1. Introduction

A challenge for optical transceivers in short-reach data communications is a mass-manufacturable, temperature-independent, power-efficient, high-speed optical transmitter. Mach–Zehnder modulators (MZMs) are often preferred over electro-absorption modulators (EAMs), since the latter typically require active thermal stabilization. EAMs rely on the steep change in absorption near the band-edge, which red-shifts as the temperature is increased [1]. Monolithically integrated silicon (Si) MZMs emerged as a promising technology [2] and are now of significant commercial interest[3]; howeverthe modulation efficiency (V${ }_{\pi }$L) of typical lateral PN junction based Si MZMs is limited to about 2.5 V$\cdot$cm. Si MZMs utilizing more complex geometries have been demonstrated, with V${ }_{\pi }$L as low as 0.2-0.5 V$\cdot$cm [4, 5, 6], but these often come at the expense of increased optical losses or limited bandwidth. Modulator performance on silicon platforms can be further improved by the heterogeneous integration of materials that demonstrate higher electro-optic (EO) modulation bandwidths (BWs) and efficiencies, such as LiNbO${ }_3$ [7] and EO polymers [8, 9, 10]. Although such χ${ }^{\left(2\right)}$-based EO materials can support high BWs in excess of 100 GHz [11, 12] and a linear EO response, forming a complete transmitter would still require the additional use of III-V compounds for the laser.

For optical gain, as well as refractive index modulation, III-V quaternary alloys are excellent candidates for heterogeneous integration with Si. The low effective mass of electrons in these materials increases the plasma dispersion effect, which, when combined with the band filling effect, results in electron-induced refractive index changes that are 5 to 10 times larger in indium phosphide (InP) than in Si in the O and C bands [13, 14, 2, 15]. Heterogeneously integrated InP-on-Si capacitive phase modulators were first demonstrated for tuning the wavelength of a microring laser [16]. Recently, InGaAsP-on-Si capacitive MZMs were reported at 1550 nm [17, 18] and 1310 nm [19], with record-low V${ }_{\pi }$L near 0.05-0.1 V$\cdot$cm. These modulators have a similar geometry to the Si semiconductor-insulator-semiconductor capacitive (SISCAP) modulator in [4] and exploit the carrier accumulation on the Si-oxide interfaces for the effective index modulation. However, the EO BWs of the C-band devices were limited to < 4 GHz and modulation at 25 and 32 Gb/s required pre-emphasis and an optical amplifier. No high frequency measurements were reported for the the O-band device.

In this article, we report on the first high speed heterogeneous InP-on-Si capacitive MZMs for the O-band [20]. MZMs with 250 and 500 μm long phase-shifters have EO BWs of 30 and 11 GHz, respectively, IL of ~0.5-1 dB, andV${ }_{\pi }L$ of 1.3 V$\cdot$cm. Open eye patterns at up to 25 Gb/s are demonstrated without optical amplifiers or pre-emphasis. The InP electrode of the MZM utilizes the same layer of the N-contact as in InP-on-Si lasers, and the modulator fabricationprocess is fully compatible with the heterogeneous InP-on-Si laser integration process [21, 22].

2. Device design

The cross-section and waveguide parameters of the nominal InP-on-Si phase modulator design are shown in Fig. 1(a). The hybrid waveguide phase-shifters are composed of a 400 nm or 500 nm wide (w${ }_{\text{rib}}$) P-doped Si rib with an N-doped InP slab separated by a thin layer of oxide (t${ }_{\text{ox}}$). The doping concentrations were 5$\times$10${ }^{17}$ cm${ }^{-3}$, 2$\times$10${ }^{19}$ cm${ }^{-3}$, and 3$\times$10${ }^{18}$ cm${ }^{-3}$ for the P, P+, and N+ regions, respectively. The targeted oxide thickness between the Si and InP was 20 nm. The InP slab overhangs the rib waveguide to improve variation tolerance to the InP processing, which depends on the lithography alignment and the InP etching. Figure 1(b) shows a scanning electron micrograph (SEM) cross-section of a fabricated modulator with an InP overhang (w${ }_{\text{top}}$) of ~0 nm. Figure 1(c) shows the simulated mode profile of a 400 nm wide hybrid waveguide with an InP overhang of 300 nm at a wavelength of 1310 nm. The optical mode is primarily confined within the Si waveguide due its higher refractive index.

 

Figure 1 (a) Cross-section of hybrid InP-on-Si capacitive modulator platform. (b) SEM cross-section of a modulator with a w${ }_{\text{top}}$ of ~0 nm. (c) Optical mode supported by the hybrid InP-on-Si waveguide.

