1. Introduction

It is currently axiomatic in the bioelectronics field that wireless medical implants should move to smaller dimensions (mm or even smaller sub-mm) to enable implantation with minimal tissue damage and long-term impact on surrounding tissue and to ease subsequent removal of such devices. Given this, many researchers have recently focused on developing wireless ultrasmall implant systems to revolutionize healthcare by improving their applicability in tissue for diagnostics and therapies.

Most existing implantable wireless systems use radio frequency (RF) electromagnetic (EM) waves in the GHz frequency range to transfer power to and data from the implant [1,2]. However, coupling these EM waves through an RF coil (antenna) to sub-mm implants can be inefficient [3]. This arises because GHz EM waves experience high attenuation in tissue [4], and the miniaturization of the RF antennas to sub-mm scales results in a small coil impedance, making it challenging to harvest enough RF energy to power up the implant [5]. Ultrasound waves with sub-mm wavelengths have recently gained attention as they enable comparably more efficient coupling to mm-scale implants placed deep in tissue [6,7]. However, the miniaturization of ultrasonic implants to sub-mm scales is challenging, and scaling down the piezoelectric bulk crystal (antenna) used in ultrasonic implants degrades power transfer efficiency and data transfer reliability of the acoustic link in tissue [8,9]. Light, on the other hand, is very attractive to use for designing wireless implants at the sub-mm or even $\mu m$ scales given the comparably short wavelengths (sub-$\mu m$ and $\mu m$ scale), enabling efficient coupling and conversion of light to electrical energy through a semiconductor device [10,11]. In biosensing applications, the wireless optical implants consisting of a sensor and an integrated circuit (IC) incorporating interface electronics, power management, and data communication components have been demonstrated [12,13].

This work focuses specifically on an optical wireless data communication modality intended for low-power wireless implants with ultrasmall footprints. In practice, uplink data (data emitted by the implant) can be transmitted from the inside to the outside of the body by using SiGe optical modulators or light sources (LEDs or lasers) [12–16]. However, SiGe modulators require large reverse bias voltages (1-3V), restricting the circuitry that can be used to drive them. Therefore, light sources, including LEDs and lasers, have recently gained popularity for optical data transfer, mainly due to their ease of integration with CMOS IC. However, these light sources require turn-on voltages with specialized drive circuitry to operate and consume substantial amounts of power.

In this work, we present a high-$Q$ electro-optic modulator (EOM) based on piezoelectric actuation that is fabricated using standard microfabrication techniques. In prior work, we demonstrated a low-$Q$ optical voltage sensor based on the same operating principle as the EOM for measuring high voltages in the order of tens of volts [17]. Here, the device incorporates high-reflectance mirrors on both sides of a piezoelectric film and an anti-reflection coating (ARC); this enables high-$Q$ operation and hence operation at small voltage amplitudes. This EOM shows great promise for establishing an optical wireless link to an implant requiring high-bandwidth uplink data transfer while avoiding the drawbacks of the current optical data transfer methods mentioned above. The EOM was designed to operate at a wavelength of ~1300nm to minimize light attenuation due to scattering in tissue [18]. The device behaves as a simple capacitor, eliminating the need for active power consumed by the device to initiate uplink data transfer. The device can be heterogeneously integrated with a low-power CMOS IC, required for sensing, energy harvesting, and data communication to achieve extreme implant miniaturization, Fig. 1(a), and be operated at voltages down to 0.5V, compatible with CMOS electronics, with extremely minimal drive circuitry requirements. The device size is demonstrated to be scaled down to micrometers (<100$\mu m$).

 

Fig. 1. (a) Example integrated sensing system incorporating separately-fabricated optical modulator. Excitation beam is coupled into a resonant modulator. The intensity of this beam reflected from the modulator surface can be modulated by the voltage provided by the sensing chip applied between the top and bottom contact. (b) The device operates on the principle of light intensity modulation, wherein light at a wavelength $\lambda _{in}$ near the resonance dip (that is, the steepest slope of the reflectance $R$ curve) is incident on the device surface. A voltage applied across the device results in a shift in its resonant wavelength (dashed line), enabling to modulate the amount of light reflected off the device.

