1. Introduction

Quantum interconnects and networks, and all-optical quantum computing attempts, need such vast resources that their optical components are likely to be based on the technology of integrated optics. There are many material platforms for integrated quantum photonics [1]; among these, silicon is likely to yield the lowest cost, as long as the performance is good [2,3]. While some integrated microchip quantum key distribution (QKD) transmitters have been demonstrated, those based on silicon photonics rely on a separate external laser [4–8] whereas integrating the laser either requires advanced active-passive III–V semiconductor fabrication [9] or hybrid Si photonics [10], both of which are challenging and expensive. Co-packaging a laser with integrated optoelectronics is difficult and is one of the most expensive parts of optical transceiver technology [11].

Light with quantum properties can be generated using the nonlinear optical properties of materials such as silicon [12]. This report shows that it is possible to do so without a dedicated laser, using instead a fraction of a 10 Gbps NRZ data stream as an optical pump source for photon pairs and heralded single photons, which have applications in entanglement-based QKD, quantum repeaters, memories and quantum optical computing. There are millions of such channels in optical communications and data center networks, many of which can spare a bit of extra headroom, at least for brief durations. Our result shows that many of today’s optical networks may be able to conveniently generate quantum resources such as heralded single photons on a massive scale, without requiring purchase, installation and maintenance of dedicated lasers, which are among the most costly and least reliable components in an optical communications link. Thus, a significant step is made towards low-cost, massive-scale generation of entanglement by tapping into the existing resources of today’s fiber-optic networks. The key component here is the room-temperature integrated silicon microring resonator photon-pair generator, which has recently improved both brightness and purity significantly [13], achieving photon rates and heralded single-photon quality comparable with solid-state single-photon sources [14], which require cryogenic conditions for operation. The approach of tapping a small fraction of classical 10G data stream at standards-compliant power levels for photon-pair generation would not result in comparable rates of entangled photons if using optical fibers or traditional nonlinear crystal waveguides based on spontaneous parametric down conversion (SPDC).

Non-classical light generation in silicon integrated photonics uses the nonlinear optical process of spontaneous four-wave mixing (SFWM) to generate photon pairs [13,15–37]. Many of these demonstrations use microring resonators, which are attractive SFWM devices for a number of reasons, including compact size, simple fabrication, convenient monitoring and thermal tuning, low pump power requirement, and inbuilt spectral selectivity (i.e., non-resonant wavelengths in the bus waveguide bypass the device). In contrast to solid-state single-photon sources (quantum dots etc.), the brightness and single-photon purity of SPDC and SFWM sources are fundamentally inter-related, with the best photon purity obtained at low brightness. However, the raw performance numbers for single SPDC and SFWM source devices have steadily improved, and moreover, proposals for multiplexed sources [38] appear feasible with regard to wafer-scale manufacturing technology such as silicon photonics. The oulook is made optimistic by recent advances in improving the pair generation rate (R, the number of photon pairs generated per unit time per unit pump power) which have shown R ≃ 10 − 100 MHz.mW⁻², and therefore, photon pair fluxes between several hundred kilohertz and a few megahertz achieved by using much less than a milliwatt of pump power from a laser diode [31]. (In the classical domain, low power four-wave mixing has been shown for single and coupled silicon microring resonators [39–42]). However, researchers have not benefitted from these advantages yet, since the apparatus used to pump SFWM experiments using silicon photonics are basically the same as used in pair generation experiments for the last several decades.

External-cavity laser diodes and mode-locked lasers which are used in many SFWM experiments are large, expensive and power-hungry devices, compared to the microring resonator, or the small integrated-optics components inside a typical optical communications transceiver used in fiber-optic networks. Doped-silica [43], nitride [44], and poled-crystal waveguide [45] platforms for quantum photonics inherently lack an integrated pump-source technology and rely on external lasers. III–V semiconductors can be used as optical integration platform [46], but pair generation rates and quality are currently superior in silicon photonics technology and the manufacturing costs of the latter are likely to be lower as well. Recently, wafer-scale fabricated hybrid III–V-Si lasers have been used as the pump, which makes the implementation more compact and less expensive, but the fabrication is a specialized process and complete integration yet has to be achieved [47].

