since a shallow silicon core is preferred very much to avoid any damage of graphene at the sidewall edges, in this paper we choose the silicon-core thickness as h_co=50 nm, and the silicon-core width is chosen as w_co=1.0 µm. Accordingly, the graphene absorption coefficient α_G for the fundamental mode of TE polarization is about 0.182 dB/µm, which is improved by 2.5 times compared to the regular hybrid silicon-graphene waveguide with h_co=220 nm. For this comparison, we choose the same core-width as well as the same upper-cladding material (PMMA) for them.

 

Fig. 2. (a) Calculated graphene absorption coefficient in the proposed hybrid silicon-graphene as the silicon-core thickness h_co varies. Insets: the fundamental mode of TE polarization for the cases with different thicknesses; (b) Calculated modal-field profiles and (c) the absorption along the graphene sheet for the cases with different core-thicknesses (h_co=50, 70, 150, 220 nm). Here the mode power is normalized to be 1 mW, and the width of the waveguide is 1µm.

Download Full Size | PDF

In order to further enhance the graphene absorption, we introduce a loop mirror at the end of the hybrid silicon-graphene waveguide, so that the residual light can be reflected back to be absorbed in the active region. Figure 3(a) shows the schematic configuration of the loop mirror, which consists of a broadband Y-branch for splitting optical power into two waveguides equally. The bending radius is chosen as R=40µm to guarantee low bending loss (<0.001 dB) and we choose θ=60° in the design. Figure 3(b) shows the calculated reflection of the designed loop mirror, and the inset shows the simulated light propagation in the designed Y-branch. For this Y-branch, the excess loss is about 0.3 dB and the power splitting ratio is 50%: 50% in an ultra-broad bandwidth. Accordingly, the loop mirror works well in a broad bandwidth from 1500 nm to 1600 nm with a total excess loss of ∼0.6 dB. With the loop mirror, the graphene absorption coefficient is doubled equivalently, i.e., α_G=0.36 dB/µm, which is improved by 5 times compared to the 220-nm-thick hybrid silicon-graphene waveguide without a loop mirror. As a result, the length of the absorption active region can be shortened by half. With the present design, one can achieve more than 80% optical power absorption approximately when choosing a 20 µm-long graphene absorption region. In contrast, for those hybrid plasmonic waveguide photodetectors reported previously, the graphene absorptance is 34%-44% only [8,30], which is much lower than that obtained in this paper. One should realize that there are some advantages when the absorption region length L is shortened. First, in MGM-type graphene photodetectors, most power consumption comes from the dark current I_d. When the graphene absorption region is shorter, the graphene-sheet resistance and the contact resistance become larger, and thus the dark current is lower. Consequently, the power consumption P is reduced according to the formula P = I_dV_b. Second, a shorter graphene absorption region leads to a lower capacitance C, which is proportional to the length L. This is helpful to increase the RC-limited 3-dB-bandwidth, even though the total resistance R of the graphene photodetector increases because it is inversely proportional to the length L. This can be explained quickly according to the simplified equivalent circuit-model for the photodetector [31]. Third, a shorter graphene absorption region is helpful to lower the dark current and thus the noise level according to the theory [32]. For a photodetector, the noise i_n mainly comes from the thermal noise i_nJ and the dark-current shot noise i_nd, which are both inversely proportional to R_g^1/2 [32]. As a result, the dark current and the noise level can be reduced by reducing the length L.

 

Fig. 3. (a) Schematic configuration of a loop mirror; (b) Calculated reflection of the designed loop mirror in the band of 1500-1600 nm. Inset: simulated light propagation in the designed Y-branch for the loop mirror.

