Introduction

Devices that have the ability to control multiple optical modes at high speeds are revolutionizing many fields of science and technology, including laser ranging1, quantum information processing2, imaging3,4, optogenetics5, artificial intelligence6,7, and augmented and virtual reality (AR/VR)8,9. In particular, the availability of multimodal and/or high-speed light modulators has been instrumental in the rapid advancement of atom-based quantum technology over the last years2,10,11,12,13. Optical control of atomic quantum systems poses the following requirements: (R1) operation at a design wavelength of choice in the visible (VIS) to near-infrared (NIR) spectrum, (R2) a scalable technology with a high channel density (>100 mm−2), (R3) a high modulation speed (>100 MHz), and (R4) a high extinction ratio (>20 dB). This motivates the development of a new modulator technology, meeting these requirements to continue driving progress. Requirement (R1) stems from the fact that many optical transitions in atoms or atom-like emitters in solids lie in the visible (VIS) to near-infrared (NIR) wavelength range. VIS to NIR wavelength operation is also essential for bioimaging14,15, optogenetics5, and AR/VR8. Requirement (R2), a scalable technology with a high channel density, originates from the need to individually control hundreds to thousands of atoms to unlock quantum technology’s full potential. High modulation speeds (R3) and high extinction ratios (R4) are required to optically control the state of atoms or atom-like emitters with high fidelity in a short time compared to the lifetime of the state13. A high-extinction and high-channel-density optical control technology acting as an ultrafast spatial light modulator is also an immediate benefit to the varied fields of research which already rely on devices with thousands to millions of times slower modulation rates.

Table 1 compares state-of-the-art multichannel light modulator technologies in terms of requirements (R1)-(R4). None of the existing technologies meets all requirements. Spatial light modulators (SLMs) operating in the VIS to NIR range based on microelectromechanical systems (MEMS)16,17,18 and liquid crystals19,20 are commercially available, but their modulation rates are limited. In the IR wavelength range, alternative SLM technologies with  > 1 MHz modulation rates have been explored based on electrical gating of graphene21,22, the Pockels effect in organic electro-optic molecules23,24,25 and dielectric thin films26, the quantum-confined Stark effect in III-V semiconductor quantum-well-based material27,28, and charge carrier effects in indium tin oxide (ITO)29 and silicon30,31. The refractive index changes associated with these fast effects tend to be small, thus long interaction lengths are needed to realize devices with a high modulation efficiency. To create compact, free-space-coupled devices, one therefore often relies on optically resonant structures, such as metasurfaces21,22,24,28,29, Fabry-Perot cavities27, resonant waveguide gratings23,25,30, or photonic crystal cavities31.

Table 1 Comparison of multichannel light modulator technologies in terms of requirements (R1)-(R4)
Full size table

To fulfill requirements (R1)-(R4), we introduce an SLM technology based on piezoelectrically actuated silicon nitride (SiN) resonant waveguide gratings (Fig. 1). The piezoelectric actuators induce strain in the gratings, resulting in a shift of the resonant wavelengths and thereby realizing free-space-coupled modulators. This technology not only enables SLMs with \({{{\mathcal{O}}}}(MHz)\) to \({{{\mathcal{O}}}}(GHz)\) modulation rates, it also operates in the VIS to NIR wavelength range. Moreover, our devices are fabricated on 200 mm diameter silicon wafers using deep ultraviolet (UV) optical lithography and CMOS-compatible (complementary-metal-oxide-semiconductor-compatible), low-temperature processes (see Methods for fabrication details). Hence, we can fabricate the piezoelectric actuators and resonant waveguide gratings directly on top of electronics to drive the devices32 or on (multiple) metal layers to enable large-scale electrical interconnects. These integration schemes are advantageous for the commercial production of large-scale SLMs.

