Nano–opto-electro-mechanical switches operated at CMOS-level voltages

Combining reprogrammable optical networks with complementary metal-oxide semiconductor (CMOS) electronics is expected to provide a platform for technological developments in on-chip integrated optoelectronics. We demonstrate how opto-electro-mechanical effects in micrometer-scale hybrid photonic-plasmonic structures enable light switching under CMOS voltages and low optical losses (0.1 decibel). Rapid (for example, tens of nanoseconds) switching is achieved by an electrostatic, nanometer-scale perturbation of a thin, and thus low-mass, gold membrane that forms an air-gap hybrid photonic-plasmonic waveguide. Confinement of the plasmonic portion of the light to the variable-height air gap yields a strong opto-electro-mechanical effect, while photonic confinement of the rest of the light minimizes optical losses. The demonstrated hybrid architecture provides a route to develop applications for CMOS-integrated, reprogrammable optical systems such as optical neural networks for deep learning.

Electrically reconfigurable photonic netE works have the potential to enable technological advances in many fields such as optical neural networks used to process information with low power at the speed of light (1), optical metrology to feed multiple sensors with a single light source (2), all-optical routing to avoid the current bottleneck of optical-electrical-optical conversion (3), and integrated quantum optical circuits (4). However, to make such reconfigurable photonic networks practical, they need to be up-scaled into large circuits and co-integrated with complementary metal-oxide semiconductor (CMOS) electronics. To achieve this level of scaling and integration, the elementary electro-opticalswitch unit needs to feature compact footprints (~1 mm2), CMOS driving voltages (~1 V), short switching times (~1 ns), low optical losses (≤0.1 dB), and low power consumption (>>1 mW) (5).

Electro-optical switches typically rely on interferometric waveguide configurations to divert light to different outputs by means of constructive or destructive interference. This is achieved by changing the refractive index (Dn) of the waveguide material. State-of-the-art networks control Dn by the electro-thermo-optical effect (6); however, the milliwatt power consumption per switch limits scalability of this approach (3). Furthermore, all-optical (7), phasechange (8, 9), and electro-optical (10–13) switching approaches show notable results (e.g., compact footprint or low loss) but struggle to excel in all requirements simultaneously. For instance, under a CMOS driving voltage of 1 V, electro-optical materials currently yield Dn > 10−2, requiring device lengths of >100 mm to achieve full switching (10, 13). Resonant approaches reduce the footprint by leveraging the high finesse of micrometer-sized cavities. Yet, the frequency tunability of lowest-loss resonators is ≤50 GHz/V, limiting the switches’ optical bandwidth (13, 14). Moreover, power-hungry stabilization is required because similar resonance frequency shifts occur for single-kelvin temperature fluctuations (15).

Opto-electro-mechanical (OEM) switches provide an alternative way to control the flow of light by mechanically changing the waveguide geometry rather than modulating the material’s intrinsic refractive index (16), with waveguide motions leading to local Dn on the order of unity. Because of the strong Dn, small actuations suffice to induce large effective refractive index changes (Dneff). Importantly, OEM switches consume negligible amounts of energy in stand-by, because the mechanical geometry is controlled by electrostatic forces that are not accompanied by static currents. Photonic OEM devices have been switched by actuating hundreds of nanometer-scale gaps between two silicon waveguides, using a remote electro-mechanical driver (17, 18). The all-photonic approach yield slow optical losses(>0.1dB), whereas the large gap size requires high driving voltages (>10 V). By contrast, the subwavelength confinement of light (19) in all-plasmonic devices enables stronger OEM responses, which reduces the drive voltages (20). All-plasmonic switches utilize two metal surfaces to form a tens-of nanometers-wide gap, where light is confined and its phase is modulated by electro-mechanical actuation of the gap width (21). However, such confinement comes with considerable metal induced optical losses (~1 dB/mm) (22), which have limited the realization of large-scale plasmonic switching networks.

Here, we introduce a hybrid photonicplasmonic (HPP) OEM technology that benefits a strong plasmonic OEM-effect to fully switch light with a CMOS-level voltage (≈1.4 V), low optical losses (0.1 dB), and a compact footprint (≈10 mm2).