Download Full Size | PDF

Figure 2 shows the simulated EO efficiency, as given by the V${ }_{\pi }$L, the propagation loss due to free carrier absorption, and the high-frequency capacitance (simulated at 1 MHz), which were computed using Lumerical DEVICE and MODE. For the simulations in Fig. 2, the fixed waveguide parameters were set to w${ }_{\text{rib}}$ = 400 nm, t${ }_{\text{ox}}$ = 20 nm, and w${ }_{\text{top}}$ = 300 nm. We used the data in [15] and [23] for the change in the real and imaginary parts of the InP refractive index, respectively. Figure 2(a) shows the simulated DC EO efficiency and propagation loss of a 400 nm wide InP-on-Si capacitive phase-shifter as a function of the oxide thickness. The oxide thickness between the InP and Si sets the device capacitance (Fig. 2(b)). A higher capacitance, due to a thinner oxide layer, leads to a larger carrier density change per volt at the oxide interfaces and a higher EO efficiency. The V${ }_{\pi }$L shown in Fig. 2 is computed using the effective index difference between a 0 V and 4 V bias to the device, and the absorption loss is taken as the average between a 0 V and 4 V bias. The V${ }_{\pi }$L varies between 0.6 and 1.2 V$\cdot$cm and is roughly inversely proportional to t${ }_{\text{ox}}$. To further increase the device efficiency, the oxide spacer can be replaced with a high-$\kappa$ dielectric [17, 16, 24].

 

Figure 2 Simulated efficiency, absorption loss, and high-frequency capacitance as a function of (a)-(b) oxide thickness, (c)-(d) Si rib waveguides width, and (e)-(f) InP overhang.

Download Full Size | PDF

Figures 2(c) and 2(d) show the simulated DC EO efficiency, propagation loss, and high-frequency capacitance of the hybrid waveguide with a 20 nm thick oxide spacer as a function of the Si rib width. A rib width of 400-500 nm provides a balance between the EO efficiency, loss, and capacitance (bandwidth). In narrower waveguides, the optical confinement within the waveguide is reduced, leading to lower efficiency and higher propagation losses. As the waveguide width is increased, the device capacitance increases but the carrier density at the interface is unchanged. Thus, the slight increase in efficiency, due to a better mode overlap between the optical mode and accumulated carriers, comes with an increase in capacitance which may limit the $RC$ time constant. Figures 2(e) and 2(f) show the modulator characteristics as a function of the InP overhang. Once the InP overhang exceeds ~300 nm, the modulator performance becomes nearly independent of w${ }_{\text{top}}$.

3. Device fabrication

The devices were fabricated on a 200 mm silicon-on-insulator (SOI) wafer. Prior to waveguide etching, the Si was implanted P and P+ in two steps. The etched waveguide was then encapsulated in low-temperature deposited SiO${ }_2$, which was planarized and thinned down to a targeted thickness of 20 nm using chemical mechanical polishing (CMP). Post-CMP wafer-wide ellipsometry measurements indicated that the SiO${ }_2$ thickness where the InP wafer was bonded ranged from 10 nm to 40 nm. An InP substrate with an epitaxially grown 300 nm thick N+ doped InP layer was then directly bonded to the SOI at room temperature and annealed at relatively low temperatures. Following the removal of the InP substrate, the N+ doped InP was etched to form the N electrode of themodulator. The electrical contacts were formed on the electrodes, which were then encapsulated in benzocyclobutene (BCB). Vias were formed in the BCB to connect a top metal layer with the electrical contacts.

4. Mach–Zehnder modulators

The capacitive phase-modulators were integrated into MZMs. Two sets of devices were evaluated, one with a nominal rib width of 400 nm and length of 250 μm and one with a nominal rib with of 500 nm and length of 500 μm. The devices were driven as lumped elements, without traveling-wave electrodes. An optical micrograph of the fabricated 500 μm MZM is shown in Fig. 3. The MZM is composed of two 3 dB multimode interference (MMI) couplers, a path imbalance $\mathrm{\Delta }L$ using 400 nm wide Si rib waveguides, and a hybrid InP-on-Si phase-shifter section. The path imbalance was introduced to ease the testing.

 

Figure 3 Optical micrograph of the fabricated 500 μm long MZM.

Download Full Size | PDF

To form the hybrid waveguide, a 60 μm long, linear, adiabatic taper was used to introduce the InP slab over the Si waveguide. The transition losses due to the adiabatic tapers were checked by comparing the transmission spectra of cascaded chains of 2, 4, and 8 transitions from passive Si waveguide to the hybrid InP-on-Si section to equivalent lengths of hybrid InP-on-Si waveguide. The spectra showed no distinguishable difference, suggesting that the transition losses were negligible.