Fig. 1: Concept and device images.
Fig. 1: Concept and device images.
Full size image

a, b Concept figure of our spatial light modulator (SLM) technology based on piezoelectrically actuated SiN resonant waveguide gratings. The piezoelectric actuators are based on AlN thin films. c Device operation principle. Applying a voltage shifts the reflection spectrum. d Chip with 4 × 4 SLM wire bonded to a printed circuit board (PCB). See Methods for packaging details. e Microscope image of 4 × 4 SLM with a 50 μm pixel pitch and 1 μm wide trenches in between pixels. f Scanning electron microscope (SEM) image showing a grating surrounded by trenches.

Our work builds on a platform developed at Sandia National Laboratories consisting of a SiN optical waveguide layer on top of an aluminum-nitride-based (AlN) piezoelectric actuation layer33. While previous works on this platform focused on photonic integrated circuits13,33,34,35, here we demonstrate densely packed free-space-coupled modulators arranged in a 4 × 4 grid with a high fill factor of  ~96% as a proof of concept for scalable SLM technology (Fig. 1). To boost the modulation efficiency of our devices we not only make use of optically resonant devices, namely resonant waveguide gratings36, but we also take advantage of the mechanical resonances of our pixels, which greatly enhances the piezoelectrically induced displacement37. Resonant waveguide gratings, also known as guided mode resonant gratings or waveguide-mode resonant gratings, couple light incident at specific wavelengths, incidence angles, and polarization to leaky guided modes, giving rise to resonances in the reflection and transmission spectra (in amplitude and phase)36.

Results

Design

We designed our SiN resonant waveguide gratings to operate at a wavelength λ of  ~ 780 nm (R1), with a pitch Λ of 0.490 μm for a SiN thickness of 0.3 μm. The simulated reflection spectrum of our gratings, \(R[dB]=10\times {\log }_{10}\left(\frac{| {E}_{out}(\lambda ){| }^{2}}{| {E}_{ref}(\lambda ){| }^{2}}\right)\), in Fig. 2a shows a steep dip at a wavelength of  ~ 800 nm with an extinction ratio  > 50 dB (R4) with respect to the reflection at  ~ 770 nm. We test our devices in a cross-polarized light setup, as schematically depicted below (see Supplementary Fig. 1 for more details). Our simulation of the reflection spectrum captures the cross-polarized testing method (see Methods for more information). The steep dip in the reflection spectrum results from the phase difference between TE and TM polarized reflected light shifting from 0 to π while the reflection amplitude of TE and TM fields remain similar (Fig. 2c), where TE polarized light corresponds to the electric field being parallel to the grating lines. When the TE and TM polarized light waves reflected off the grating are in phase with equal amplitude, the reflected light is polarized at 0° before the polarizing beamsplitter and no light is reflected towards the detector. For a phase difference of π all light is reflected. Notably, the effective interferometer governing this effect is contained within  ~ 1 μm of thickness of the device, requiring no long-pathlength stabilization.

Fig. 2: Device design, optical.
Fig. 2: Device design, optical.
Full size image

a Simulated reflection spectrum compared to the measured spectrum, with \(R[dB]=10\times {\log }_{10}\left(\frac{| {E}_{out}(\lambda ){| }^{2}}{| {E}_{ref}(\lambda ){| }^{2}}\right)\). b Influence of the grating pitch Λ on the reflection spectrum. These simulations are all for a duty cycle of 63%. c The simulated amplitude ratio and phase difference between TE and TM polarized reflected waves. The dots indicate the wavelengths at which there is a 0 and  − π phase difference, corresponding to the minima and maxima in (a). d Simulation of reflection spectrum for different grating sizes L. For a fair comparison the Gaussian input beam was rescaled such that the 1/e field full width is kept at 45% of the grating size. This way, equal proportions of the input beam interact with the grating structure for all pixel sizes.

A decrease in grating pitch blue-shifts the wavelength of the dip in the reflection spectrum (Fig. 2b). A pitch of 492 nm gives a maximum extinction ratio of >60 dB. The reduced extinction ratio at different pitches results from the ratio between the TE and TM reflection amplitudes being further away from 1 where the phase difference is 0.