Download Full Size | PDF

3. Experiments and discussions

The designed waveguide photodetectors were fabricated with a series of steps. The silicon top-layer of a silicon-on-insulator (SOI) wafer was thinned from 220 nm to 50 nm by using the processes of thermal oxidization and buffered-oxide-etch (BOE) wet-etching. The ultra-thin silicon-core strips and the grating couplers were fabricated by the processes of electron-beam lithography (EBL) as well as the ICP etching. After that, 10nm-Ti/50nm-Au electrodes were formed by using the electron-beam evaporation technique. Then a CVD-grown monolayer graphene sheet was transferred on the silicon core layer by using the wet transfer method and then patterned by using the O₂ plasma etching process. Another 50-nm-thick Au thin-film was deposited on the graphene sheet as the top metal-layer to form the metal-graphene-metal sandwiched structure for the ground-electrodes, which helps reduce the metal-graphene contact resistance. A 10 nm-thick CVD hBN thin film was transferred on the top of the graphene sheet to insulate the graphene sheet and the PMMA upper-cladding. Finally, a PMMA upper-cladding was formed by using a spin-coating process, so that the TE₀ mode can be well supported in the optical waveguides. Figure 4 shows the microscope images of the fabricated hybrid silicon-graphene waveguide photodetectors. There are two photodetectors on the same chip, i.e., samples S₁ and S₂. For Sample S₁, there is a 20 µm-long graphene absorption region (i.e., L=20 µm) as well as a loop mirror, as shown in Figs. 4(a) and 4(b). In order to give a comparison, sample S₂ has a 40 µm-long graphene absorption region (i.e., L=40 µm) and there is no loop mirror, as shown in Figs. 4(c) and 4(d). The measured graphene absorption rate is about 0.185 dB/µm for sample S₂ and correspondingly sample S_1­ has α_G=∼0.36 dB/µm, which is consistent with the theoretical prediction.

 

Fig. 4. Microscopy images for the fabricated graphene photodetectors (e.g., samples S₁ and S₂). (a) Sample S₁ with a loop mirror; (b) an enlarged view for the graphene absorption region of sample S₁; (c) Sample S₂ without a loop mirror; (d) an enlarged view for the graphene absorption region of sample S₂.

Download Full Size | PDF

In order to characterize the static performance for samples S₁ and S₂, a low-frequency measurement setup was established, as shown in Fig. 5. Here the 1550 nm laser was modulated with a low frequency of 1 kHz, generated easily by using a tunable laser (e.g., HP 8163A). The modulated light was then coupled to the chip through a polarization controller and the input grating coupler working for TE polarization. The bias voltage was applied by a pre-amplifier (SR570), while the electrical signal was received by using a lock-in amplifier (SR830) with the help of the reference clock signal from the laser. The measured voltage from the lock-in amplifier was used to evaluate the responsivity of the photodetectors.

 

Fig. 5. Setup for characterizing the fabricated graphene photodetectors.

Download Full Size | PDF

Figure 6(a) show the measured photocurrent under different optical powers as the bias voltage V_b varies from −0.3 V to 0.3 V for sample S₁. It can be seen that the photocurrent is negative relative to V_b and increases linearly as V_b increases. Figure 6(b) shows the measured photocurrent and responsivity for the cases with different optical powers when V_b=0.3 V. It can be seen that the photocurrent increases linearly with the optical power when operating at a fixed bias voltage. The responsivity is about 22.2-24.7 mA/W as the input optical power varies from 0.48 mW to 0.12 mW. Figure 6(c) shows the responsivity for sample S₁ as the bias voltage V_b varies from −0.3 V to 0.3 V. When operating with a positive bias voltage (i.e., V_b=0.3 V), the photocurrent is negative, which is a typical property for the bolometric effect in graphene photodetectors [33]. For the bolometric effect, the local temperature for the graphene sheet increases as the incident optical power increases, and thus the graphene sheet resistance is changed. The bolometric effect exists commonly in graphene photodetectors [33,34] because no gradient doping is needed, which is different from the photovoltaic or thermoelectric effect [35]. When operating at zero bias (i.e., V_b=0), the responsivity is about 1.4 mA/W. In this case, there is no in-plane electric field in the absorption region, and thus the photo-voltage effect is not to be the dominant one. Instead, the photoresponse is mainly from the photo-thermoelectric (PTE) effect. In the overlay fabrication process, the fabrication error introduces some asymmetry for the metal electrodes due to the misalignment (e.g., ∼100 nm). Consequently, the asymmetry of the metal electrodes introduces some electronic temperature gradient, and thus the PTE effect was observed in the photodetector with homogeneous graphene [36]. As a comparison, sample S₂ was also characterized and the measured photoresponse is shown in Fig. 6(d). It can be seen that the responsivity is about 22.5-26.7 mA/W at V_b=0.3 V with different input optical powers. Sample S₂ has a slightly higher responsivity than sample S₁ since the Y-branch used for sample S₁ introduces a little excess loss. From the comparison between samples S₁ and S₂, it can be seen that the reflector at the rear indeed helps improve the light absorption in graphene even when the active region is reduced by half.