Figure 1A illustrates the dynamic routing of light by two nano-OEM (NOEM) switches, respectively biased to two different resonance states (wavelength lres): drop state (foreground device, 0 V) and through state (background device, 1 V). Incident light (wavelength l0) guided in the through port is transmitted (lres ≠ l0) or dropped (lres = l0) depending on the individual, bias-dependent lres of the encountered resonant switches. The HPP resonator comprises a thin gold membrane partially suspended above a silicon disc forming an air gap (z0) (see Fig. 1B) (23). The air HPP waveguide combines low-loss propagation in the silicon waveguide with strong field enhancement at the metal surface in the gap (24, 25). Additionally, gold and silicon form an air capacitor able to actuate z0 by means of an electrostatic force generated by an applied voltage (Vdrive). The gold membrane bending (dz) induces a resonance shift (Dlres) by changing the mode index (Dneff) (Fig. 1C).

Figure 1D indicates that the large tunability of the resonance wavelength and dz as small as 4 nm already provide Dlres larger than the resonance’s loaded full-width half-maximum (FWHM). Low-loss coupling to the drop port requires that the waveguide-resonator coupling rates are larger than the resonator’s plasmonic loss rate (e.g., intrinsic FWHM) (fig. S3) (25). Thus, the switching efficiency is limited by the ratio of Dlres to the intrinsic FWHM. Dlres exceeds the FWHM by more than an order of magnitude when reducing z0 (Fig. 1E). A tunability of >1 THz/V (>10 nm/V) and a subnanometer FWHM is achieved for z0 ≈ 35 nm, enabling a large extinction ratio (ER) and low-loss switching (25).

This strong tuning can be understood by separating the OEM effect into its two subprocesses. First, the opto-mechanical coupling (GOM º dlres/dz) increases for decreasing gaps because of the plasmonic confinement of light to the gap (fig. S2) (25). Thus, more light experiences the strong Dn between air and metals upon actuation (26). The goldÕs skin depth for infrared light is ≈25 nm; thus, thin and low-mass membranes suffice as high reflectors to concentrate light in the gap. Second, the electro-mechanical coupling (GEM = dz/dV) reaches large values because the voltage, which is applied over nanometer-scaled gaps, 2) induces strong electrostatic forces (º1/z0 (fig. S4) (25).

Furthermore, the dynamics of the NOEM switch are determined by its geometrical parameters similar to those of a ruler that extends beyond the edge of a table. Shorter suspensions (i.e., stiffer spring) and a lighter mass result in faster ruler oscillations (fres). Here, we make the overhang as short and thin as possible. The combination of small moving mass, large forces, and small mechanical quality (Q) factors enables tens of nanosecond switching at CMOS driving voltages.

The fabricated resonators are shown in Fig. 2. The drop port was omitted to probe the resonatorÕs intrinsic OEM properties. Vertical HPP waveguide geometries were uniformly created by depositing and selectively removing a sacrificial alumina layer, by wet-etching to a typical undercut value of ≈1.1 mm. Here, atomic layer deposition provides z0 with atomic-level precision. The critical feature size is the lateral waveguide-disc separation (w > 120 nm), which is achievable with low-cost photolithography.

The cavityÕs intrinsic Q factor (º1/FWHM) was measured by varying w (Fig. 2C). At critical coupling, Qintrinsic = 2·Qloaded ≈ 7000, which translates to propagation lengths and losses of 395 ± 70 mm and 0.01 ± 0.002 dB/mm, respectively (fig. S6) (25). The reason for such low plasmonic losses is multifold. First, ohmic losses are proportional to the fraction of the optical-mode energy penetrating the metal. This fraction drops with decreasing permittivity of the gap dielectric; air or vacuum provides a low dielectric permittivity, minimizing loss (fig. S7) (25). Second, excess losses induced by typical adhesion layers for gold (27) are minimized as air exposure oxidizes the 2-nm titanium adhesion layer used here. Third, the interaction of the HPP mode with the gold membrane is mostly restricted to the smooth metal surface facing the gap, reducing scattering losses (28).