With the spectrum of the input/output grating couplers removed, the device insertion loss was ~0.5 dB and 1 dB for the MZM with 250 μm and 500 μm long phase shifters, respectively. Assuming insertion losses of 0.1 dB for the MMIs, the propagation loss of the hybrid InP-on-Si waveguide was 12-16 dB/cm. This is in good agreement with the simulated value of about 12 dB/cm in Fig. 1(c) plus about 2.5 dB/cm due to the sidewall roughness of the Si waveguide.

4.1. DC characterization

Figures 4(a) and 4(b) show the shift in the transmission spectrum as a function of a DC bias voltage applied to the top arm of the 250 μm long (w${ }_{\text{rib}}$ = 400 nm) and 500 μm long (w${ }_{\text{rib}}$ = 500 nm) MZMs, respectively. The response of the grating couplers has been removed from the data in Figs. 4(a) and 4(b). The DC MZM transmission spectra showed extinction ratio (ERs) around 25 and 35 dB for the 250 and 500 μm long modulators, respectively, suggesting that the losses in the two arms of the modulator were nearly balanced. The measured phase shift per unit length as a function of voltage is plotted for the two devices in Fig. 4(c) (solid lines) and agrees well with simulation (dashed lines). We extract the V${ }_{\pi }$L by considering the wavelength shift of the transmission minima near 1310 nm when the bias is varied between 0 V and 4 V. For both modulators, the measured V${ }_{\pi }$L was 1.3 V$\cdot$cm.

 

Figure 4 The DC tuning spectra of (a) the 250 μm long MZM and (b) the 500 μm long MZM, respectively, when a voltage is applied to the top arm. The response of the grating couplers has been removed from the data in (a) and (b). (c) Accumulated phase shift per mm as a function of applied voltage for the measured (solid lines) and simulated (dashed lines) hybrid waveguides. (d) Measured and simulated capacitance of the 400 nm wide hybrid waveguide, as a function of applied voltage, for an oxide thickness of 20 nm. (e) Measured and simulated capacitance of the 500 nm wide hybrid waveguide, as a function of applied voltage, for an oxide thickness of 20 nm. The measurement with a lower modulation frequency shows a higher capacitance in depletion.

Download Full Size | PDF

Figures 4(d) and 4(e) compare the measured and simulated capacitance per unit length for the 400 nm wide and 500 nm wide capacitors, respectively, at a frequency of 1 MHz and an oxide thickness of 20 nm. The capacitance increases with increasing bias as the capacitor reaches accumulation. Although the capacitance decreases in reverse bias (i.e., in depletion), which can support a higher BW, this comes at the cost of a decrease in efficiency. In both devices, the measured capacitance was higher than the simulated value. It is unlikely that the discrepancy is due solely to a difference in the simulated and fabricated oxide thickness, as the device efficiency agrees quite well with simulation. The low-temperature deposited TEOS SiO${ }_2$ used in the modulator is expected to have a higher dielectric constant than thermally grown SiO${ }_2$ [25], which could contribute to a higher than expected capacitance in both the depletion and accumulation regimes. The measured capacitance of the 500 nm wide capacitor at a modulation frequency of 100 kHz, which is included in Fig. 4(e), is higher in depletion than the measurement at 1 MHz, indicating that there are traps at the oxide interface. This may further account for some of thediscrepancy between the simulated and measured capacitance. Lastly, the difference in the simulated and measured capacitance is roughly twice as large for the capacitor with w${ }_{\text{rib}}$ = 500 nm than for the capacitor with w${ }_{\text{rib}}$ = 400 nm, suggesting there may be some fabrication non-uniformity.

The access resistance of the 250 μm and 500 μm long modulators were about 38 Ω and 22 Ω, respectively, with roughly half coming from the P contact. The EO 3 dB BWs of the devices, which are equivalent to the electrical 6 dB BWs due to the conversion from optical power to voltage at the receiver, can be approximated by $f_{3dB}=\sqrt{3}/2\pi RC$[26]. Using the measured capacitances in Figs. 4(d) and 4(e), the $RC$ limited EO 3 dB BWs of the 250 μm and 500 μm long modulators with a 0 V bias are about 29 GHz and 17 GHz, respectively.

4.2. High-frequency characterization

Figures 5(a) and 5(b) show the measured scattering parameters, the electrical S11 and EO S21, of the MZMs at 0 V and 2 V bias taken with a 67 GHz vector network analyzer (Agilent N4373C) with the input laser wavelength tuned to the MZM -3 dB transmission point. No optical amplifiers were used in the measurement. The top arm of the modulator was driven without termination. The EO S21 is plotted as half of the electrical S21 [26]. The 3 dB frequencies of the 250 and 500 um long modulators were 30 GHz and 11 GHz with a 0 V bias and 26 GHz and 9 GHz with a 2 V bias, respectively. In forward bias, the devices were more efficient, but their BWs were reduced due to the increased capacitance. The peaking in the EO S21 of the250 μm modulators at low frequencies (~3 GHz) may have been caused by reflected electrical power, which increased the effective voltage across the MZM. In the longer device, the peaking is no longer present.