Here, we show that the light from a standard 10 Gigabits per second (Gbps) non-return-to-zero (NRZ) data stream can replace the role of the dedicated pump laser for SFWM-based pair generation. A fraction of a classical communications channel between two off-the-shelf SFP+ transceivers was tapped (i.e., a small fraction was diverted) and used for photon-pair generation in a silicon microring device (see Fig. 1). The average power level of the pump before the microring was −5.5 dBm, which is a small fraction of the typical 0 dBm power level used in a classical communication link. The untapped light is free to continue to establish the classical link.

 

Fig. 1 (a) A fraction of a 10 Gbps NRZ data stream is used to perform photon-pair generation using a silicon microring (b) Eye diagram of the classical data stream. (c) Microring transmission resonances showing the locations of the pump (P), signal (S) and idler (I) resonances used in this experiment. (d) The experimental setup for photon pair generation. (ASE: amplified spontaneous emission. TEC: thermo-electric controller)

Download Full Size | PDF

2. Experimental details

The two computers involved in the classical communication were PC workstations [see Fig. 1(a)], running a standard operating system, with plug-in PCI-e network interface cards (NICs) supporting plug-in 10 Gbps SFP+ transceivers with single-mode fiber (SMF) pigtails for optical communications near 1.55 μm wavelengths. The eye diagram of the classical channel is shown in Fig. 1(b) (receiver jitter has not been deconvolved) and exhibits a few tens of picoseconds of jitter, typical in transceivers compliant with IEEE 802.3 and similar standards. Off-the-shelf DWDM-grade SFP+ transceivers were used in both NICs, whose wavelengths were discretely tunable in steps of 50 GHz. The discrete wavelength step is much larger than the linewidth (approximately 2 GHz) of the silicon microring resonators used in this experiment, and therefore, the precise wavelength alignment was achieved by tuning the microring, not the transceiver. The microrings can be tuned thermally over several tens of nanometers in wavelength, so the precise wavelength of the transceiver is not important.

The microring resonator was fabricated using a foundry silicon photonic process on silicon-on-insulator (SOI) wafers, using ridge waveguides of width 0.65 μm, height 0.22 μm, and slab thickness 70 nm. This cross-section is designed for low-loss transmission in the lowest-order (i.e., fundamental) mode of the transverse electric (TE) polarization defined relative to the device plane. The microring had a radius R = 10 μm. The slab regions of the ridge silicon waveguides were doped, followed by contact and via formation and metalization, to form a p-i-n diode for monitoring the optical power circulating in the microring under reverse-bias, which helped in monitoring and stabilizing the resonance, as described below. The waveguides used in the feeder waveguide and microring had a propagation loss (measured on test sites consisting of waveguides of different lengths) of approximately 0.7 dB/cm at 1550 nm, resulting in an intrinsic Q-factor of approximately 9 × 10⁵, and a resonance lifetime τ ≈ 76 ps (loaded Q-factor of 9.2 × 10⁴, with a spectral full-width at half-maximum (FWHM) of approximately 2.1 GHz) as shown in Fig. 1(c). The microresonator’s loaded-cavity photon lifetime thus approximately matches the pulse duration of a 10 Gbps non-return-to-zero (NRZ) data stream, and the classical bits have the appropriate time duration for SFWM pumping.

For photon-pair generation measurements reported here, a bare-die silicon photonic chip containing the microring resonator was mounted on a temperature-controlled stage with a thermo-electric controller (TEC) in feedback with a thermistor on the stage mount. The experimental configuration including ancillary telecommunications components is shown in Fig. 1(d). Filters and polarization controllers have already been integrated into a silicon photonics platform for classical optics applications, and may be a convenient future simplication. The temperature of the microchip was maintained slightly above room temperature and the TEC was used to bring the microresonator into stable spectral alignment with the wavelength of the incoming data stream used as the optical pump for SFWM. The spectral alignment of the pump to the microring was continuously monitored during measurement using the reverse-biased photo-current of a silicon p-i-n junction diode fabricated across the microring [29], and confirmed using high-magnification infrared camera images of the microring. Light was coupled to and off the silicon chip using polarization-maintaining lensed tapered fibers (with anti-reflection coating). The state of polarization of the input data stream, though not actively controlled, was quite stable over a duration of several tens of minutes. Variations in the state of polarization change the generated rate of the signal and idler photons similarly, and may not affect certain applications such as quantum key distribution except in time-to-completion. For applications where a uniform brightness of photon pairs must be sustained, an active polarization controller or tracker should be incorporated into the chip or system design.

