Introduction

Optical interconnects are the most promising option to overcome transmission bottlenecks in data centres and high-performance computers, and energy consumption is one of the most important parameters of the associated photonic–electronic interfaces. Targeted figures are tens of femtojoule per bit for transmitters in off-chip connections, and a few femtojoule per bit for on-chip links.1 Key requirements are low drive voltages that can be provided by standard CMOS (complementary metal oxide semiconductor) circuitry without further amplification. Apart from energy efficiency, modulators need to provide fast electro-optic (EO) response along with large optical operation bandwidth to ensure high-speed transmission and flexibility in wavelength-division multiplexing systems. Moreover, dense integration is essential, calling for a small device footprint.

Silicon photonics is currently the most promising technology to realize such devices, leveraging mature high-yield CMOS processing and offering the potential of photonic–electronic co-integration on large-area silicon wafers. However, second-order nonlinearities are absent in bulk silicon due to inversion symmetry of the crystal lattice. Hence, current silicon-based modulators2,3 have to rely on free-carrier depletion or injection in p–n, p–i–n or metal-oxide-semiconductor structures. This leads to various tradeoffs when realizing fast and energy-efficient devices with small footprint: while carrier injection in forward-biased p–i–n structures enables compact modulators with voltage-length products as low4 as UπL=0.36 V mm, free-carrier lifetime currently limits the modulation speed to 25 Gbit s−1 and requires strong pre-emphasis of the drive signal.5 Moreover, energy efficiency of these devices is intrinsically limited to the pJ bit−1 range due to the permanent injection current flowing through the diode section. In contrast, carrier depletion in reverse-biased p–n junctions enables negligible bias currents and symbol rates6 of up to 50 GBd, but typical voltage-length products are beyond 10 V mm and thus, much larger than those of injection-type devices. For a silicon-based Mach–Zehnder modulator (MZM), the lowest reported energy consumption amounts to 200 fJ bit−1, achieved in a 5-mm-long depletion-type device.7 Modulation energies and device footprint can be significantly reduced by using resonant structures such as microrings, microdiscs or photonic–crystal waveguides.2,8 For resonant silicon-based modulators, the lowest energy consumption reported to date amounts to 3 fJ bit−1 and has been achieved with a microdisc device8 operated at a data rate of 12.5 Gbit s−1. However, the optical bandwidth of resonant devices is inherently limited, and resonance wavelengths are often subject to strong temperature-induced drifts.

Our work aims at overcoming the intrinsic limitations of all-silicon devices by combining conventional silicon-on-insulator (SOI) slot waveguides9 with organic cladding materials in a hybrid approach. The concept of silicon–organic hybrid (SOH) integration10,11 leverages both the benefits of large-scale standardized CMOS processing and the wealth of optical properties provided by theory-guided molecular design of organic materials.12 In particular, the SOH concept lends itself to highly efficient EO modulators that feature a large modulation bandwidth, a broad range of operation wavelengths and a small device footprint.13,14,15 First, SOH EO devices have been demonstrated using guest-host systems of polymer materials that are doped with EO molecules, so-called chromophores.16,17,18,19,20,21 When used as thick layers between transparent electrodes, such materials exhibit high EO coefficients22 of up to 118 pm V−1. However, for nanophotonic SOH modulators, in-device EO coefficients are found to be much lower with typical values in the range15,17,23,24,25,26 of 20–60 pm V−1, leading to energy consumptions19 of 300 fJ bit−1 or more. These figures are comparable to those of all-silicon devices, but still far from exploiting the full potential of the SOH approach.

Here, we demonstrate that an ultralow energy consumption of down to 0.7 fJ bit−1 can be achieved on the silicon integration platform even with non-resonant devices. In a proof-of-concept experiment, a 1-mm-long SOH MZM is operated with drive voltages as small as 80 mVpp at a data rate of 12.5 Gbit s−1 with a bit-error ratio (BER) below the threshold for hard-decision forward-error correction.27 This demonstration is enabled by combining an optimized SOH slot-waveguide modulator with a novel class of organic EO materials, which consists of pure chromophores instead of a chromophore–polymer guest–host system. The modulator exhibits an in-device EO coefficient of 180 pm V−1, and enables modulation with peak-to-peak drive voltages of down to 80 mV at 12.5 Gbit s−1. The devices support data rates of up to 40 Gbit s−1 with an extinction ratio exceeding 10 dB. At DC (direct current), the UπL product is as low as 0.5 V mm.