Download Full Size | PDF

2. Design and fabrication

The electro-optic modulator (EOM) operates as a normal-incidence reflective optical intensity modulator requiring a narrowband tunable light source. Fundamentally, the EOM is a resonant cavity with a half-wavelength piezoelectric active layer sandwiched between two high-reflectance mirrors. In this work, the active layer is formed by Aluminum Nitride (AlN), and the mirrors are distributed Bragg reflectors [19]. Figure 1 (b) shows the operating principle of the EOM. When light is incident on this device at its resonance wavelength ($\lambda _{r}$), the reflected light intensity is at a minimum. Moving away from $\lambda _{r}$, the intensity increases until it reaches a maximum after roughly a linewidth. By using incident light at a wavelength between these two extremes, a slight shift in the $\lambda _{r}$ can be made to result in a change in the reflected light intensity. An applied voltage ($V_{in}$) across the device produces an electric field inside the active AlN layer that causes it to change its thickness, resulting in a shift in the $\lambda _{r}$ and hence a change in the reflectance ($R$). In practice, it is easier to work with the relative change in reflectance, $\Delta R$. For the EOM operated at a wavelength ($\lambda _{in}$) where the $R$ slope is the steepest near $\lambda _{r}$, the modulation depth $\Delta R$ can be approximated by [17]:

The EOM operates without the need of a bias current or voltage, and acts as a simple capacitor. Electrically, the device contains no active components, does not operate using charged carriers, and is well-modeled as a single capacitor. Any electric field applied within the inner piezoelectric layer causes it to deform, but this requires no bias current, other than leakage through the layer.

The device was designed to operate at a wavelength of 1310nm that is a minimally-scattering wavelength for transmission through biological tissue. A central problem for optical communication through tissue is overcoming absorption and scattering in tissue [20]. At increasingly longer wavelengths, Mie and Rayleigh scattering becomes less problematic, but at wavelengths above ~2$\mu m$, absorption by water begins to dominate the optical loss [21]. Along with the availability of efficient optical detectors in this region [22], this has led researchers to conclude that the $1.0 \mu m$ to $1.7 \mu m$ near-infrared (NIR) region is an optimal one for imaging in biological tissue [18]. Fortunately, the NIR is also a minimally-absorbing wavelength in glass used in optical fibers [23], and optical telecommunications components are readily available in this region. Operation at wavelengths >1$\mu m$ also minimizes the absorption of several materials commonly used in microfabrication (among them $SiN$, $SiO_2$, $\alpha -Si$). For these reasons, we chose a target wavelength of 1310nm.

The electro-optic modulator (EOM) was fabricated using CMOS-compatible micromachining techniques. The EOMs were fabricated in the Marvell nanofabrication facility at the University of California, Berkeley. Figure 2 shows (a) the fabrication steps and (b) the cross-sectional SEM of the device. Briefly, we deposited an anti-reflection (AR) coating of 50nm titanium (Ti) and 200nm amorphous Si ($\alpha -Si$) on a bare p-Si wafer. We then deposited nine quarter-wavelength layers of alternating 240nm silicon dioxide ($SiO_2$) and 90nm $\alpha -Si$ using PECVD at 350$^{\circ }C$. We then sputtered 310nm-thick aluminum nitride (AlN) layer, and deposited a second set of nine alternating $SiO_2$ and $\alpha -Si$ layers for the top mirror. We then lithographically patterned circular regions with sizes between $5 \mu m$ and $1280 \mu m$ and etched the top mirror and the AlN layer. We confirmed full clearing of the AlN layer with an AlN etchant (dilute TMAH) drop test. Next, we sputtered 10nm Ti/100nm Al on the backside of our wafer to compensate for film stress and provide an ohmic contact to the p-Si substrate. After this, we sputtered a 60nm-thick layer of indium tin oxide (ITO) on top of the entire structure, lithographically patterned it, and ion milled all but a circle overlapping the original devices. Finally, we patterned a final liftoff layer and evaporated 300nm-thick Al on the surface of the wafer to form bond pads. The fabricated 5${\times }$mm$^2$ die containing different size devices and test structures, shown in Fig. 2(c), was bonded with conductive silver epoxy and wire-bonded to a printed circuit board (PCB) for testing. Note that multiple devices were placed on the same die to enable fair performance comparison of different size devices. See Supplement 1 for further details of the fabrication steps.