In the laboratory, the insertion loss of each fiber-to-waveguide coupler was estimated as −3.5 dB. The amplified spontaneous emission (ASE) background of the input light from the transceiver was suppressed at the wavelengths to be used for pair generation (i.e., the signal and idler wavelengths) using a relatively broad bandpass filter (FWHM of 1 nm at 1550 nm). The level of ASE background in a communications link varies widely with the specific implementation; however, let us assume, for the sake of performing a quick estimation, that the ASE level is about −50 dBm. A rate of generated photons at 1.55 micron wavelengths of 100 MHz translates to an average power level of about −79 dBm. Thus, it is important to suppress the ASE background at the signal and idler wavelength spectral windows by at least 40 dB, in this representative example, in order to be able to detect the photon pair above the background of photons already present from ASE. While a dedicated optical filter was used in this experiment, a wavelength-division multiplexing (WDM) add/drop element, widely used in practical networks, would achieve the same result [47].

Under SFWM, energy-conservation between the pump and the generated Stokes (S) and anti-Stokes (aS) photon pair dictates the frequency relationship, 2ω_p = ω_S +ω_aS, so that all three frequencies (wavelengths) lie within the band used in communication networks near 1.55 μm wavelengths. The microring provided simultaneous resonance for all three frequencies across adjacent free-spectral ranges. Output light from the chip was routed through cascaded filters to select one pair of spectral lines of Stokes and anti-Stokes photons positioned symmetrically around each pump wavelength. For these experiments, signal and idler wavlengths near 1535 nm and 1575 nm, respectively, were selected for convenience, since such photons can be easily separated using a standard telecommunications C/L band splitter component.

Photons were detected using fiber-coupled superconducting (WSi) nanowire single photon detectors (SNSPD), cooled to 0.8 K in a closed-cycle Helium-4 cryostat equipped with a sorption stage. The detection efficiencies were measured to be about 90% at 1550 nm wavelengths, with a timing jitter full-width at half-maximum of about 130 ps. Cryogenically-cooled detectors are expensive to build and operate, but quantum communication protocols such as MDI-QKD can be implemented by placing detectors (without transmitters) only at interior nodes of the network, and inexpensive transmitters (without detectors) at the edges. These detectors were not gated and operated in a simple DC-biased mode with an RF-amplified readout. Coincidences were measured using a multi-input time-to-digital converter (TDC) instrument, with 0.08 ns minimum bin width, in start-stop mode. To prevent binning artifacts when accumulating histograms, at the cost of a factor-of-two in temporal resolution, two adjacent hardware bins were summed, according the manufacturer’s suggestions, resulting in the 0.16 ns bin width used for all coincidence measurements.

3. Measurements

3.1. Singles and coincidence rates

Figure 2(a) shows the scaling of the singles rates for the signal and idler photons as a function of the average pump power, i.e., the average power of the 10 Gbps NRZ data stream used as the pump. These are the raw measured singles rates, and include uncompensated experimental loss contributions (−5 dB loss for coupling from the microchip to fiber, −7 dB insertion loss from filters, −1.9 dB detector efficiency); the on-chip singles rates are about 14 dB higher. As expected, [48] the scaling of the singles rates is quadratic with the average pump power. The scaling of the (loss-compensated) coincidence rate is shown in Fig. 2(b) and follows the expected quadratic trend (the vertical axis is shown on a logarithmic scale), until a drop-off was noticed at higher pump powers, most likely due to free-carrier absorption and /or thermal detuning of the microring resonator. The fitted efficiency of pair generation was 14.6 MHz.mW⁻², which is not the record result [13], but is comparable to most other reports of SFWM in silicon microrings (see Table 1 in 13).