Materials and methods

SOH EO modulator

A schematic and a cross-sectional view of an SOH MZM are displayed in Figure 1a and 1b, respectively. The device consists of two 1-mm-long SOH phase modulators that are driven in push–pull configuration by a single coplanar transmission line with a ground-signal-ground configuration. A cross-sectional view of a single SOH phase modulator is depicted in Figure 1c along with the electric field profile of the fundamental quasi-TE (quasi-transverse-electric) mode. The phase modulator section consists of an SOI slot waveguide that comprises two silicon rails separated by a 160 nm wide slot. The waveguide is covered and the slot is filled with an organic EO material that is deposited by spin coating. The rails of the slot waveguide are electrically connected to a copper radio frequency (RF) transmission line by thin n-doped silicon slabs and tungsten vias. A voltage applied to the transmission line thus drops across the narrow slot, resulting in a strong modulation field that has a large overlap with the guided optical mode.

Figure 1
figure 1

Schematic of an SOH MZM. (a) Two 1-mm-long silicon slot waveguides act as phase modulators and are driven in push–pull mode by a ground–signal–ground coplanar transmission line. Further details of the photonic integrated circuit are explained in the Supplementary Information. (b) Cross-section of the MZM. The transmission line consists of 600-nm-thick copper (Cu) electrodes connected to the silicon slot waveguides by 900-nm-high tungsten (W) vias. The organic cladding (not shown) is deposited into the slot region of the SOI waveguide and is then poled at an elevated temperature by applying a poling voltage Upoling between the floating ground electrodes of the coplanar transmission line. This leads to a strong poling field Epoling=Upoling/2Wslot in each slot that aligns the dipolar chromophores as indicated by dark blue arrows. The alignment is antisymmetric with respect to the RF modulation field marked by red arrows, resulting in a push–pull operation of the MZM. (c) Schematic and simulated quasi-TE optical mode of a slot waveguide that is filled with an EO organic material. The two rails of the slot waveguide are connected to the transmission line by 60-nm-thick n-doped silicon slabs. The optical mode and the RF modulation field are both well confined to the slot region. This leads to strong EO interactions and therefore to a high modulation efficiency. (d) Chemical structure of the EO cladding material22 DLD164. The molecule consists of a ‘ring-locked’ EO chromophore core (red), and two coumarin-based pendant side chains (blue). EO, electro-optic; MZM, Mach–Zehnder modulator; RF, radio frequency; SOH, silicon–organic hybrid; SOI, silicon-on-insulator; TE, transverse electric.

Full size image

For high-modulation efficiency, an organic material featuring a large macroscopic EO coefficient r33 is essential. Besides the optical properties of the individual chromophore molecule, the macroscopic EO coefficient depends on the orientation of the chromophore ensemble with respect to the modulating RF field. After deposition of the organic material, the EO coefficient is zero (r33=0) due to the random orientation of the dipolar chromophores. For achieving a non-centrosymmetric orientation (acentric order), the material must be poled. This is achieved by heating the organic material close to its glass-transition temperature Tg while applying an external electric field to align the dipolar chromophore molecules. The material is subsequently cooled back to room temperature while maintaining the poling field, thereby conserving the acentric order of the chromophores and the macroscopic EO coefficient of the material. For poling of our device, we apply the poling voltage Upoling between the two floating ground electrodes of the RF transmission line,19 as depicted in Figure 1b. This way, half of the poling voltage drops across each silicon slot, resulting in an orientation of the dipolar chromophores (dark blue arrows), which is parallel to the RF modulation field (red arrows) in the left half of the ground–signal–ground transmission line and anti-parallel in the right half. This leads to a positive phase shift in the left arm of the MZM, and to a negative phase shift in the right arm, thereby resulting in push–pull operation of the device.

The optical fiber-to-fiber insertion loss of the modulator amounts to 16.5 dB, dominated by fiber-to-chip coupling losses of the non-optimized grating couplers28 of approximately 10.5 dB for both interfaces. The on-chip loss of the device is approximately 6 dB for maximum transmission of the modulator and is dominated by scattering losses due to sidewall roughness in the slot waveguides. A more detailed breakdown of the on-chip insertion losses can be found in Supplementary Information. A reduction of sidewall roughness by optimization of the fabrication process and shorter device lengths is expected to enable SOH MZM with on-chip insertion losses below 2 dB.

