1. Introduction

Silicon photonics has emerged as a promising technology platform for low-cost solutions to optical communications and interconnects [1, 2]. Recently, several important breakthroughs have been reported. This includes silicon modulators [3–5], amplifiers [6–10], and lasers [11–13]. All-optical wavelength conversion based on coherent anti-Stokes Raman scattering or four–wave mixing (FWM) in silicon waveguides is another attractive area that has been explored [14–16]. However, the conversion efficiencies reported remain in the range of -50 dB to -35dB. One of the major limitations is the strong nonlinear absorption that occurs at high pump powers. Although linear optical absorption in silicon at wavelengths of 1.3-1.7 μm is small [17], two-photon absorption (TPA) induced free carrier absorption (FCA) causes the optical loss to increase with the pump power [18–20]. We have demonstrated that the TPA induced FCA in silicon can be significantly reduced by introducing a reverse biased p-i-n diode structure embedded in a silicon waveguide, shortening the free carrier life time [10]. Using such p-i-n waveguides we have achieved FWM conversion efficiency of up to -8.5 dB, which is comparable to LiNbO₃ based devices [21–23]. We show high-speed optical data stream at 10Gb/s on one DWDM channel in the C-band can be converted to another channel with practically no waveform distortion.

2. Device description

The silicon rib waveguides are fabricated on the (100) surface of a silicon-on-insulator (SOI) substrate using standard photolithographic patterning and reactive ion etching techniques. The rib waveguide width (W) is 1.5 μm, the rib height (H) is 1.55 μm, and the etch depth (h) is 0.7 μm. A typical scanning electron microscope image of the waveguide cross-section is shown in Fig. 1. The effective core area of the waveguide is calculated to be ~1.6 μm² by using a fully vectorial waveguide modal solver [24]. In our experiments, we use waveguides with three different lengths: a straight waveguide of 1.6 cm long, an S-bend waveguide of total 4.8 cm long, and a double S-bend waveguide of 8 cm long (see Fig. 2). All bends have bend radius of 400 μm. The straight sections of the waveguides are oriented along the [011] direction. To reduce the nonlinear optical loss due to the TPA induced FCA, a reverse biased p-i-n diode structure is fabricated in the waveguides by implanting boron and phosphorus in the slab on either side of the rib waveguide with a doping concentration of ~1×10²⁰ cm^-3. The separation between the p and n type doped regions is 6-8 μm. Aluminum films are deposited on the p and n doped regions to form ohmic contacts.

 

Fig. 1. Scanning electron microscope image of a typical p-i-n waveguide cross-section.

Download Full Size | PDF

It has been experimentally verified that the presence of these doped regions and metal contacts has negligible effect on the waveguide loss, which is due to the tightly confined mode for the waveguide being used. The linear optical transmission loss of the waveguides is 0.4±0.1 dB/cm; measured using the Fabry-Perot resonance technique [2] prior to anti-reflection coating the silicon waveguide facets. The 0.1 dB/cm uncertainty in the linear optical loss includes the experimental error and waveguide-to-waveguide variations.

3. Experiments and results

A schematic of the experimental setup is shown in Fig. 2. The pump and signal lasers are combined with a wavelength multiplexer into a lensed single-mode fiber whose output is coupled into the waveguide under investigation. The coupling loss between the lensed fiber and the waveguide is measured to be 4 dB. The output beam of the waveguide is coupled into another lensed single-mode fiber. An optical spectrum analyzer (OSA) is used to analyze the spectrum of the output light. To measure the converted optical signal and to generate eye diagrams using a digital communications analyzer (DCA), the converted signal at the new wavelength needs to be discriminated from the pump and the input signal. For this purpose, a band-pass filter at the converted signal wavelength is used to separate the converted signal from the pump and input signal beams. Fiber polarization controllers are used to align the polarization of the pump and signal beams. The device under test is mounted on a thermo-electric cooler and kept at a constant temperature of 20 °C. The pump laser is a CW external cavity laser emitting around 1550 nm, which is amplified using an erbium doped fiber amplifier system to a maximum output power of 1.6 W. The input signal laser is a 4 mW, CW external cavity tunable diode laser with a tuning range over 60 nm. The polarization of the signal and the pump beams are aligned with the TE mode of the waveguide.

 

Fig. 2. Schematic of the experimental setup for wavelength conversion experiment

Download Full Size | PDF

Figure 3 is a typical spectrum of the output beam from the silicon waveguide. In this example, the pump is at λ₁=1549.39 nm, and the input signal at λ₂=1548.37 nm. A new wavelength is generated at λ₃=1550.41 nm, which is exactly where the FWM process predicted, namely 1/λ₃=2/λ₁-1/λ₂. For easy comparison with other published works, we follow the definition of wavelength conversion efficiency as the ratio between the peak levels of the converted signal at λ₃ and the original signal at λ₂ in Fig. 3. The conversion efficiency shown here is -11.5 dB for an 8-cm long waveguide with pump power of 320mW into the waveguide.

 

Fig. 3. Spectrum of the output beam from an 8 cm long waveguide. Coupled pump power is 320 mW, and the conversion efficiency is -11.5 dB

Download Full Size | PDF

The small peak at λ₄ corresponds to the FWM condition 1/λ₄=2/λ₂-1/λ₁, and is ~37 dB below the signal peak at λ₂ in this example. Such FWM components will grow with increased input signal and pump powers, and may cause channel crosstalk for synchronous multichannel conversions with strong input signals. Depending on network configurations and requirements this effect needs to be taken into account in the WDM system design.