Characterization of the switching capability (see Fig. 2D) yields a Dlres > 6 nm, which equals five times the FWHM. The nonlinear red shift is expected from electro-mechanical effects as the growing proximity of the metal membrane increases Dneff (25). Furthermore, the nonlinear dependence allows one to enhance the voltage sensitivity of the resonance shift (Dlres/V) to ≈10 nm/V (i.e., ≈1.25 THz/V) by biasing the device with 1 V (25). The demonstrated tuning capability enables the compensation of thermally induced shifts in lres, which are typically hundreds of picometers per kelvin (14, 29). The large values of GOM, GEM, and Q factors allow reduction of the required actuation distance to a few nanometers and, correspondingly, the switching time to tens of nanoseconds (see Fig. 3).

The suspended membrane features a fres of ≈12 MHz. The small mechanical Q factor and the roll-off in modulation at lower frequencies is attributed to squeeze-film damping and stiffening; for example, air compression increases the stiffness at higher frequencies and smaller gaps and thus reduces the actuation (25). This effect can be overcome by combining vacuum packaging and advanced driving signals (30). Figure 3B shows a two-step driving scheme, in which the applied drive voltages (I) and (III) exceeded the steady-state voltages (II) and (IV) at the start of the individual on- and offswitching pulses. This resulted in rise and fall times of 60 and 100 ns, respectively, where the difference is due to electrostatic actuation forces that are larger than the spring restoring forces. The devices were switched with megahertz frequencies over hours (billions of switching cycles) without degradation of the signal quality. The estimated electrical power consumption was ≈600 and 12 nW for a peak-to-peak driving voltage of 1.4 and 0.2 V, respectively (25). Further optimization could yield fall and rise times approaching ≈10 ns (fig. S5) (25).

Subsequently, we performed 1-by-2 switching experiments (see Fig. 4A). The through-and drop-port transmission spectra are plotted in Fig. 4B. Coupling the resonator to the drop port broadens the FWHM (optical bandwidth) from ≈1 nm (125 GHz) to ≈2.5 nm (350 GHz) (see Fig. 4B). Still, a 1.4-V driving voltage yielded a Dlres (≈6.2 nm) that exceeded multiple FWHMs. This enabled light routing with a cross-talk below Ð15 dB, drop-port insertion loss (ILD) of ≈2 dB, and through-port insertion loss (ILT) of ≈0.1 dB (fig. S10) (25). Optimization may further reduce ILT to ≈0.01 dB and ILD to ≤0.5 dB (fig. S9) (25). This loss asymmetry is ideal for applications in N-by-N cross-switching grids, where N is any number, as envisioned in Fig. 4D. Because light experiences ILD only once when propagating through the network, accumulated losses (i.e., path of the red light beam) are dominated by through-port loss (i.e., overall loss ≤ 2N·ILT + ILD). This would result in an average loss-toport count ratio of 0.12 dB per port for an optimized 15-by-15 network.

We present devices that challenge the common presumption that opto-electro-mechanics is a slow and bulky technology that requires high driving voltages. We demonstrate NOEM switches whose distinct compactness paves the way for high-density optical switch fabrics that are directly cointegrated with CMOS driving circuits. For instance, 200 switches and their electrical drivers could be integrated on an area as small as the cross section of a single human hair. Beyond that, the strong OEM interaction and low-loss could enable nonresonant functional units such as phase shifters and intensity modulators for purposes such as light detection and ranging (LIDAR) applications. The performance of phase shifters is typically evaluated by the voltage-length product VpL, which states the minimal combination of p phase-shift voltage times device length. The HPP prototypes demonstrated here already feature a VpL = 27 ± 4 V·mm, which, in combination with low propagation losses (a = 0.026 ± 0.006 dB/mm), represents a substantial improvement over the state-of the-art electro-optical switches (fig. S1) (25). These switches could form the building blocks of optical field–programmable gate arrays and trigger a technological revolution similar to the one enabled over the past few decades by electrical field–programmable gate arrays.