Organic material with increased EO activity

In general, the EO coefficient of the organic cladding depends on the nonlinear polarizability of a single chromophore molecule, on the volume concentration of the chromophores in the material, and on the average orientation of the molecules with respect to the applied electric field.29 This leads to an expression30 of the form

where βzzz denotes the first-order hyperpolarizability of the molecule along its dipole axis and N is the chromophore volume number density. For a single chromophore molecule, the quantity θ denotes the angle between the molecular dipole axis z and the direction of the external electric field, and the average degree of chromophore orientation is described by the ensemble average <cos3θ>, the so-called acentric order parameter. The quantity g denotes the Lorentz–Onsager local field factor12 that accounts for partial field screening in the material, and n is the refractive index. Based on Equation (1), the EO coefficient can be enhanced in three ways: first, by increasing the chromophore hyperpolarizability βzzz, second, by maximizing the chromophore density N, and third, by inducing the highest possible acentric orientation <cos3θ>.

Chromophore hyperpolarizability βzzz has continuously been increased over the last years, driven by the advancement of computational methods that allow to theoretically predict the hyperpolarizability of molecular structures.12,30,31 However, chromophore density N and average acentric orientation <cos3θ> are coupled quantities that cannot be maximized independently:12 increasing chromophore density leads to strong electrostatic interactions between the dipoles, thereby counteracting the desired acentric orientation of the ensemble. For high densities, these intermolecular interactions lead to partial crystallization of the material, to vanishing EO activity and to increased scattering loss. The conventional approach to mitigate these interactions is based on using a small chromophore load of typically less than 25 wt-% in a polymer host matrix. SOH integration was so far largely based on these guest–host systems,15,17,23,24,25,32,33 for which in-device EO coefficients were far below values achieved for parallel-plate poling of bulk reference samples. The incomplete transfer of EO coefficients from bulk materials to devices appears to be a general problem, which does not only occur in SOH devices: EO coefficients of up to 450 pm V−1 can be achieved in bulk materials,12 whereas the highest reported in-device EO coefficients amount to r33=59 pm V−1 for SOH devices26 and to r33=138 pm V−1 for an all-polymer MZM.34

In our experiments, the constraints of low in-device EO coefficients are overcome by using the novel monolithic organic material DLD164. The design of the molecule allows to simultaneously increase the chromophore density N and acentric orientation <cos3θ>. The chemical structure of DLD164 is depicted in Figure 1d and its synthesis is reported in Ref. 22 The chromophore molecules are designed to have an EO active core (marked in red) with pendant coumarin-containing side groups (marked in blue). The interaction of these side groups effectively suppresses head-to-tail orientation of the dipolar chromophores and thus inhibits partial crystallization of the material. For a more detailed discussion on this principle, the reader is referred to Refs. 22, 35 and 36. This approach allows for applying the neat chromophore as cladding material without the need for an insulating polymer matrix. Moreover, as discussed in Ref. 35, the intermolecular interaction of the coumarin-pendant side groups reduces the number of rotational degrees of freedom of the chromophores from three to two. Based on a theoretical model, this interaction increases the average acentric order <cos3θ> by a factor of 2, due to a lower poling energy required to achieve a given average chromophore orientation, see Ref. 36 for a more detailed discussion of this effect.

Results and discussion

Poling of the organic material and DC characterization

SOI waveguides are fabricated by standard CMOS processes using optical lithography, and EO materials are deposited by spin coating (see Supplementary Information for a more detailed description). The EO coefficient of the cladding after poling is studied by measuring the voltage-dependent transmission of an MZM at DC (Figure 2a). The depicted trace refers to a 1-mm-long device which was poled by a voltage of 40 V per phase modulator section. This corresponds to a remarkably high poling field of Epoling≈250 V μm−1 in the 160-nm-wide slots. The ability to withstand extraordinarily high poling fields appears to be a beneficial result of the SOH slot waveguide structure—for a bulk reference sample of monolithic organic cladding material, dielectric breakdown was already observed22 at 75 V μm−1. We attribute the improved stability to thin-film effects and to a low defect probability in the slot region. The devices are characterized at infrared telecommunication wavelengths around 1550 nm. The 1-mm-long device reveals a π-voltage of Uπ=0.5 V measured at DC under a 4 V bias. The bias avoids partial field screening by free charges in the cladding that may occur at small DC fields37 (Figure 2a). The corresponding voltage–length product amounts to UπL=0.5 V mm at an optical insertion loss of 6 dB. This is a very promising result, especially when comparing the SOH MZM to a typical carrier-depletion all-silicon modulator38 that features a π-voltage of Uπ=3.1 V at a length of 6 mm and an insertion loss of 9 dB. Taking into account the measured waveguide dimensions and the associated overlap between optical mode and RF field, we can derive the in-device EO coefficient of the cladding to be r33=180±20 pm V−1 (see Supplementary Information for details of the calculation). This coefficient clearly exceeds the maximum value of r33=137 pm V−1 previously achieved in the corresponding bulk material reference,22 where poling fields were limited by dielectric breakdown.