 

Fig. 2. (a) Device fabrication process. 1. Deposited backside AR-coating and mirror. 2. Deposited AlN layer and frontside mirror. 3. Deposited backside Ti/Al metal contact. 4. Etched device mesas, stopping on back mirror. 5. Sputtered, patterned top ITO contact. 6. Evaporated, patterned Al bond pad layer using a liftoff process. (b) Cross-sectional SEM of the fabricated device. The measured thicknesses of the layers are as follows: ITO: 77nm, $SiO_2$: 212nm (mean, top), 207nm (mean, bottom), $\alpha -Si$: 109nm (mean, top) and 103nm (mean, bottom), AlN: 314nm, $\alpha$-Si ARC layer: 229nm, Ti ARC layer: 93nm. (c) Optical micrograph of the fabricated 5${\times }$mm$^2$ die containing different size devices, and (inset) close-up top views of the $80 \mu m$ and $320\mu m$ and diameter devices.

Download Full Size | PDF

Fabrication of high-quality and extremely smooth $SiO_2$ and $\alpha -Si$ layers allows the entire structure to be deposited at once. A major challenge with deposition of Bragg mirrors is that as the number of layers increases, the mirror roughness tends to increase [24]. This degrades the reflectivity (${R}$) of the mirror, lowering the device $Q$. To avoid this, we performed a 17-run definitive screening experiment for $SiO_2$ films deposited on bare Si substrates and for $\alpha -Si$ layers deposited on thermally oxidized substrates. We varied the temperature, pressure, and available gas flows and measured the deposition rate with cross-sectional SEM, film roughness using AFM, and stress using wafer curvature (see Supplement 1). Both optimal $SiO_2$ and $\alpha -Si$ recipes achieved rms roughnesses of <1nm and film stresses of ~200MPa. Fully-deposited 9-layer mirrors maintained rms roughnesses of 0.8nm, and the roughness did not increase substantially with an increasing layer number.

3. Experiment results

We measured reflectance (${R}$) spectra of the electro-optic modulator (EOM) using a beamsplitter-based cage system reflectometer. Figure 3(a) shows a schematic diagram of the setup used to measure the EOM reflectivity. In this setup we used a fiber-coupled thermal broadband light source (Thorlabs SDS201L). The light was collimated with a 15mm fixed-focus collimation package L1, reflected with a 50/50 plate beamsplitter BS1, and focused onto the device with a 40mm plano-convex lens L2. The reflected light from the device was then coupled through a focusing lens L3 into a $100 \mu m$ core / 0.1 NA fiber, which was measured by an optical spectrum analyzer (OSA; Yokogawa AQ6370C). To obtain a ${R}$ spectra of the EOM, the measured spectrum data was normalized to the data collected in the same manner using a 100nm gold-coated reference sample; the sample has approximately unity reflectance above 800nm [25]. Figure 3(b) shows the measured and predicted ${R}$ spectra of the EOM with anti-reflection coating (ARC), showing that the experiment data agree well with the data predicted from a transfer matrix model (Supplement 1) using the device parameters provided in Table S5. We believe the difference in the measured and predicted spectra mainly results from difficult-to-predict variations in the oxide absorption coefficient throughout the layer stack.