We thank C. Hierold, J. Strait, M. Davanco, J. A. Liddle, N. Zhitenev, and U. Drechsler for valuable discussions. Funding: ERC grant PLASILOR (640478). C.H. acknowledges support under a Cooperative Research Agreement between the University of Maryland and the National Institute of Standards and Technology Physical Measurement Laboratory, award 70NANB14H209, through the University of Maryland. C.H. acknowledges support by the Hans-Eggenberger Foundation. M.B. acknowledges funding from the SNSF Ambizione grant (173996). Author contributions: C.H. conceived the concept and supervised the project. C.H., A.J., F.M., and D.C. designed the switch and performed numerical optimization. C.H., A.J., and Y.F. fabricated the modulator and developed the required process technology. C.H., A.H., M.D., M.B., M.M., and V.A.A. designed and performed the experiments. All authors discussed and analyzed the data. C.H., A.J., M.M., H.J.L., V.A.A., and J.L. wrote the manuscript. Competing interests: J.L. is involved in activities toward commercializing high-speed plasmonic modulators at Polariton Technologies Ltd. Part of the work is subject to a patent application by C.H., V.A.A., H.J.L., D.C., J.L., M.M., and A.J. The remaining authors declare no competing interests. Data and materials availability: All data are available in the manuscript or the supplementary materials.

science.sciencemag.org/content/366/6467/860/suppl/DC1

Materials and Methods

Supplementary Text

Figs. S1 to S10

Table S1

References (31–64)

• 5 August 2019; accepted 21 October 2019
• 10.1126/science.aay8645

DIAGRAM: Fig. 1. Operating principle of plasmonic NOEM networks. (A) Incident light guided in the through port is switched to a drop port if its wavelength (l0) matches the node’s resonance wavelength (lres), whereas off-resonance (lres ≠ l0) light continues along the waveguide and bypasses the plasmonic resonator, thereby avoiding ohmic losses (29). (B) HPP disc resonators (radius 2 mm) are formed by a thin gold membrane suspended above a silicon disc forming a gap (z0). (C) Doped silicon and gold bridges are used to apply a voltage across the gap, thus inducing an electrostatic force that bends the membrane and prevents light from coupling to the resonator. (D) Through-port spectra for various dz. (E) Calculations (25) show that Dlres increasingly exceeds the intrinsic FWHM when reducing z0.

DIAGRAM: Fig. 2. False-colored scanning electron microscopy images and measured device properties. (A) Perspective view and transmission spectrum. The small cavity volume results in a free spectral range (FSR) of 45 nm. (B) Focused–ion beam cross section. Air gaps (z0) of 35 or 55 nm have been realized. Gap length, 600 nm. The inset shows a simulated optical field, which is strongest in the gap. Enorm, absolute value of the electric field; a.u., arbitrary units. (C) ER (blue triangles) and loaded Q factor (red circles) versus waveguide-disc separation (w) for z0 ≈ 55 nm. The ER peaks at ≈200 nm, indicating critical coupling. (D) Dlres (blue to green) and FWHM (red to yellow) as a function of voltage for z0 ≈ 35 nm. The inset illustrates complete optical switching with a 200-mV difference. For (C) and (D), the 95% confidence intervals are approximately equal to the symbol size.

DIAGRAM: Fig. 3. Time dynamics. (A) Modulation response for a sinusoidal driving signal. The inset shows the mode shape of the fundamental mechanical eigenfrequency. k, thousand; M, million. (B) Utilizing more complex driving signals (red) enables optical (blue) rise and fall times on the order of tens of nanoseconds. The optical contrast between on and off state exceeds 90%.

DIAGRAM: Fig. 4. Performance of 1-by-2 NOEMS. (A) Perspective false-colored scanning electron microscopy image of two fabricated NOEMS. GND, electrical ground. (B) Measured power spectrum of light coupled to the through port (blue) and the drop port (red) under 0 V (solid) and 1.4 V (dashed) bias. (C) Through-port (blue circles) and drop-port (red crosses) transmittance over voltage. (D) Low through-port losses are beneficial for switching architectures such as cross-grid networks envisioned here, where light (various rainbow colors) only needs to be switched once to a drop port while propagating through a 15-by-15 network. In (C), the 95% confidence interval is smaller than the symbol size. l0, probing wavelength.