By reducing the grating pitch to 0.27 μm, we can shift the optical resonance all the way down to  ~ 470 nm, the short wavelength end of the visible spectrum (see Supplementary Fig. 3). Our technology can operate at a design wavelength of choice in the VIS to near-IR spectrum simply by altering the pitch of the grating (R1).

Our simulations take into account the finite size of the gratings and the incident light beam. While the fabricated and tested devices have a grating size L of 49 μm (Fig. 1e), simulations indicate that we could reduce the size down to  ~ 20 μm without significant changes in extinction ratio and the overall shape of the reflection spectrum (Fig. 2d), resulting in even higher channel densities (R2).

The material platform with which we are working allows fabrication of undercut devices (see Methods), giving us an additional degree of freedom in engineering the mechanical device properties. We can produce pixels of length L supported by a pillar of size l underneath (schematically shown in Fig. 3a, SEM image in Fig. 3b), where L and l can be chosen freely for every pixel. The pillar also contains electrical vias between the piezoelectric actuator and a bottom metal routing layer, connecting the pixels to bond pads (see Fig. 1d, e). Figure 3c shows the simulated mechanical mode with an eigenfrequency fr of 5.9 MHz for a device with L = 49 μm and l = 10 μm. This closely matches the resonance observed experimentally, Fig. 4b, c. By altering the pixel and pillar size L and l, we can set the mechanical resonance to a frequency of choice (Fig. 3d).

Fig. 3: Device design, mechanical.
Fig. 3: Device design, mechanical.
Full size image

a Schematic cross-section of a released device with grating size L. Released devices are undercut and anchored to the substrate with a silicon dioxide pillar, with diameter l, in the middle of the grating. b SEM image of a cross-section of a released device (at an angle of 57.5°). We used focused ion beam milling (FIB) to create a hole in the device so we could image the cross-section. This is an image of the middle of a grating, where it is anchored to the substrate. c Simulated mechanical mode with an eigenfrequency fr of 5.9 MHz for a released device with L = 49 μm and l = 10 μm. Simulated in COMSOL. d Simulated eigenfrequency fr as a function of pillar diameter l for grating size L = 49 μm (top) and simulated fr as a function of L for l = 5 μm (bottom).

Fig. 4: Device characterization.
Fig. 4: Device characterization.
Full size image

a The reflection spectrum of the grating (blue) overlapped with the modulation amplitude A as a function of wavelength (orange), measured on an unreleased device at a fixed AC frequency f of 20 MHz for a peak-to-peak voltage of 10 V. b Wavelength shift Δλr/ΔV versus AC frequency f for a released and unreleased device. We used a detector with a 3 dB bandwidth of DC-400 MHz for the released device and 5−1000 MHz for the unreleased device. c Measurement of the different mechanical resonance frequencies fr for each pixel of the strong resonance around 5 MHz and the associated enhancement factor k. 13 of the 16 pixels were functioning in this device. All measurements were done at a wavelength of 791 nm. d Ringdown measurement of the mechanical mode at  ~ 5 MHz of a released device. For this measurement we first apply an AC frequency at resonance, then we switch it off and observe how the AC-coupled photodiode signal PAC decays over time t on an oscilloscope. Fitting gives a ringdown time τ of 3.1 μs. e Schematic depiction of our cross-polarized light setup. PBS: polarizing beamsplitter, λ/2: half-wave plate, λ/4: quarter-wave plate.