 

Fig. 6. Experimental results of the ultrathin silicon/graphene waveguide photodetectors. (a) Measured photocurrent as the bias voltage varies from −0.3 V to 0.3 V with different optical powers. (b) Measured photocurrent and responsivity for sample S₁ operating with 0.3 V bias voltage and different optical powers. Measured responsivity for sample S₁ (c) and sample S₂ (d) operating with a bias voltage varying from −0.3 V to 0.3 V and different optical powers.

Download Full Size | PDF

The frequency responses of samples S₁ and S₂ are also characterized by using the setup shown in Fig. 7(a). In the experiments, light from a tunable laser was modulated by a high-speed Mech-Zehnder modulator and then amplified by an EDFA (Thorlabs EDFA100P). The signal was then coupled into the chip through an TE-type grating coupler, which has a coupling loss of about 10.5 dB. The grating coupler loss is much higher than the theoretical simulation result because of the dimension deviation and some contamination introduced in the process. It is actually possible to achieve efficient grating couplers even for ultrathin silicon waveguides. For example, the loss is about 3.7 dB for the grating coupler demonstrated in [37]. When it is desired to further improve the coupling efficiency, an ultrathin silicon photonic waveguide can easily be connected with a mature grating coupler based on a silicon photonic waveguide with a different thickness (e.g., h_co=220 nm) by using bi-level tapers [24]. When the silicon-graphene waveguide photodetector received the signal, the generated electrical signal was amplified by using a RF amplifier (Centellax OA4MVM2) and finally analyzed by a vector network analyzer (VNA). A careful calibration was performed by using a commercial high-speed photodetector.

 

Fig. 7. High frequency characterization of the ultrathin waveguide graphene photodetectors. (a) Experimental setup for the high-frequency measurement. (b) Normalized response of S₂₁ for sample S₁ and sample S₂ under the same condition. Here V_b=0. Inset: the equivalent circuit model for graphene photodetector. (c) Measured and fitted results for the reflection S₁₁ for sample S₁. Inset, calculated response S₂₁ from the equivalent circuit model.

Download Full Size | PDF

Figure 7(b) shows the measured results of the normalized frequency responses under zero bias for samples S₁ and S₂ operating in the same condition. Here the input optical power is P_in=∼3 mW and the operation wavelength is λ=1550 nm. In addition, the S₂₁ response of sample S₁ has less ripples than that of sample S₂ due to the lower noise level. As it is well known, the noise i_n mainly comes from the thermal noise i_nJ and the dark-current shot noise i_nd. Both of them are inversely proportional to R_g^1/2 [32], where R_g is the total resistance of the graphene photodetector. Since the graphene absorption region length in sample S₁ is reduced by half compared to sample S₂, the resistance R_g of sample S₁ is doubled accordingly. As a result, sample S₁ has a lower noise level than sample S₂ when operating in the same condition. One has i_n1/i_n2=$1/\sqrt 2 $, where i_n1 and i_n2 are the noise levels for sample S₁ and S₂. For photodetectors, the noise equivalent power (NEP) is given by NEP = i_n/R_P (where R_P is the responsivity). Accordingly, the NEP of sample S₁ is $1/\sqrt 2 $ lower than that of sample S₂. In the inset of Fig. 7(b), the established equivalent circuit model [31,38] is shown. Here, R_s is the series resistance (mainly contributed by the contact resistance), C_Sub is the substrate capacitance, R_Sub is the substrate resistance, and C_air is the capacitance between the metal electrodes. In order to further explore the limitations of the photodetector speed, we tested the S₁₁ parameters of two samples (S₁ and S₂) and fitted the curves to extract the parameters for the equivalent circuit model. Figure 7(c) shows the measured and fitted results for S₁₁. These RCL parameters extracted from the measured reflection coefficient S₁₁ are shown in Table 1.

Table 1. Parameters for the equivalent circuit extracted from the measured S₁₁ and the 3dB-bandwdith BW_3dB estimated from the equivalent circuit model.