Figure 2
figure 2

Static characterization of SOH MZM. (a) Transmission vs. applied DC voltage at a wavelength of 1546 nm for a 1-mm-long SOH MZM in push–pull configuration. At a bias of 4 V, we find a π-voltage of Uπ=0.5 V. For zero bias, free charges in the cladding lead to partial field screening of the applied fields, which leads to increased π-voltages. (b) EO coefficient r33 vs. poling field Epoling for a monolithic chromophore cladding (DLD164, red) and for a guest–host cladding containing 25 wt-% YLD124 in a PMMA matrix (green). For weak poling fields, r33 increases linearly with the poling field, as indicated by fitted straight lines. DLD164 reveals a poling efficiency of r33/Epoling=1.17 nm2 V−2 which is significantly higher than the value of 0.23 nm2 V−2 obtained for the YLD124/PMMA guest-host cladding. Crosses mark previously reported values for chromophore-doped polymers.15,17,23,24,25,26 For the YLD124/PMMA guest–host cladding, breakdown occurs at field strengths exceeding 200 V μm−1, whereas DLD164 withstands all applied voltages, leading to an EO coefficient r33 of up to 180 pm V−1. EO, electro-optic; MZM, Mach–Zehnder modulator; SOH, silicon–organic hybrid.

Full size image

To verify that the extraordinarily high in-device EO coefficient is in agreement with theoretical predictions, we compare the poling efficiency r33/Epoling of the DLD164 cladding with that of the guest–host cladding YLD124 (25 wt-%) in PMMA (polymethyl methacrylate) using nominally identical SOI waveguide structures. In terms of molecular hyperpolarizability βzzz, the chromophore cores of DLD164 and YLD124 are identical. However, by utilizing the neat chromophore DLD164, we achieve a 2.3 times higher number density N of EO active dipoles as compared to the commonly used mixture of 25 wt-% YLD124 in PMMA (see Supplementary Information for details). For comparison of the acentric order parameters in both materials, we derive the EO coefficients r33 for various poling fields (Figure 2b). For weak poling fields, the EO coefficient increases linearly with Epoling. From the measurements, we deduce a poling efficiency of r33/Epoling=1.17 nm2 V−2 for the DLD164 material and of 0.23 nm2 V−2 for the YLD124/PMMA composite. Using Equation (1), we find that the product N<cos3θ> for DLD164 is increased by a factor of 4.8 as compared to YLD124/PMMA: The calculated quantities βzzz are equal in both cases, while the g/n4 ratio for YLD124/PMMA is larger by a factor of 1.07 compared to DLD164.22 Considering the 2.3-fold higher chromophore concentration N, we conclude that the acentric order parameter <cos3θ> for DLD164 must be a factor 2.1 times larger than that of the YLD124/PMMA composite, which is in excellent agreement with theoretical predictions.36 These findings are also consistent with independent experimental investigations of an identical chromophore concept in parallel-plate poled reference measurements,35 revealing enhancements of the acentric order parameter <cos3θ> by a factor 2.1 when using coumarin-based pendant side groups. We hence conclude that the concept of reduced dimensionality through harnessing engineered intermolecular interactions can be directly transferred from parallel-plate poled reference samples to devices and allows to significantly enhancing the performance of SOH devices.