 

Fig. 2 (a) Singles rates (raw measurements) versus average input power in the bus waveguide (b) Coincidence rate (scaled by losses, approximately −14 dB in each of the signal and idler photon pathways) versus average input power in the bus waveguide. The indicated fit, excluding the last few points, and shown by the solid line, is used to infer the pair generation rate. The right-hand side vertical axis reports the measured (raw) coincidence rates. (c) Coincidences to accidentals ratio (CAR) versus average input power in the bus waveguide.

Download Full Size | PDF

3.2. Coincidences-to-accidentals ratio (CAR)

Figure 2(c) and 3 report the measurements of CAR versus input average pump power. Raw two-fold coincidence counts (C _raw) and accidental coincidence counts (A_raw) between the generated photon pairs were measured and binned into a histogram (one for each input pump power level). The uncertainties in A_raw are one standard deviation values of counts in bins away from the peak (start-stop delays were measured up to 100 ns time difference), and were propagated to generate the error bars in the CAR plot. Coincidences due to dark counts were measured separately, but since their contribution was negligible, they were not subtracted from the measurements. Each histogram peak was fitted by a Gaussian function, whose full width at half-maximum (FWHM) was typically 0.31 ns, as shown by the representative example in Fig. 3(b).

 

Fig. 3 (a) Start-stop histogram for the measurement with the highest CAR value, 2873 ± 1415. (b) Plot of the coincidence peak. (c) A section of the accidental coincidences trace, showing the low level of background noise in the measurement.

Download Full Size | PDF

The histogram of start-stop coincidences (measured bin counts divided by the measurement time in seconds) which resulted in the highest CAR is shown in Fig. 3, along with a segment of the accidental coincidences in the inset figure. The peak was well fit by a Gaussian function. Its two-sided root-mean-square width defined the time window over which the (fitted) coincidence counts were integrated to yield C, and the same width was used to calculate the integrated averaged accidentals count, A, with CAR defined as CAR = C/A. (This number is less than C_raw/A_raw, since the accidentals counts are more or less flat over the integration window, whereas the coincidences are peaked.) The highest CAR was 3, 000 ± 500 measured using an integration time of 3,000 seconds. As expected, CAR decreased at higher pump powers, with a fitted trend shown in Fig. 2(c). The fit is based on the ratio-of-polynominals functional form CAR = P · (aP² + bP + c)⁻¹ which, at low pump powers, saturates at a certain maximum value and then drops to zero. [20] Other fitting forms have been used in the literature, which do not show evidence of CAR saturating at low P values, and sometimes, the saturating and non-saturating behavior can be seen for the same device at different wavelength regimes [36]. Here, we use the functional form that recognizes that, in an optical communications network, “true” coincidences generated by very low pump powers are likely to be masked out by the imperfectly-filtered background ASE level, since optical filters do not have infinite extinction ratio, and thus CAR should be low as P decreases further. For clarity, the extrapolation of the fit to the regime of lower pump powers than was actually used is shown in Fig. 2(c) by the dotted line. At the higher end of the pump powers used in these measurements (which were much less than 1 mW, or 0 dBm, the typical average power level of a data stream in an optical network), the error bars in the calculation of CAR were lower because the singles counts were higher. At an average input pump power of 0.23 mW, a CAR of 193 ± 11 was measured. While there is no fixed rule on what CAR values should be, values in excess of 50 have been recently considered satisfactory for typical applications of photon pairs in communications and quantum key distribution [21].

These measurements, based on optical pumping by a 10 Gbps data stream, are reassuringly similar to well-known trends and quantitative values in pair generation using silicon photonic microrings pumped by laboratory-grade diode laser instruments [21,30,49]. These measurements show that high-quality pair generation can be obtained by tapping a 10 Gbps NRZ data stream from an optical network, with sub-milliwatt average power levels, to serve as the pump.

