Results – Device Fabrication and characterization
Fig.3a shows a schematic of the G-OEM. It consists in a CPW in ground-signal-ground (GSG) configuration, operating in the sub-THz range (see Supplementary Information, Section 1B). The electrical CPW embeds at the center of the signal electrode (S) an hBN-encapsulated single crystal graphene layer, transferred on top of a SiN photonic waveguide. Fig.3b shows a schematic detail of the active area of the device (top and section view), indicated with a dashed circle in Fig.3a. The graphene bottom layer (graphene channel) has dimensions L=4 mm and W=50mm . A second graphene layer acts as electrostatic gate on top of the hBN-encapsulated graphene and is contacted with a metallic electrode (indicated as "Gate electrode" in Fig.3a-b). The graphene gate is used to set the operating point of the device: the chemical potential mc of the bottom graphene layer must be set far from the Charge Neutrality Point (CNP), i.e. mc ~ 150 meV for optimal operation, since at such electrostatic doping the photo-bolometric effect, and so the phoresponsivity, is maximized43(see Supplementary Information, Section 1D-E).
The hBN/graphene/hBN/graphene stack is placed on top of a transverse electric (TE) SiN photonic waveguide of 1500 nm width (WGw) and 270 nm thickness (WGt), allowing single mode operation at 1550 nm (see Supplementary Information, Section 1C). The waveguide is covered by a BPTEOS film used as top cladding. After deposition, the cladding is selectively thinned down to 50 nm on top of the waveguide in the region where the hBN/graphene/hBN/graphene stack is transferred. In this way, the optical waveguide TE mode is evanescently coupled to the graphene layer placed on top of the waveguide. The obtained effective absorption in the graphene active layer is ~0.57 dB/um (see Supplementary Information, Section 1C). We used graphene as top gate to reduce the optical power losses compared to a metallic gate. From simulations, we estimated < 3 dB total optical power loss due to absorption in the graphene gate and metal contacts (see Supplementary Information, Section 1C).
An adiabatic inverted taper has been designed to access the photonic waveguide through butt-coupling via a lensed single mode fiber with mode field diameter ~3 um (Supplementary Information, Section 1C). Fig.3c shows an optical microscopy image of the device. The full fabrication flow is presented in details in Fig. S9 of Supplementary Information, Section 2. We use a van der Waals 'pick-and-flip' technique50 to expose a CVD-grown graphene crystal (previously transferred from Cu to SiO2/Si36,51 ) on top of a ~30 nm thick hBN flake (mechanically exfoliated from bulk crystals). After depositing Ni/Au (7/60 nm) top contacts by thermal evaporation37, we place a second graphene/hBN stack (obtained via the same pick-and-flip procedure) onto the first one, realizing both full encapsulation of the device channel and electrical-isolation of the top graphene layer.
Halfway through the assembly, we use standard scanning microRaman spectroscopy to evaluate the quality of the hBN-supported graphene channel. Fig.3d shows a representative Raman spectrum from a spatial mapping over the active area (50×4 mm2). We observe the typical spectral features of high-quality single-layer graphene53,54, such as narrow Lorentzian 2D peak and large intensity ratio between the 2D and G peaks, in addition to a sharp E2g peak from the hBN flake. As a standard indicator for the quality of graphene/hBN stack55, we analyze the full width at half maximum (FWHM) of the 2D peak, obtaining a statistical distribution peaked at 19.5 cm-1 (see Fig.3d inset). The decisive improvement over the response of nominally identical graphene crystals placed on conventional substrates (~23 cm-1 on SiO2/Si in Ref.37) points toward a reduction of strain fluctuations56, which is known to crucially enhance the carrier mobility57,58.