Data transmission and energy consumption

The fabricated SOH modulators were tested in transmission experiments using a non-return-to-zero (NRZ) on–off keying format. We perform two sets of experiments, aiming at modulation with lowest possible energy consumption, or with highest possible data rate. Both experiments were performed at a wavelength of 1546 nm. For the low-energy experiments we use a 1-mm-long device operated at data rates of 12.5 Gbit s−1 and 25 Gbit s−1; for the high-speed experiments, we use also devices as short as 250 µm. The corresponding experimental set-up is sketched in Figure 3a and is explained in more detail in the Supplementary Information along with the models that are used for estimating the dissipated energy per bit.

Figure 3
figure 3

Data generation and reception at lowest energy consumption. (a) Experimental set-up for transmission experiments. A PPG is used to drive the modulator without additional RF amplifier. The drive signal is fed to the DUT by using a GSG RF probe. A bias-T is used to adjust the operating point. A second RF probe at the end of the modulator can be used to terminate the device. Light from an external-cavity laser is coupled to the MZM using grating couplers. The modulated light is amplified by an EDFA and detected by a DCA and a BERT. (b) Measured BER as a function of drive voltage.27 For transmission at 25 Gbit s−1 (magenta), the DUT is always terminated by a 50 Ω load, whereas for a rate of 12.5 Gbit s−1, the device is driven with (green) and without (blue) termination. Data points in the gray-shaded areas were measured at the quadrature operating point of the device. (c) Corresponding energy per bit for the various drive voltages27 of (b). (d) Equivalent circuit of the terminated traveling-wave modulator along with the corresponding eye diagrams measured at the quadrature operating point at 25 Gbit s−1 (Udrive=430 mVpp, Wbit=38 fJ) and at 12.5 Gbit s−1 (Udrive=270 mVpp, Wbit=30 fJ). The internal resistance of the source and the terminating impedance amount to Rs=RL=50 Ω. (e) Eye diagrams at 12.5 Gbit s−1 of a non-terminated device27 along with the respective equivalent lumped-element circuit. The right inset shows an eye diagram at Udrive=310 mVpp (Wbit=12 fJ) when operating the MZM at the quadrature point. The left inset shows an eye diagram at Udrive=120 mVpp (Wbit=1.8 fJ) when operating below the quadrature point. BER, bit-error ratio; BERT, bit-error-ratio tester; DCA, digital communications analyzer; DUT, device under test; EDFA, erbium-doped fiber amplifier; FEC, forward-error correction; GSG, ground–signal–ground; MZM, Mach–Zehnder modulator; OOK, on-off keying; PPG, pseudo-random pattern generator; RF, radio frequency.

Full size image

For the low-energy experiments, we operate the MZM in different driving modes and measure the BER along with the energy consumption per bit as a function of drive voltage (Figure 3b and 3c). For high data rates beyond 20 Gbit s−1, the 1-mm-long device acts as a travelling-wave modulator and the transmission line needs to be terminated by a matched load resistor of RL=50 Ω to avoid detrimental back-reflections of the RF wave (Figure 3d). In this case, the RF power is dissipated along the lossy transmission line and in the termination. For a mean-free drive signal switching between −Udrive/2 and +Udrive/2, the energy per bit Wbit is obtained by dividing the dissipated average power by the data rate r, Wbit=(Udrive/2)2/(RL×r).

For data rates below 20 Gbit s−1, the modulator is much shorter than the RF wavelengths, and the device acts as a ‘lumped’ element. In this case, the MZM may be operated without termination (Figure 3e). This significantly decreases the energy consumption, since the drive signal is no longer dissipated in a 50-Ω termination.39 Instead, energy consumption is now dominated by ohmic losses that occur in the internal resistor Rs of the source and in the series resistor Rd of the device during charging and de-charging of the capacitor1 Cd (see Figure 3e for an equivalent-circuit model). For a terminated device, the drive voltage Udrive is half the open-circuit voltage U0 of the 50-Ω source, whereas for the non-terminated device, the drive voltage Udrive can approach the full open-circuit voltage (see equivalent circuit in Figure 3e). This effect must be taken into account when calculating the drive voltage and the associated energy consumption of the terminated and non-terminated device. Moreover, the device capacitance Cd is an important parameter for estimating the energy consumption of the unterminated modulator. The device capacitance can be measured using a vector network analyzer, leading, e.g., to a value of Cd=(400±20) fF for a 1-mm-long device (see Supplementary Information for a more detailed description of the measurement). The measured results are in good agreement with the values predicted by electrostatic simulations (CST Microwave Studio). In general, the device capacitance Cd is a frequency-dependent quantity due to its dispersive dielectric. When calculating the energy consumption, this can be taken into account by integration over the dissipated power spectrum in the frequency domain (see Supplementary Information).