 

Figure 5 The measured electrical S11 and EO S21 of (a) the 250 μm long modulator and (b) the 500 μm modulator. The EO 3 dB BWs in (a) and (b) are 30 GHz and 11 GHz with a 0 V bias applied and 26 GHz and 9 GHz with a 2 V bias, respectively.(c) The eye pattern of the 500 μm long modulator biased at 0 V and driven with 2.4 V${ }_{\text{pp}}$ at 10 Gb/s using the PRBS-9 pattern. No optical amplification or pre-emphasis was used. The ER and OMA were 4.8 dB and -2.3 dBm, respectively. (d) Theeye pattern of the 250 μm long modulator biased at 0 V and driven with 4 V${ }_{\text{pp}}$ at 25 Gb/s using the PRBS-9 pattern. The ER and OMA are 2.3 dB and -5.1 dBm, respectively.

Download Full Size | PDF

Figures 5(c) and 5(d) show the measured eye patterns of the MZMs using a 2${ }^9$-1 pseudo-random binary sequence (PRBS) signal, without optical amplification or pre-emphasis. For both eye diagrams, the laser power was 15 mW. The laser was set to a wavelength corresponding to -3 dB from the MZM transmission maximum. The optical output from the MZM was directly detected using a 12 GHz photoreceiver (New Focus 1544) connected to a 80 GHz BW sampling oscilloscope (Keysight 86100D, 86116C module). For the eye patterns in Fig. 5, the MZMs were differentially driven. The large EO BWs of the devices enabled high-frequency operation without pre-emphasis. Figure 5(c) shows the eye diagram from the 500 μm long modulator operating at 10 Gb/s when biased at 0 V and driven at 2.4 V${ }_{\text{pp}}$. The extinction ratio and optical modulation amplitude (OMA) were 4.8 dB and -2.3 dBm, respectively. Figure 5(d) shows the eye diagram from the 250 μm long modulator operating at 25 Gb/s when biased at 0 V and driven at 4 V${ }_{\text{pp}}$. The ER and OMA were 2.3 dB and -5.1 dBm, respectively.

5. Discussion

Compared to prior works on hybrid InP-on-Si capacitive MZMs [17, 18], we have achieved higher EO BW operation (30 GHz and $\ge$ 25 Gb/s) without pre-emphasis. The high EO BW was enabled by a lower series resistance and a lower capacitance, though the latter led to a lower EO efficiency. The ERs of the MZMs are consistent with the DC V${ }_{\pi }$L, the lengths of the hybrid waveguide sections, and the applied voltage at 10 Gb/s, but lower than expected at 25 Gb/s. Inferring from the DC EO phase shift efficiency, the maximum achievable DC ER for the 500 μm long modulator shown in Fig. 5(c) with a drive voltage of V${ }_{\text{PP}}$ = 2.4 V is 5.3 dB, which is close to the measured value of 4.8 dB. The maximum achievable DC ER for the 250 μm long modulator shown in Fig. 5(d) with a drive voltage of V${ }_{\text{PP}}$ = 4 V is 4.4 dB, much higher than the measured ER of 2.3 dB. This is most likely due to the limited bandwidth of the photoreceiver (12 GHz), since the efficiency roll-off of the modulator should not be significant at 25 Gb/s.

The performance of these devices can be significantly improved in the future. Using a silicide on the P contact would roughly halve the access resistance of the device, enabling higher BWs without sacrificing capacitance. Optimization of the layer thicknesses and waveguide geometry can provide higher efficiencies by improving the optical mode overlap with the accumulation region. The mode overlap can also be improved by using a higher index III-V quaternary alloy, which would more evenly distribute the mode between the two materials. Alternatively, the efficiency could be improved by introducing a III-V quaternary alloy that has been optimized to provide the highest index change per electron. To reduce the V${ }_{\pi }$ of the device, longer phase modulation sections can be used in conjunction with traveling wave electrodes. Lastly, the use of an alternative dielectric material between the Si and InP electrodes, with improved thickness accuracy, could reduce the device variability.

6. Conclusion

In summary, we have demonstrated the first high BW InP-on-Si capacitive MZMs operating in the O-band. The hybrid phase modulators achieved an efficiency of 1.3 V$\cdot$cm and optical losses of 12-16 dB/cm. An EO 3 dB BW of up to 30 GHz was achieved and open eye patterns were observed at up to 25 Gb/s. The performance of these devices can be significantly improved in future fabrication iterations, to achieve both higher efficiency and higher EO BWs. Importantly, the fabrication is compatible with hybrid laser integration, and laser integration is currently underway.