1. Introduction

As clock speeds in computing machines increase, metal wire interconnects become inadequate in terms of power consumption and wire density and optical solutions are envisioned. Silicon photonics in particular is expected to provide an exceptional platform for chip-level interconnect technology [1–3]. The power budget required is projected to be sub-pJ per bit for an optical link [4,5]. Another constraint for the optical approach is the need for compactness to conserve silicon chip area. In order to provide high bandwidth and low power consumption, wavelength division multiplexing (WDM) is required, as proposed for several on-chip optical networks [5–8]. Low power and compact multiplexing filters are perceived to be one of the most critical components and are yet to be demonstrated with acceptable performance. There are at least three major challenges to achieve high performance silicon WDM filters in optical interconnects. First, in order to achieve compactness, submicron silicon waveguides, which can have bending radii down to a few microns, are commonly used in silicon WDM filters and these waveguides are very sensitive to fabrication tolerance since the effective index and group index have high variation with dimensional errors from fabrication. For example, 1 nm of dimensional error in waveguide cross section in a ring resonator may result in a wavelength shift of ~100 GHz for a waveguide geometry of ~0.5 µm by ~0.25 µm [9]. Such index variations also increase the channel crosstalk in WDM filters based on gratings or Mach-Zehnder interferometers (MZIs). Second, silicon has a relatively high thermo-optic coefficient (1.86x10⁻⁴ /°C), which results in a resonance shift of about 0.1 nm/°C (this occurs for all types of WDM filters). Third, even with the use of submicron silicon waveguides, previously demonstrated WDM filters based on arrayed waveguide gratings and echelle gratings have an area usually >5000 µm² per channel [10–12]. Increasing the channel number may significantly increase the device footprint since the areas for these grating-based devices are usually proportional to the square of channel number. Significant reduction of filter area is required for inter- and intra-chip optical interconnects [4].

Cascaded thermally tunable microcavities may address all of these challenges [13]. Large tunable ranges can be realized by utilizing silicon’s thermo-optic effect, which can compensate resonance variation from both fabrication and environmental temperature fluctuations. Silicon microcavities can have a bending radius down to a few microns, resulting in a channel area possibly less than 100 µm². In addition, as the resonant wavelength can be tuned individually for each microcavity, the filters can be configurable in both wavelength and channel spacing. Another important advantage is that each microcavity can have independent add-drop functionality and therefore be flexibly positioned on chips. The key challenge, however, is to reduce tuning powers. Previously demonstrated thermally tunable microcavities usually have relatively large tuning powers of 20 mW-100 mW to achieve one free spectral range (FSR) [13–19]. For a transmission data rate of 25 Gbps, this thermal tuning power will add up to 1-4 pJ per bit in an optical link. Recently, it was demonstrated that the tuning power can be as low as 2.4 mW per FSR using free-standing silicon microcavities [20,21]. With this amount of power, the energy consumption is about 10-100 fJ/bit for a data rate of 25 Gbps, depending on how large a tuning range is required. These results further validate the use of silicon microcavities in chip-level optical interconnects.

In this paper, we report a 1x4 reconfigurable demultiplexing filter based on cascaded thermally tunable silicon microcavities with ultralow tuning powers. The use of free-standing silicon microcavities with undercut structures significantly reduces the tuning power, with a figure of ~2.9 mW per free spectral range. Even with thermal crosstalk between two adjacent resonators, we demonstrate multiplexing functionality for channel spacings of 200 GHz, 100 GHz, and 50 GHz, with channel wavelengths aligned to International Telecommunication Union (ITU) grid specifications. The achieved channel crosstalk values for 200 GHz and 50 GHz channel spacings are less than −20 dB and −11.5 dB, respectively.

2. Device structure and fabrication

(De)Multiplexing filters can be constructed by cascading silicon microcavities with one common bus waveguide but with individual drop waveguides. Our thermally controlled resonators are realized by putting micro-heaters on top of the resonator cladding [13]. In our previous paper [21], we have presented the design and fabrication of thermally tunable single microcavities with undercut beneath the cavities. Due to excellent thermal isolation by the air gap between the microcavities and substrate silicon, the heating efficiency can be improved about 10 times, compared with devices without undercuts [21]. As shown in Fig. 1 , we have fabricated a four channel WDM filter with cascaded racetrack resonators with undercut structures. In these devices, the resonators have a racetrack shape with straight coupling lengths of 11 µm and bending radii of 10 µm. The edge-to-edge distance between neighboring resonators is 13 µm. There are always concerns about the mechanical stability of the suspended structures. It is seen that the suspended membranes are supported by oxide beams to avoid bending. In general, the thicker the suspended membranes, the more stable the structures in terms of mechanical stability. Our suspended membranes have a total thickness of about 5 µm [21], and very few devices are found broken after full fabrication and polishing. The optical waveguides are designed to be protected by the upper oxide cladding and buried oxide [21]. However, the stress accumulated during the film depositions and undercut release may degrade the optical performance of the micro-resonators.

 

Fig. 1 Tilted top-view SEM for a fully fabricated WDM filter using free-standing resonators.