View Table

According to the established circuit model with these extracted RCL parameters, one can estimate the RC-limited 3 dB-bandwidth. For the present case, the RC-limited 3 dB-bandwidths for samples S₁ and S₂ are estimated to be about 118 GHz and 95.9 GHz, respectively. On the other hand, the RC-limited 3 dB-bandwidth is much higher than the measured results shown in Fig. 7(b). It indicates that the bandwidth of the present photodetectors is limited due to some other reasons. According the results reported previously, the 3dB-bandwidth is also often limited by the transit time of carriers [39]. When assuming the saturation velocity is ∼10⁵ m/s for the carrier in graphene on a SiO₂ insulator [40], the calculated transit-time-limited bandwidth for our structure is about 20 GHz, which is close to the measured result. The reduction of the saturation velocity might be due to the phonon scattering induced by the SiO₂ insulator layer [41] and the slowed carrier transport speed caused by the wrinkles at the CVD hBN-graphene interface [42]. Furthermore, the wet-transferring process might introduce some contamination from the PMMA residue and the copper etchant, which also lowers carrier mobility and prevents ultrafast responses [43]. It is possible to shorten the carrier transit time by reducing graphene channel length and improving the fabrication process as well as introducing a high-quality insulator layer of e.g. Al₂O₃. Definitely the present photodetectors should be improved further with more efforts when compared to Ge/Si photodetectors, which are now well mature with high responsivity and high speed when operating in the wavelength-band of 1.3/1.5 µm. Note that the realization of Ge/Si photodetectors usually requires expensive and complicated fabrication processes [44], such as the ion implantation, the thermal annealing, and the CMP for the Ge layer. It is also not easy to fabricate Ge photodetectors on some other popular optical waveguide systems with e.g. SiN [45], LiNbO₃ [46], etc. In contrast, graphene can be combined easily with these optical waveguides for realizing waveguide photodetectors as well as other active devices. Besides, regarding that light absorption in Ge becomes weak for the wavelength longer than 1.55 µm and special process is needed to cover the L-band (1.56-1.62 µm) [47] due to the bandgap limitation, graphene is a promising option for realizing photodetection beyond 1.55 µm, which will be very useful for many applications in potential.

4. Conclusion

As a summary, in this paper we have demonstrated compact and high-speed hybrid silicon/graphene photodetectors by introducing an ultra-thin silicon waveguide integrated with a loop mirror. For the present design, the ultrathin silicon core really helps avoid the damage of the graphene sheet at the sidewall edges. This robustness is really important for further developing photonic integration. It has been shown that the graphene absorption rate can be greatly enhanced by ∼5 times for TE polarization because of the evanescent field enhancement of an ultrathin silicon waveguide and the feedback of the on-chip loop mirror at the rear. As a result, the graphene absorption region can be shortened greatly, which is helpful to improve the RC-limited 3-dB bandwidth, lower the dark current, reduce the power consumption, lower the noise level, and improve the signal-noise-ratio (SNR). The present approach of using the ultrathin waveguide with a loop mirror can be extended to work compatibly together with some other excellent structure (e.g., nano-slot waveguides [23]), and it is expected to achieve photodetectors with further enhanced light absorption. A possible issue is that light reflected by the loop mirror is not absorbed completely if the absorption region is not sufficiently long. Note that it is often desired to have light absorption of at least 90% for realize high-responsivity photodetectors. Therefore, the power reflected back from the photodetector to the input side is usually low, and thus the influence should usually be insignificant to the other components (e.g., laser sources) in the link. Particularly, there is usually an isolator after a laser source to isolate the reflection. Another possible issue related to the reflection is that the reflected signal interferes with the next incoming bit. Fortunately, the time delay between the forward- and backward-propagating lightwaves in the present design is less than 2 ps, estimated by assuming the delay is ∼400 µm long (which can be reduced further). Such a tiny time-delay can be negligible even when the bit rate is as high as 100 Gbps. For the fabricated photodetectors, the measured responsivities are ∼1.4 mA/W under zero bias and ∼25 mA/W under a 0.3V bias voltage when operating at 1.55 µm. The measured 3dB-bandwidth is about 17 GHz even under zero bias. The measured and calculated results indicate that the bandwidths of the present photodetectors are possibly limited by the carrier transit time. It is expected to achieve very high bandwidth by introducing a high-quality insulator layer of e.g. Al₂O₃ and using a narrow graphene channel length. Graphene hetero-structure can also be introduced to replace the monolayer graphene sheet used in the present photodetectors, which will be helpful for a further dark current suppression.