 

Fig. 3. (a) Device characterization setup. Light is coupled in either from the tunable laser source or the broadband thermal source into the reflectometer and collimated by lens L1 (f=15mm). The beam then reflects off beamsplitter BS1, is then focused and de-magnified (M=0.38) by lens L2(f=40mm) onto a device. It then reflects back by lens L3 (f=15mm) and is coupled into a second fiber cable. That cable is then either routed to a TIA / oscilloscope (for narrowband/modulation experiments), a power meter (for measuring reflected power / attenuation) or an OSA (for broadband reflectance). (b) Reflectance of planar film with ARC measured using a thermal source and an OSA. Dashed line represents simulation results using TMM package (see Supplement 1). Inset near resonance, theoretical inset is offset to force overlap of resonances. Inset data measured using swept-source narrowband laser for more accurate $Q$-factor estimation. Best-fit $Q$ is 2900 theoretical using lossless $SiO_2$ ($\kappa = 0$) and 2900 measured.

Download Full Size | PDF

The AR-coated devices achieved a maximum quality factor ($Q$) of 2900, but optical coupling reduces modulation gain ($\beta$) at smaller device sizes. We measured the $Q$ of each device by sweeping the tunable laser source (Santec 550) near resonance, and fitting the resultant measured spectra to a Lorentzian at a fixed $Q$ using least-squares linear regression. A typical fit curve is shown in Fig. S5. The $Q$ values measured for the five different device sizes are shown in Fig. 4(a), and range from ~2500 to ~2900. These are representative of devices measured elsewhere on the same wafer and across wafers, with typical $Q$ values in the range of 1000-3000. The largest device $Q$ (~2900) measured with the tunable laser was almost three times that (~1000) measured with the optical spectrum analyzer. We suspect this is due to the poor beam quality factor ($M^2$) of the fiber-coupled thermal source compared to the tunable laser, and the resultant much wider range of incident angles present when using the thermal source, leading to a smeared-out spectra near resonance.

 

Fig. 4. (a) Measure quality factor ($Q$) values for devices with differently-sized diameters, from $80 \mu m$ to $1280\mu m$. $Q$ was measured by sweeping the tunable laser source near resonance, and fitting a Lorentzian lineshape to the resulting spectra using least-squares linear regression. (b) Measured modulation gain ($\beta$) for differently-sized devices.

Download Full Size | PDF

While the $Q$-factor for smaller devices remained high, the device $\beta$ decreased substantially from $2.3 * 10^{-3} V^{-1}$ to $9 * 10^{-5} V^{-1}$ (see Fig. 4(b)). We believe this is primarily due to the poor optical coupling from the beam to the device, due to the relatively high reflectance observed when aligning to the device.

The anti-reflection coating (ARC) enables achieving high $Q$ by eliminating back reflection of light due to the transparent substrate. One challenge of operating the device at a wavelength of around 1300nm is that Si substrate is transparent in this region. This is because the incident light near resonance that passes through the device layers will be reflected off the bottom electrode surface, degrading the device $Q$ and therefore capability of modulating light intensity. We alleviated this problem by adding an ARC based on a $\alpha -Si$ ($n \approx 3.5$) layer on top of a Ti layer underneath the bottom mirror; these materials have high compatibility with standard fabrication tools. Any light near the device resonance which makes it through the bottom mirror is absorbed by the ARC, eliminating the effect of the substrate on the device performance. Fig. S7 shows the measured and predicted reflectance ($R$) spectra of the device without ARC, revealing that the device without ARC has a substantially lower $Q$ (440) than that (2900) of the device with ARC as expected.