For small data rates far below the 3 dB cutoff frequency of the modulator, power dissipation associated with charging/de-charging the capacitor Cd from an initial voltage 0 V to a final voltage Udrive=U0 and vice versa is independent of the series resistance and amounts to1 . When driving a modulator with a digital NRZ signal, we need to take into account that charging and de-charging only occurs during 1–0 and 0–1 transitions, but not for 0–0 and 1–1 sequences. On average, charging or de-charging of the capacitor occurs every second bit, and the energy consumption per bit hence amounts to1 Wbit=. For data rates r approaching the 3 dB cutoff frequency of the modulator, however, the device capacitor is no longer fully charged within the timeslot of one symbol, and the energy consumption can no longer be calculated by the simple expression Wbit=. Instead, the dissipated power has to be computed for each frequency component of the NRZ power spectrum separately. The total dissipated power is then obtained by integrating over the corresponding power spectrum (see Supplementary Information for a more detailed description). For a frequency-independent device capacitance Cd, this leads to an expression of the form

In this expression, T=1/r denotes the symbol duration, τ=(Rs+Rd)Cd is the resistance-capacitance time constant of the circuit, and Rs and Rd are the series resistances of the source and of the device, as defined in Figure 3e. This model can be extended to the case of a frequency-dependent capacitance Cd (see Supplementary Information). In addition to the low-frequency approximation and the exact frequency-domain model, we also investigate a slightly simplified method based on a time-domain approximation. We find good agreement between the various methods for estimating the energy consumption of our devices and hence conclude that the results are reliable (see Supplementary Information).

For transmission at 12.5 Gbit s−1, we measure comparable BER for the terminated and non-terminated device at small drive voltages (Figure 3b and 3c). The associated energy consumption, however, is significantly lower for the non-terminated device. A drive voltage of only 80 mVpp is required to keep the measured BER below the hard-decision forward-error correction threshold40 of 4.5×10−3, corresponding to an energy consumption of 0.7 fJ bit−1. For remaining at a BER below 10−9, we find a drive voltage of Udrive=460 mVpp (27 fJ bit−1) for the non-terminated MZM and Udrive=300 mVpp (40 fJ bit−1) for the terminated MZM. These energy consumptions are one to two orders of magnitude below the 200 fJ bit−1 which were previously demonstrated for a 5-mm-long silicon-based MZM7 operated at 630 mVpp. The performance of our MZM is even comparable to best-in-class resonant structures,2,8 where 3 fJ bit−1 were reported at drive voltages of 1 Vpp. For transmission at 25 Gbit s−1, slightly worse BER are measured, and the energy consumption amounts to 7 fJ bit−1 (190 mVpp) at a BER of 4.5×10−3 and to 52 fJ bit−1 for a BER of 10−9. These figures do not include the energy consumption of any external components, such as laser, optical amplifier and electrical drivers. Currents flowing due to DC voltages are measured to be below 2 nA, contributing only a negligible amount of energy (< 1 aJ bit−1) to the budget.

For the high-speed transmission experiments at data rates of up to 40 Gbit s−1, we use 500-µm-long and 1-mm-long terminated devices and a 250-µm-long non-terminated device. For the 1-mm-long device, optimum signal quality at 12.5 Gbit s−1 was obtained for a peak-to-peak drive voltage swing of Udrive=950 mVpp measured at the output of the pulse-pattern generator. This voltage was also used for the other data rates to facilitate comparison. Note that, due to slight overmodulation and due to the frequency response of the device and the deployed RF components, the drive voltage swing is bigger than 500 mVpp that would be expected from the DC voltage-length product of UπL=0.5 Vmm. Eye diagrams for various data rates are depicted in Figure 4a. The extinction ratios exceed 10 dB in all cases. At 12.5 Gbit s−1 and 30 Gbit s−1, we measure excellent Q2-factors of 22 dB and 19 dB, respectively, and obtain error-free operation with BER values below 10−12. At 40 Gbit s−1, we measure a Q2 of 15 dB and a reasonably low BER of 1×10−8. We demonstrate that 40 Gbit s−1 data rates can also be generated with 500-µm-long terminated MZM and with 250-µm-long non-terminated MZM, which were operated with drive voltages of Udrive=2 Vpp and Udrive=3.4 Vpp, respectively. These findings demonstrate that SOH modulators support data rates comparable to those of carrier-depletion p–n modulators while featuring small device footprints comparable to those of carrier-injection p–i–n modulators.