We measured the two-terminal resistance (R) of the sample as a function of the top gate potential (Vtg), at room temperature and in air. Fig.3e shows the typical ambipolar behavior of graphene field-effect transistors, with a CNP at Vtg ~ 0.5 V. We estimate low doping level of n ~ 3×1011 cm-2 at Vtg = 0 V, (assuming 30 nm hBN dielectric with out-of-plane relative permittivity εr = 3). Using the standard fitting procedure from Ref.59 (light blue lines in Fig.3e) we obtain a electron (hole) mobility of 27000 cm2V-1s-1 (23000 cm2V-1s-1) and a contact resistance of 1.7 k W ∙mm (1.2 k W ∙mm). The clear asymmetry of the field-effect curve indicates a contribution from gate-dependent contact resistance60,61, which likely affects the mobility estimate (proper four-probe measurements are however not possible in the current device design). Finally, the magnitude of charge inhomogeneity in the CNP region, estimated via the procedure in Fig.3e inset57, falls well below the 1011 cm-2 range, further confirming the expected high device quality57. Fig. S6 of the Supplementary Information shows that the achievement of these results is promoted by post-fabrication thermal annealing of the device in inert atmosphere.
The operating principle of the G-OEM is detailed in Supplementary Information, Section 1A,D-E. The basic concept allowing optoelectronic frequency mixing relies on the fundamental physical mechanism of hot carrier generation in a graphene layer coupled to an optical field. The chemical potential (mc) is set far from the CNP (mc> 150-200 meV, depending on charge carrier mobility43). At this electrostatic doping, the dominant effect allowing photodetection is carrier heating, which translates into a decrease of the carrier mobility, i.e. into a decrease of graphene electrical conductivity42,43. The photogenerated carriers' cooling dynamics are very fast in graphene, with measured relaxation times of ~2ps45. This means that graphene conductivity can be modulated at very high frequencies (up to ~500 GHz) by means of an optical field. Therefore, as shown in Fig.3a, to obtain optoelectronic mixing a baseband electrical signal (intermediate frequency (IF) signal, few GHz bandwidth) is coupled to the CPW, and two optical wavelengths separated by the desired frequency (LO frequency) are coupled to the graphene channel. The optical LO induces a time-varying conductivity of the graphene channel in the sub-THz range (see Supplementary Information, Section 1D). When the baseband electrical signal passes through the portion of the CPW exhibiting time-varying conductivity (i.e. where the graphene channel is present), it is modulated at the frequency of the LO, i.e., it is upconverted. Graphene photo-bolometers can be used as PDs by applying a DC bias voltage across the graphene channel62. This produces a large dark current in the mA range62. We stress that in our case, i.e, when a graphene bolometer is operated as OEM, no DC bias voltage is needed across the graphene channel41(see Supplementary Information, Section 1D).
The fabricated OEM was first characterized in terms of photodetection bandwidth, i.e. to determine the maximum LO frequency that can be generated by the device. As detailed in Supplementary Information, Section 3, we used two different experimental setup. The first allowed to characterize the device response up to 67 GHz using a Vector-Network-Analyzer (Keysight PNA-X 5247B). We then characterized the response in the range 92-96 GHz using a calibrated commercial electronic downconverter (Mi-Wave 970W-94/387) which downconverts high-frequency electrical signals from the 92-96 GHz range down to the 4-8 GHz range. We used a 44 GHz bandwidth electrical Spectrum Analyzer (Anritzu MS2850A) to measure the downconverted electrical signal. The result is shown in Fig.4a and reveals a frequency response higher than 96 GHz, since no roll-of was measured up to this value. We then characterized the G-OEM as a mm-wave upconverter using the same aforementioned receiver (Mi-Wave 970W-94/387). The experimental setup is detailed in Supplementary Information, Section 4. Fig.4b reports the upconversion of an IF signal that has been mixed with the optical LO. The IF was swept between 1 and 5 GHz, and the LO was kept at the fixed frequency of 91 GHz. The 92-96 GHz frequency window was set by the downconverter working range. The IF input power was 3 dBm, while the input optical power was 13 dBm. The measured upconverted power has an average of ~-41 dBm corresponding to an upconversion efficiency of Peff[dB]=Pout[dBm]-Pin[dBm]~-44dB (see Supplementary Information, Section 2E and Section 4). Finally, the inset in Fig.4c shows the characterization of the linearity of the G-OEM versus the input electrical power. The measurement was performed by sweeping the input power of a 3GHz sinusoidal wave that was mixed with a 91 GHz, 13 dBm optical LO. The response was highly linear for electrical input powers up to 1 dBm.