Characterization

Figures 4a and 2a show the measured reflection spectrum of a grating and its agreement with simulations, with a measured extinction ratio  > 20 dB (R4). We measure the modulation amplitude A and speed of the packaged devices by reflecting laser light off a pixel and detecting the reflected light (with electric field Eout) on a fast photodiode (Fig. 4e), which is connected to an electrical spectrum analyzer (ESA). When applying an AC voltage to the illuminated pixel, a peak appears on the ESA at the applied AC frequency f, as anticipated for our piezoelectrically actuated device. The height of this peak, i.e. the modulation amplitude A, depends on the laser wavelength. Figure 4a overlaps the reflection spectrum with the modulation amplitude as a function of wavelength λ. The peak of the modulation amplitude occurs where the slope in the reflection spectrum is steepest (~791 nm). The devices have an optical bandwidth of several nanometers (8 nm peak to trough of resonance, 3 dB bandwidth of 2.3 nm in the modulation amplitude measurement). This corresponds to photon-lifetime-limited modulation bandwidths on the order of 1 THz, which does not pose a limitation for the envisaged applications. Moreover, the target applications typically only require modulators with narrowband operation at a discrete set of wavelengths (e.g. a set wavelengths corresponding to certain atomic transitions). Our modulator technology is very well suited to accommodate this requirement. Either one can use separate modulator chips designed to operate at different wavelengths, or pixels designed to operate at different wavelengths can be integrated on the same chip (in separate chip areas or interleaved).

Figure 4b compares the modulation amplitude and speed for two different types of devices, released and unreleased. Released devices are undercut and anchored to the substrate with a 10 μm wide silicon dioxide pillar in the middle of the pixel (see Methods for fabrication details and Fig. 3b for an SEM image of a device cross-section). Unreleased devices have no undercut. The goal of releasing is to make the devices more mechanically compliant, resulting in lower drive voltages for the same level of induced strain. Additionally, released devices have high-quality-factor mechanical resonances. Resonant enhancement reduces the required drive voltage even further. Figure 4b shows the wavelength shift Δλr/ΔV of the reflection spectrum versus the applied AC frequency f for a released and unreleased grating. We calculate the wavelength shift from the measured modulation amplitude and the slope of the reflection spectrum. The numerous peaks and troughs in Fig. 4b are attributed to mechanical resonances38,39. The more compliant nature of the released devices results in low-frequency wavelength shifts that are more than an order of magnitude larger compared to the unreleased devices. It also leads to the appearance of mechanical resonances at lower frequencies. Finite element analysis confirms the existence of mechanical resonances and identifies the mechanical mode at fr ≈ 5 MHz for the released device (Fig. 3c). The resonance at fr ≈ 5 MHz gives a large wavelength shift of 4 × 102 pm V−1, which corresponds to an enhancement of a factor \(k=\frac{{(\Delta {\lambda }_{r}/\Delta V)}_{{f}_{r}}}{{(\Delta {\lambda }_{r}/\Delta V)}_{DC}}\approx 50\) compared to the low-frequency shift. Up to a frequency of  ~ 100 MHz, the data shows wavelength shifts larger than the low-frequency response for the released device (R3). The stiffer nature of the unreleased device shifts the resonances to higher frequencies. Up to  ~ 700 MHz, wavelength shifts are comparable to the low-frequency response. The frequency response measurements are limited by the photodetector bandwidths (DC-400 MHz for the released device and 5−1000 MHz for the unreleased device measurement) and the 900 MHz bandwidth of the voltage amplifiers on the printed circuit boards (see Methods).

From the frequency response measurements (Fig. 4b) and the measured reflection spectrum (Fig. 4a), we can estimate the voltages required to shift the spectrum by 8 nm, that is, the wavelength difference between the trough and peak in the spectrum. We find a voltage of  ~ 20 V for a released device at the  ~ 5 MHz mechanical resonance and  ~ 1 × 103 V off-resonance (100 kHz driving frequency). For an unreleased device, we estimate a voltage of  ~ 40 × 103 V off-resonance (at 10 MHz). We recognize that the off-resonance and unreleased device voltages are high in this proof-of-concept demonstration and suggest ways of lowering these significantly in the discussion. Also, we note that several technologies exist today to deliver voltages in the \({{{\mathcal{O}}}}(10\,V)\) to \({{{\mathcal{O}}}}(kV)\) range at \({{{\mathcal{O}}}}(MHz)\) to \({{{\mathcal{O}}}}(GHz)\) frequencies, such as Si LDMOS (laterally-diffused metal-oxide semiconductor)40, SiC41, and GaN-based devices42.