To extract the modulation depth ($\Delta R$) of the electro-optic modulator (EOM), we utilized the same cage system with a fiber-coupled tunable laser source. As this device has a high ${Q}$, the optical source linewidth must be much smaller than the resonant peak linewidth for light intensity amplitude modulation to function properly. The only light source capable of providing this at sufficient power is a laser. Therefore, we used a tunable O-band laser (Santec 550) as a light source, coupled to a $8 \mu m$ core diameter fiber. We measured the change in device photocurrent using an integrated fiber-coupled photodiode and a transimpedance amplifier (TIA) (Thorlabs PDB450C) with a gain of $10^5$V/A and a bandwidth of 4MHz, limited by the TIA. The TIA output was digitized using an oscilloscope (Keysight DSOX2024A) with a 5MHz sampling rate and then sent to a computer for data storage. Prior to $\Delta R$ measurements, we swept the laser wavelength around the resonant wavelength to precisely identify the operating wavelength. To maximize the $\Delta R$ during operation, the laser wavelength was set slightly off-resonance where the resonant slope was the steepest. In the measurement, we applied a 100Hz sinusoidal signal with various peak amplitudes from 10mV to 10V to the EOM and measured the change in optical power reflected off the device, a representative power spectrum of which is shown in Fig. 5. The measured power value at 100Hz was normalized to the DC level to extract the $\Delta R$. Figure 6(a) depicts the measured $\Delta R$ versus voltage amplitude applied to the EOM with a diameter of $1280\mu m$, showing that the device is highly linear ($R^2 = 0.9999$) in the typical CMOS voltage range of 0.5-5V and maintains linearity even well above the typical maximum voltage tolerated by CMOS electronics. The measured device gain $\beta$ of $0.8 * 10^{-3} V^{-1}$ is slightly lower than the predicted value from analytical model value of $2.5 * 10^{-3} V^{-1}$ (see Table S6). The most likely cause of the discrepancy is the relatively poor fit of a Lorentzian lineshape to our measured reflection spectra - leading to an inaccurate estimation of the true slope. It is also possible that the deposited AlN had slightly lower values of $d_{33}$ and/or $r_{33}$ due to direct deposition on oxide.

 

Fig. 5. Measured input-referred power spectral density with an input 10mV amplitude sinewave applied to the device. The red arrow indicates the location of the input sinewave, and the blue and orange curves represent the unfiltered and 1kHz high-pass filtered data, respectively. High-frequency (>10kHz) PSD was 153$\mu V / \sqrt {Hz}$.

Download Full Size | PDF

 

Fig. 6. (a) Measured modulation depth ($\Delta R$) versus input voltage in free space. (b) 1kHz highpass-filtered time-domain data from an 0.5V$_{pp}$, 9kHz square wave applied to device. (c) BER versus SNR, obtained by varying incident optical power from 0.3mW to 10mW. The data was collected at 5MHz sampling rate with a 4MHz TIA bandwidth and a 50% duty-cycle, 100 kHz square-wave input modulation signal. The BER was predicted using an AWGN channel model, and calculated treating all individual sample points as bits, excluding the square wave HIGH->LOW and LOW->HIGH transition points. (d) $\Delta R$ versus input voltage measured through 0.8mm-thick chicken skin resting on the device. The device used for these measurements has a $1280\mu m$ diameter.

Download Full Size | PDF

The input-referred voltage noise is 280$\mu V / \sqrt {Hz}$ with 1mW of incident optical power and is nearly shot-noise limited. In this work, there are three main sources of noise: shot noise, excess laser noise, and TIA noise. The DC photocurrent was typically around $4 \mu A$, yielding a theoretical shot noise of $1.1 pA / \sqrt {Hz}$. The TIA used had a specified typical noise of $1.45pA / \sqrt {Hz}$ of $10^5 \Omega$. The total measured noise above the $1/f$ corner of the laser of 10kHz was $1.46 pA / \sqrt {Hz}$, about 3dB above the shot noise limit, and consistent with the two aformentioned noise sources.

The integrated noise current from 2Hz to 10kHz was $4.8 nA$, corresponding to a laser relative intensity noise (RIN) of 0.1%. The input-referred noise spectrum from 2Hz-500kHz with a $10mV$ input amplitude sinewave is shown in Fig. 5. Note that we computed the input-referred spectral noise from the data that we obtained by dividing the measured photocurrent by $i_{max}*\beta$, where $i_{max}$ is the photocurrent due to a perfectly reflective surface ($R=1$).