3.3. Hanbury Brown - Twiss (HBT) measurement

Figure 4(a) shows the schematic for the measurement of the single-photon second-order self-correlation function, g⁽²⁾, obtained by measuring the self-correlation of the signal photon, either with or without using the idler photon as a herald, as two separate experiments. Coincidences were defined as simultaneous detections within a 5 ns time window, measured directly by the TDC hardware (calculating coincidences between combinations of input channels without software post-selection). One of the photons in the SFWM pair is detected in path ‘A’ as labeled in the figure, and may, or may not, be used as a ‘herald’ for a coincidence measurement. If heralded, the remaining single photon is expected to be anti-bunched, whereas when the signal arm is measured without regard to a herald, the photon is expected to be bunched [44] as shown by the start-stop statistics in Fig. 4(b). The peak value, g⁽²⁾(0) = 2.21 is close to the expected value g⁽²⁾(0) = 2 for a single-mode thermal state [50]. Generally, the impact of SNSPD timing jitter is to lower the peak value of g⁽²⁾(0) [32]. Evidence of weak superbunching indicates that some of the two-photon coincidences arise from more than one pathway, and have been observed in experiments on pseudothermal light generated by intensity-modulated, rather than continuous-wave, light incident on rotating-glass plates [51]. The TDC instrument performs these calculations in real-time with no post-selection; however, a very long acquisition time is needed for an accurate g⁽²⁾(t) measurement, which is currently limited here by the polarization and/or coupling drifts of the fibers to the unpackaged silicon photonic microchip. For comparison, Fig. 4(c) shows the HBT measurement for the SFP+ light (without pair generation) attenuated to the single-photon level (−100 dBm, resulting in singles count rates of about 1.25 × 10⁵ s⁻¹, and no evidence of bunching is seen.

 

Fig. 4 (a) Schematic of the measurement setup for second-order auto-correlation statistics on the signal photon. (b) For the unheralded measurement, g⁽²⁾(t) shows characteristic bunching, with a peak value slightly exceeding the expected value of 2.0. (c) No bunching is seen for the light from the transceiver data stream, attenuated to single-photon levels. (d) The heralded (by the idler photons) second-order autocorrelation of the signal photons at zero time difference, g⁽²⁾(0), decreases with pump power and since g⁽²⁾(0) ≪ 0.5, shows evidence of (heralded) single-photon character. (e) In comparison, g⁽²⁾(0) ≈ 1.0 for the light from the transceiver data stream itself, as expected.

Download Full Size | PDF

3.4. Heralded single-photon generation

Detecting one photon of the generated photon pair results in a heralded single-photon source, since the other photon is then expected to show non-classical anti-bunching behavior. This is expected to be true even if the pump pulses arrive at the microring at random times (i.e., pumped by a classical 10 Gbps NRZ data stream). Fig. 4(d) shows the scaling with pump power of the heralded (i.e., conditional) single-photon second-order self-correlation function, $g_H^{(2)}(0)$, obtained by detecting one of the generated photon pair as a herald, and measuring the self-correlation of the other photon in the presence of the herald. The normalized value of the photon correlation measurement on the heralded single photons at zero time delay was calculated using the formula [52] $g_H^{(2)}(0)=\frac{N_{ABC}{N}_A}{N_{AB}{N}_{AC}}$, where N_A is the average photon detection rate on the heralding SNSPD detector, double coincidences N_AB and N_AC correspond to average rates of simultaneous events on one of the detectors (B or C) and the heralding SNSPD detector (A), and triple coincidences N_ABC correspond to average rates of simultaneous events on all three detectors. The arrival times of events at the TDC module were synchronized by selecting appropriate lengths of BNC cables.

Counting times varied from 100 seconds for the higher pump powers to 600 seconds for the lowest pump power. The scaling trend seen in Fig. 4(d) is consistent with results obtained for pumping SFWM in the microring resonator with a laboratory continuous-wave laser, and g⁽²⁾(0) is proportional to the biphoton rate [53], which, in SFWM, is quadratic in the pump power, P, at low values. At the highest average power values (0.15 mW) used in this sequence of measurements, $g_H^{(2)}(0)=0.11\pm 0.051$ (the errorbar is one standard deviation uncertainty) at a heralding rate of N_A = 340 kHz, showing deeply sub-Poissoinian statistics. This is already well below the classical threshold. Even lower values, as low as $g_H^{(2)}(0)=0.005\pm 0.02$ were directly measured for a measured heralding rate of N_A = 18 kHz. The heralding efficiency is about 3–4%, unchanged from our previous experiments for this microring [13] and can be improved by changing the coupling coefficient between the microresonator and the bus waveguide, and lowering the insertion losses of the filters and detectors [30].