Figure 4
figure 4

High-speed modulation. (a) NRZ OOK eye diagrams27 of a 1-mm-long modulator for data rates of 12.5 Gbit s−1, 30 Gbit s−1 and 40 Gbit s−1. The modulator is biased at the quadrature point. The drive voltage is 950 mVpp in all cases. The ERs exceed 10 dB even at 40 Gbit s−1. Up to 35 Gbit s−1, the BER is below 10−12. At 40 Gbit s−1, we measure a low BER of 1×10−8. (b) 40 Gbit s−1 eye diagrams of a 500-μm-long terminated modulator (Udrive=2 Vpp) and of a 250-µm-long modulator without termination (Udrive=3.4 Vpp). Measured Q2 factors, ER and BER are denoted in the respective figures. The pseudo-random binary sequence length is 231−1. BER, bit-error ratio; ER, extinction ratio; MZM, Mach-Zehnder modulator; NRZ, non-return-to-zero; OOK, on–off keying.

Full size image

Conclusions

We have shown that SOH integration offers the potential to implement EO devices with unprecedented performance while maintaining many of the advantages of standardized large-scale CMOS processing. We experimentally demonstrate non-resonant EO modulation at 12.5 Gbit s−1 with a peak-to-peak voltage swing of 80 mV and an energy consumption as small as 0.7 fJ bit−1. The in-device EO coefficient of the organic cladding amounts to r33=180±20 pm V−1 and is significantly larger than the values that can be achieved by parallel-plate poling of the corresponding bulk material. Moreover, our SOH modulators feature small footprint and are capable of operating at 40 Gbit s−1. Our findings demonstrate that SOH integration allows overcoming intrinsic limitations of all-silicon devices: the concept combines low UπL products comparable to those of carrier-injection devices with large EO bandwidths known from carrier-depletion devices, while the energy consumption is less by at least one order of magnitude. We believe that there is still significant room for enhancing the performance of SOH modulators, e.g., by taking advantage of the continuously improving EO cladding materials.

One important aspect of organic EO materials is the long-term stability, both regarding thermal and photo-induced degradation. Within the experiments of this paper, we have not encountered any photo-degradation or bleaching effects, even when operating the device at input powers of more than 15 dBm, which corresponds to an intensity of 3×106 W cm−2 in the slot of the SOH modulators. At higher power levels, thermal effects were already beginning to impair the performance of the device. We hence believe that material stability in our device is governed exclusively by thermal reorientation of chromophores. The glass transition temperature of the currently used DLD164 amounts to Tg=66 °C, and thermally activated re-orientation of the chromophores reduces the lifetime of the current devices to roughly a month under ambient laboratory conditions. However, we have recently demonstrated that electro-optic coefficients of up to 230 pm V−1 can also be achieved by other material systems such as binary-chromophore organic glasses that combine shape-engineered dendritic molecular structures with conventional bare donor–acceptor species.37 These materials feature glass transition temperatures of more than 100 °C and are expected to exhibit better thermal stability. In addition, we expect that stability can be significantly improved by synthetically modified DLD164 chromophores that bear specific crosslinking agents for post-poling lattice hardening. The viability of this approach has already been demonstrated for similar classes41,42 of EO compounds, with materials being stable for temperatures of up to 250 °C. It has been demonstrated experimentally that these techniques can be applied to realize temperature-stable all-polymeric MZM.34 Whereas earlier demonstrations of crosslinking have led to increased optical losses by absorption or scattering in the organic material, it has recently been demonstrated that this problem can be overcome by utilizing cycloaddition reactions (Diels–Alder and related cycloadditions) that do not produce chemical side products or lead to lattice distortion. In these experiments, no increase in loss has been observed in the near infrared,43 and the values obtained for the glass transition temperatures after crosslinking are above 200 °C.44 For DLD164, the addition of crosslinking agents to the side groups can be realized without affecting the EO activity of the chromophore or the coumarin–coumarin interaction of the side chains. Thermal stability and photo-degradation of SOH devices is subject to ongoing research.