Results – sub-THz wireless link
Fig.5a shows the wireless transmission experimental setup. A Mach-Zehnder modulator (MZM) with bandwidth of 40 GHz was operated in double sideband – suppressed carrier (DSB-SC) mode63 and modulated with a sinusoidal signal of 45.5 GHz. The optical input of the MZM was a single continuous wavelength (CW) coming from a distributed-feedback (DBF) laser. The modulated output optical signal was then filtered and amplified to obtain two phase locked optical wavelengths, separated in frequency by 91 GHz. The central carrier and high order harmonics were suppressed by >30 dB (see Supplementary Information, Section 3). The optical signal was coupled to the integrated graphene OEM through a lensed fiber butt coupled to the photonic chip. A quadrature phase shift keying (QPSK) baseband signal with CF of 2 GHz was generated starting from two pseudo-random binary sequencies (PRBS), generated using a 65 GSample/s digital-to-analog converter (DAC) (Fujitzu LEIA 55-65GSa/s 8-bit DAC), delivering a digital signal with maximum total length of 2^17 samples. The data-rate of the baseband signal was swept in the range 1-4 Gbit/s. The signal was fed to the G-OEM through a GSG RF probe (MPI T110A-GSG100) with bandwidth 110 GHz ("IF" port, in Fig.5a). The baseband signal was upconverted by the mixer and collected by a second probe (MPI T110A-GSG100) ("RF" port, in Fig.5a). The probe was connected through a 1-mm connector to short (~15 cm) RF electrical cable. A 1-mm to WR-10 waveguide transition was used to connect the system to a WR-10 electrical bandpass filter (mi-wave 460W-86/94/387) with 8 GHz bandwidth and central frequency 90 GHz. The filtered signal was then amplified with an amplifier (mi-wave 955WF-35) with 35 dB gain, 75-110 GHz bandwidth, and transmitted through a 2 meter wireless link using an horn lensed antenna (mi-wave 258W) with 30 dB gain and 4 GHz bandwidth, with central frequency 94 GHz. The receiver was composed by a second antenna (mi-wave 258W), connected to a commercial electronic downconverter (mi-wave 970W-94) allowing frequency downconversion from the 92-96 GHz (RF) range to the 4-8 GHz (IF) range. A real time oscilloscope (Agilent Infinium VSA80000A) with electrical bandwidth of 12 GHz was used to acquire the downconverted signal, and visualize the received QPSK data stream constellation using a built-in software included in the instrument. A picture of the setup is present in the Supplementary Information, Section 5. We optimized the Error-Vector-Magnitude (EVM) of the received signal by adjusting the working point of the graphene OEM.
We transmitted a 1Gbit/s QPSK baseband signal through the wireless link, and adjusted the gate voltage of the device. This is shown in Fig.5b. The inset shows the received constellation for a gate voltage VG~-1.8 V, which corresponds to a voltage of ~2.3V from the CNP voltage Vtg ~ 0.5 V (see Fig.3e), that is µc~130 meV. The measurement was performed for a LO power of ~13.3 dBm and with an input baseband electrical power of ~9dBm. These two are the values optimizing the EVM, as shown in Fig.5 c-d. A minimum EVM ~22% is reached. We then tested the EVM as a function of the data-rate of the input baseband QPSK signal. The results are shown in Fig.5e, where the retrieved constellation, together with the eye diagram of one of the two quadratures (Q quadrature) composing the QPSK signal are presented. The Error-Vector-Magnitude (EVM) for 1Gbit/s is 22%. This value increases up to 24% for the 2Gbit/s data stream and to 27% for the 4 Gbit/s signal. We estimated64 a bit error rate log(BER) of 4 ∙10-6, 1.3∙10-5 and 1.3∙10-4 at 1Gbit/s, 2 Gibit/s and 4 Gbit/s datarate, respectively.