Download Full Size | PDF

3. Heater efficiency and thermal crosstalk

We characterize the heater efficiency and thermal crosstalk by measuring the drop-port transmission spectra with different heating powers. As we apply heating powers on one of the four resonators, thermal crosstalk may induce resonance red shifts in adjacent resonators. We can define C_ij (i,j = 1-4) as the resonance shift of resonator j with unit heating power applied on resonator i. Where C_ii represents the heating efficiency of heater i, C_ij (i ≠ j) presents the information on the thermal crosstalk between resonator i and resonator j. From spectra such as those shown in Fig. 2(a) –2(d), we can calculate C_ij from the resonance shift as a function of power. We list the measured results in Table 1 . The thermal tuning efficiency is ranged from 2.30 to 2.43 nm/mW. Since the free spectral range is 6.9 nm, the tuning power is on average 2.9 mW per FSR tuning. The thermal crosstalk between two adjacent resonators is about 13%, and it is negligible between resonators not next to each other (i.e., thermal crosstalk is negligible if the edge-to-edge distance between two resonators is more than 57 µm). For easy thermal control, it would be better if there was no thermal crosstalk since every microcavity can be independently tuned. Compared to thermal crosstalk without undercut structures, as reported in Ref [13], undercuts actually increase thermal crosstalk between adjacent resonators. Removing the silicon substrate underneath the resonator forces heating flux in the free standing membranes, resulting in larger thermal crosstalk. However, in our opinion, this should not be a limiting factor to achieving compact WDM filters using free standing microcavities. In chip-level optical interconnects, each add-drop resonator may locate close to its own transmitter/receiver site, therefore far away from each other to avoid thermal crosstalk.

 

Fig. 2 Drop port spectra for resonators 1-4 (a-d) with various heating powers on resonator 1.

Download Full Size | PDF

Table 1. Measured thermal efficiency

View Table

4. Demonstration of reconfigurable WDM filter

We demonstrate reconfigurable multiplexing functionality by aligning the four resonant wavelengths to ITU grid wavelengths. Four source-meters are used to drive four heaters. With the heating efficiency from the last section, the resonance shift can be expressed by

As an example, we present spectra before tuning and after tuning to achieve 200 GHz channel spacing and ITU channels 45-51 in Fig. 3 . The calculated powers from Eq. (2) are (0.30, 0.14, 0.61, 0.62) mW and the experimentally applied powers are (0.32, 0.15, 0.66, 0.71) mW. They agree reasonably well. Total power to achieve this WDM filtering functionality is 1.84 mW. The drop spectra in Fig. 3 are normalized to the maximal power at the through port. The transmissions at resonance approximately represent device insertion loss from the resonators themselves, which are 1.5-2.5 dB for four resonators. The quality factors are 9000 for the first three resonators but 12000 for the fourth resonator. The quality factor variation may come from the gap variation between the bus waveguide and resonators. The worst-case optical crosstalk among four resonators is less than −20 dB, as shown in Fig. 3(b). Reconfigurablity is demonstrated by aligning the four resonances to ITU grid wavelengths with channel spacings of 100 GHz and 50 GHz, as shown in Fig. 4 . For 100 GHz, the chosen wavelengths are channels ITU 37 – ITU 40. For 50 GHz, the chosen wavelengths are ITU 38 – ITU 39, and their 50 GHz offset wavelengths. The total powers applied are 2.4 mW and 2.16 mW for Fig. 4(a) and 4(b), respectively. As we decrease the channel spacing, the optical crosstalk between different channels increases. We achieve <-16 dB and <-11.5 dB crosstalk for 100 GHz and 50 GHz, respectively. Design of narrower bandwidth resonators would reduce the optical crosstalk further. However, this limits the data rates of transmitted signals. Careful trade-off between resonator bandwidth and channel crosstalk needs to be considered in applications. Using high-order coupled-ring resonators can reduce the channel crosstalk, at the expense of higher tuning power and larger device area [22].

 

Fig. 3 (a) Drop port spectra before tuning. (b) Drop port spectra after thermal tuning to achieve 200 GHz channel spacing.

Download Full Size | PDF

 

Fig. 4 Drop port spectra after thermal tuning to achieve 100 GHz (a) and 50 GHz (b) channel spacing.

Download Full Size | PDF

5. Discussion and conclusion

In this paper, we demonstrate the feasibility of reconfigurable WDM filtering functionality using cascaded thermally tunable microcavities. We solve the key challenge of tuning power by using free-standing silicon resonators with undercuts. The low thermal tuning power is realized due to the excellent thermal isolation from the air gap between resonators and silicon substrates. As presented in Ref. [21], the tuning time is about ~170 µs, which may be fast enough for tuning the wavelength. Considering the achieved low tuning power of 2.9 mW per FSR for an individual resonator, it would add a maximum power of 11.6 mW to realize fully reconfigurable WDM filters. The demonstrated WDM filters have advantages which include compactness, reconfigurablility, and the ability to compensate resonance variations from both fabrication and temperature variations. However, some challenges still remain, such as closed-loop control of resonances. We note that a few attempts have been made to solve this problem, by either using on-chip temperature sensors [23] or detecting scattering light from ring resonators [24]. Further optimization of thermal control techniques without costing too much power is necessary. Nevertheless, the demonstrated low power/energy consumption in this paper validates the use of silicon microcavities to realize WDM filters in chip-level optical interconnects.