Figure 4d shows the ringdown measurement of the mechanical mode at  ~ 5 MHz for a released device. Fitting the data gives a ringdown time τ of 3.1 μs. This corresponds to a mechanical Q-factor of \({Q}_{m}=\frac{{\omega }_{m}\tau }{2}=52\) (ωm = 2π × 5.3 MHz)43. This is in good agreement with the enhancement factor of  ~ 50 observed in the frequency response measurement, as expected for a harmonic oscillator.

Pixel-to-pixel fabrication-induced variations lead to differences in resonance frequencies and enhancement factors, e.g. variations in pixel size, layer thicknesses, and depth of the undercut. For the pixels in our 4 × 4 device, we determined the mechanical resonance frequencies fr of the strong resonance around 5 MHz and the associated enhancement factor k (Fig. 4c). The average resonance frequency is 5.7 MHz (standard deviation of 0.9 MHz) and the average enhancement factor is 32 (standard deviation of 13). Established post-fabrication trimming techniques such as laser trimming44 and focused ion beam (FIB) milling45 can be used to tune up the resonances. A possible approach for closed-loop laser trimming consists of combining the trimming laser, probe laser (to be modulated by the device), and imaging optics in a single optical system, as has been demonstrated for photonic crystal cavities31 and ring resonators13. Image recognition methods can locate the pixel edges to be trimmed44. We can parallelize laser trimming of multiple pixels by e.g. using a liquid crystal SLM to generate arrays of laser spots and photodetector arrays for probe laser detection.

Discussion

One of the main challenges ahead is to reduce the voltages required to achieve high-extinction-ratio modulation. We see several ways to achieve this. Firstly, we can use a piezoelectric material with an enhanced response. Scandium aluminum nitride, for instance, has a piezoelectric response up to 5 times larger compared to AlN46. Secondly, we can use optical resonators with reduced optical linewidths. Recently, a silicon-nitride-based resonant waveguide grating with a linewidth of only  ~ 0.7 pm at 779 nm has been demonstrated47. Vertically coupled photonic crystal cavities with high quality factors are another promising candidate. These have been demonstrated in silicon with linewidths  < 10 pm at a wavelength of 1.6 μm31. Silicon nitride photonic crystal cavities with linewidths down to 0.1 nm at 760 nm have been reported48, though these were not vertically coupled. By combining narrow-linewidth optical resonators with piezoelectric materials with an enhanced response, we can potentially reduce the drive voltages by three to four orders of magnitude. A significant reduction in linewidth will increase the need for post-fabrication optical resonance tuning, as the resonant wavelength is likely to deviate from the target operating wavelength. This is a common problem in integrated optics for which multiple solutions have been reported, e.g. laser trimming13,31 and on-chip heaters13,49.

Another challenge is to get all the required electrical driving signals onto the SLM chip as we scale up the number of pixels. In our proof-of-concept device we use wire bonds. The number of wire bonds is limited by the available chip perimeter. For example, on a 1 cm by 1 cm chip, for a pixel size of 50 μm and a typical bond pad spacing of 100 μm, we can fit 40,000 pixels, but only 400 bond pads around the chip perimeter. One approach to increase the number of electrical inputs is to make use of through-silicon vias50 and bond the SLM chip to an interposer. This interposer could also hold electronic driving ICs. An even more advanced solution is to integrate some of the electronics on the SLM chip.