For comparison, Fig. 4(e) shows the g⁽²⁾(0) statistics of the light from the data stream itself, attenuated using external fiber-coupled attenuators to the single-photon regime, as would be used in a BB84-type QKD protocol and similar applications. Data shown using squares were obtained by measuring photons using the setup and calculation procedure resulting in panel (d) and described in the preceeding paragraphs. Data shown using diamonds were measured from the setup used in panel (c), i.e., the HBT measurement which resulted in a feature-less g⁽²⁾(t) profile, as shown in Fig. 4(c), and whose values were averaged over the central 10 nanosecond time window. The number of photons “per pulse” was calculated from the average power level (average power levels before the detector for the indicated points range from −94 dBm to −106 dBm) and with respect to the bit period (100 ps) of the 10 Gbps data stream, multiplied by 2 since one-half of the bits are ‘0’ and do not contribute to the average power). These measurements show Poissonian statistics.

Thus, heralded single photons can be generated by pumping a silicon microring resonator with a tapped fraction of a classical 10 Gbps NRZ data stream. As Fig. 4(d) shows, about one-tenth of a milliwatt needs to be extracted from the classical data stream, leaving the rest of the light to complete the classical link. Together with the high CAR values, this suggests that information for optical communications can be cleanly encoded on single photons, with only a small possibility that a receiver will detect two or more signal photons when triggered by a single herald, or that the detected signal and herald (idler) photons will be counted across the “wrong” time bins. These observations suggest that such photon pairs could be useful resources for quantum communications in optical networks.

3.5. Comment on energy-time entanglement

The generated photon pair could be used in energy-time entanglement; however, we do not yet have a simple method of verifying such entanglement. One popular method is through a Franson-type two-photon interference experiment, by violating Bell’s inequality [54,55]. Such measurements have already been shown for several silicon photonic pair-generation devices [27, 33, 56–59]. However, we cannot verify entanglement when pumping the SFWM process with a 10 Gbps NRZ data stream. As shown by the eye diagram in Fig. 1b, the average duration of the pump pulse which performs SFWM in the microring is somewhat shorter than 0.1 ns, the bit period, and thus, the time difference between the “short” and the “long” arms of the time-bin Franson interferometer would have to be significantly shorter than 0.1 ns, and the histograms of the two bins would also need to be cleanly resolved. However, the binning resolution of the apparatus used in these measurements is 0.16 ns, and the timing jitter of the detectors is about 0.13 ns. Consequently, the fitted full-width at half-maximum (FWHM) values of the coincidence peak, shown in Fig. 3, is currently about 0.3 ns, which is much larger than the bit period. We cannot yet measure the visibility of the coincidence fringe (as a function of the phase delay) in the Franson interferometer, but hope to report on it in the future by using reduced-jitter detectors and electronics, which have already been reported in the literature [60, 61]. Using the silicon microring with continuous-wave or slower repetition-rate pumping (e.g., 1 GHz), high values of Franson visibility (> 98%) have been reported for the same microring resonator used in these experiments [13,36].

4. Conclusion

These results show that the photon pairs generated by initiating SFWM using a 10 Gbps NRZ data stream as the pump were sufficient in quantity and quality in comparison with results previously obtained using a stand-alone laser. Si microring resonators can be designed such that their pump power requirements for SFWM are compatible with a tap fraction of a typical classical communications data stream, which has an average power of about 0 dBm. The microresonator photon lifetime approximately matches the pulse duration of the 10 Gbps data stream, and thus, the NRZ classical bits are already carved into the appropriate time duration for near-optimal SFWM pumping. SFWM itself is a random process, and therefore, it is not of concern that a PRBS classical bitstream serving as the pump does not have a “1” pulse deterministically in every time slot, unlike, say, a mode-locked laser serving as the pump. The use of (relatively inexpensive) silicon photonics and the elimination of the separate pump diode requirement show progress in the same spirit of cost reductions and simplifications of practical quantum photonics devices using existing telecommunications technology.