The conversion efficiency is also measured as a function of the pump power for an 8 cm long waveguide at various bias voltages. As shown in Fig. 4, at low pump powers, the bias on the p-i-n diode does not have a strong effect on the conversion efficiency, but at high pump powers, the nonlinear absorption starts to reduce the effective pump power and the efficiency drops. Eventually it saturates when the p-i-n is open. By applying the reverse bias to the p-i-n diode in the waveguide, sweeping out the two-photon absorption generated free carriers, the conversion efficiency significantly increases. At a pump power of 640 mW or intensity of 40 W/cm² inside the waveguide, a conversion efficiency of -8.5 dB can be reached with -25 V reverse bias. More than 3dB improvement is realized by reverse biasing. At low pump powers, the slope of the conversion efficiency curves in the log-log plot of Fig. 4 is approaching 2, indicating that the conversion efficiency scales with the pump power squared, which is expected from the FWM process [25]. This also means that increasing the effective pump power inside the waveguide by reducing the nonlinear loss with reverse biased the p-i-n diode is an efficient way to enhance the conversion efficiency.

 

Fig. 4. Wavelength conversion efficiency as a function of pump power coupled into the waveguide for an 8 cm long double S-bend waveguide at different bias voltages.

Download Full Size | PDF

Figure 5 shows the wavelength conversion efficiency η as a function of the wavelength detuning Δλ, defined as the wavelength difference between the pump and the signal. These measurements are performed for a 4.8 cm long S-bend and a 1.6 cm long straight waveguide at pump power of 200 mW inside the waveguide. The η vs. Δλ curve is symmetric around Δλ =0, so only the Δλ > 0 half is plotted. As can be seen, the longer waveguide can have higher conversion efficiency but the bandwidth is narrower. In contrast, the shorter waveguide has broad conversion bandwidth but lower efficiency. This is understandable as the longer waveguide has longer interaction length and therefore can have higher efficiency for smaller detuning. When Δλ increases the longer propagation length produces large phase mismatch between the interacting beams and the efficiency drops quickly. The 3-dB conversion bandwidth is 20 nm and 32 nm for the 4.8 cm and 1.6 cm long waveguide, respectively. The sinc² function shape of the conversion efficiency curves qualitatively agrees with theoretical calculations [14, 15].

 

Fig. 5. Wavelength conversion efficiency as a function of signal wavelength detuning from the pump wavelength for a 4.8 cm long S-bend waveguide and a 1.6 cm long straight waveguide. The pump wavelength is 1550 nm and the pump power is 200 mW inside the waveguide.

Download Full Size | PDF

To demonstrate the conversion of high-speed optical data, we modulate the input signal on ITU channel 30 (1553.33nm) with a pseudo-random bit sequence at 10 Gb/s rate and we measure the converted signal on channel 34 (1550.12nm). Fig. 6 shows a comparison of the time-domain waveforms of the two normalized signals. They look identical and no distortion is observable. In this configuration with constant CW pump and low power input signal the free carrier density in the waveguide remains approximately the same level over time, which results in the conversion speed not being affected by the free carrier decay time estimated to be ~1ns with -25V bias [10]. This is in comparison to silicon wavelength converters based on pump beam modulation [16], or free-carrier dispersion effect [26].

 

Fig. 6. Time-domain waveforms of the input (blue) and converted (pink) signals at 10Gb/s

Download Full Size | PDF

Figure 7 shows the eye diagrams of the input optical data on ITU channel 30 and the converted data on channel 34 at 10 Gb/s rate. While the eye shape remains essentially the same, the signal to noise ratio is reduced on the converted signal. This mainly is due to the intensity noise of the pump and the noise from the photo-detector that is at the detection limit for the weak converted signal.

 

Fig. 7. Eye diagrams of the converted (left) signal with input (right) signal at 10Gb/s

Download Full Size | PDF

4. Conclusion

In conclusion, we have achieved practical level of wavelength conversion efficiency in silicon waveguides via FWM process which allows us to demonstrate high-speed optical data transfer from one to another ITU channel in the C-band. We have measured -8.5 dB of conversion efficiency at a single channel data rate of 10 Gb/s. Compared to previous published works based on silicon channel waveguides [15, 16], our rib waveguides have much lower propagation loss and allow p-i-n diode structure to be built in, reducing nonlinear losses at high pump powers. The low loss bends enable longer waveguides to be folded on a single chip with small footprint. Lower loss, longer interaction length and being able to take higher pump power in the p-i-n waveguides result in higher conversion efficiency. Compared to wavelength converters based on difference-frequency generation in periodically poled lithium niobate waveguides, silicon waveguide devices are not subject to photorefractive damage causing degradation of the quasi-phase matching conditions, therefore have superior reliability. In addition to modulators, amplifiers and lasers, efficient all-optical wavelength conversion further expands the active functionalities of silicon based photonics devices. By reducing waveguide dimensions and optimizing the p-i-n diode design, the carrier lifetime could be further reduced and even higher conversion efficiency is achievable. Using resonant structures such as ring resonators or Bragg grating cavities, the FWM effect can be further enhanced [16]. By proper design of the waveguide dimensions to achieve phase matching conditions [27], high conversion efficiency with broader conversion band width can be achieved.