We see many potential applications for this SLM technology. Depending on the target application, we can configure our modulator technology in different ways. For instance, driving the modulators at a strong mechanical resonance generates strong sidebands to the incident laser light (the sidebands could be resonant with an atomic transition). The generation of these sidebands can be switched on and off (individually controlled for every pixel) with response times on the order of microseconds, with the exact response time depending on the mechanical quality factor. This has applications in the optical control of atomic systems, where one wishes to address certain atomic transitions in large-scale arrays of atoms, ions, or solid-state color centers13. We can use different mechanical resonances to generate sidebands at different frequencies or engineer the pixels to alter the mechanical resonances, e.g. by varying the grating size or the depth of the undercut. Laser trimming31,44 or focused ion beam milling45 can alter the resonance frequencies post-fabrication.

Predictable time-varying and even periodically repeating light patterns can be generated in the far field38 by driving each of the pixels at a particular mechanical resonance frequency, enabling low-voltage driving, with a particular phase. The generated pattern will depend on the chosen frequencies and phases. This concept has been described for photonic integrated circuit phased arrays38 and has applications in atom control51, display technology52, and laser ranging53. The proposed SLM technology enables MHz speeds, compared to conventional MEMS laser scanners limited to sub-MHz speeds52.

Our modulator technology could even generate arbitrary waveforms requiring the full device bandwidth (i.e. not only driving at resonance frequencies) by using signal equalization techniques to compensate for the non-flat frequency response. Signal equalization has been demonstrated for photonic integrated circuit modulators with mechanical resonances in the MHz to GHz range39.

In conclusion, we have presented a scalable high-channel-density, high-speed modulator technology operating in the visible to near-infrared spectrum. This technology has the potential to drive progress across various fields in science and technology, including large-scale coherent optical control of atomic quantum systems.

Methods

Fabrication

Our devices consist of a silicon nitride (SiN) optical waveguide layer on top of an aluminum-nitride-based (AlN) piezoelectric actuation layer with a silicon dioxide optical buffer layer in between33,34,35. The devices are fabricated on 200 mm diameter silicon wafers at Sandia National Laboratories using deep ultraviolet (UV) optical lithography. Low-temperature, CMOS-compatible processes are used for thin-film deposition, namely plasma-enhanced chemical vapor deposition (PECVD; for amorphous silicon, SiN and oxide deposition) and physical vapor deposition (PVD; for AlN and metal deposition). There are three metal layers in the stack to enable device actuation. From bottom to top, these are: a routing layer which we use to connect each of the pixels to bond pads, a bottom electrode layer, and a top electrode layer. In between the two electrode layers there is the AlN layer; together these comprise the piezoelectric actuation layer. The bottom electrode layer is separated from the routing layer by a silicon dioxide buffer. The three metal layers are connected by vias. Throughout the process flow, chemical-mechanical polishing (CMP) is used for planarization. In between the bottom electrode layer and routing layer there is also a patterned amorphous silicon layer to enable undercutting of devices. After having processed the entire device stack, a release etch in XeF2 gas removes the amorphous silicon to create undercut devices. The XeF2 gas can reach the amorphous silicon layer through deep, etched trenches.

Packaging

We dice the wafers into chips and wire bond them to printed circuit boards (Fig. 1d). The devices with 4 × 4 pixels have 17 bond pads, 1 pad for a shared ground and 16 pads for the signals applied to each of the pixels. Each of our PCBs contains 16 operational amplifiers (OPAMPs; Texas Instruments THS3491) to individually address 16 pixels. The OPAMPs provide a voltage gain of 5, have a 900 MHz bandwidth, and a  ± 16 V supply voltage. Ganged microscale RF plugs (Samtec) connect the PCBs to signal generators.

Simulation

We use Ansys Lumerical’s 2D FDTD solver for optical simulations. We run the simulation for two orthogonal polarizations. Afterwards, we use transfer matrix multiplication to implement the beamsplitter and wave plates together with the simulations of both polarizations. Our simulation takes into account the Gaussian shape of the excitation and collected beams. To directly compare the simulation with the measurement, we implement the same normalization method. We perform a simulation without the grating structure to simulate the effect of reflection from the aluminum layer with the silicon dioxide and SiN layers on top (Eref). The resulting simulation is able to capture the entire measurement setup to accurately predict the results.