The electro-optic modulator (EOM) can enable uplink data communication with implants through digital amplitude modulation of the reflected light from the device with a bit error rate (BER) <$10^{-4}$ in a bandwidth of 4MHz with a $0.5V$ driving voltage. To test the device’s ability to transmit digital data, we used the same setup as was used in the $\Delta R$ experiment and operated the $1280\mu m$ diameter EOM with square-wave voltages with a peak-to-peak amplitude of 0.5V and frequencies of 9kHz and 100kHz, generated by a function generator (Agilent 33210A). To reduce the effect of laser low-frequency noise, such as 1/f noise, on the output, we post-filtered the data from the device at 1kHz through a 3rd-order butterworth high-pass filter. Figure 6(b) shows the time-domain transient optical response of the EOM operated with a 9kHz square-wave signal at a 1MHz sampling frequency.

For the BER measurement, we collected 500,000 data points from the device operated with a 0.5V$_{pp}$, 100kHz square-wave signal at a 5MHz sampling frequency and excluded the first and last five time constants (5000 points) of the data due to filtering. At voltages greater than $0.5V_{pp}$, the BER was too small to be measurable. We measured the signal-to-noise ratio (SNR) by summing the power of the first harmonic and all subsequent odd harmonics, taking this to be the signal power, and the remaining power to be the noise power. We took each individual sample of the data to be a "bit", excluding the samples during a high-to-low or low-to-high transition, as well as the adjacent samples. To experimentally calculate the BER as a function of SNR, the laser optical power was varied between 0.3mW and 10mW during the measurement. Figure 6(c) shows the measured BER versus SNR, showing perfect agreement with the BER predicted using AWGN channel model [26]. As expected, SNR increased roughly linearly with output power, from -5dB to +11dB when incresing the power from 0.3mW to 10mW. Most experiments done here, when high-pass filtered at 1kHz, were within 3dB of the shot noise limit. At frequencies <10kHz, the laser drift and 1/f noise becomes dominant (see Fig. 5). It is also important to know that the BER performance of the device can be improved when the device is operated with larger modulation voltages, as applying higher voltages increases the device $\Delta R$.

The device operated through 0.8mm-thick dry chicken skin with degraded SNR. As a proof-of-principle to determine whether this device could be implanted subcutaneously, we measured the $\Delta R$ of the $1280\mu m$ diameter EOM as a function of applied modulation voltage by placing it below the 0.8mm-thick dried chicken skin layer. For this experiment, we set the laser power to 1mW and used a measurement frequency of 9kHz. The measurement result is depicted in Fig. 6(d), showing that the device exhibited a 30 times lower $\Delta R$ than the device operated in free space. The $\Delta R$ reduction is mainly due to tissue light scattering and results in a concomitant reduction in the SNR and increase in the BER. Furthermore, the reflected light power from the device coupled into the output fiber was approximately 10 times lower than without the skin present due to the light attenuation in tissue.

The $1280\mu m$ diameter device can operate with a 5MHz bandwidth (BW), which increases to ~300MHz for the smallest $80 \mu m$ device. The underlying mechanism by which this device operates is a mixture of the piezoelectric and electro-optic effects. The speed of piezoelectric actuation is ultimately limited by the speed at which the material can deform - the speed of sound. For these devices operating in the longitudinal mode along the thickness direction, which have an active thickness of 300nm, this yields a theoretical BW of ~30GHz using a speed of sound of $10^4 m/s$ in AlN [27]. For this reason, the device operating BW in this work is set by its electrical BW, or its RC time constant. We extracted the device series resistance R and capacitance C by fitting the lumped element RC model to the measured impedance data, shown in Fig. 7(a) for the $1280\mu m$ device. As we used a thin layer of relatively low-quality ITO, our device series resistance was obtained to be ~1k$\Omega$ for all device sizes. Device capacitance values from fitting vary from ~28pF for the largest $1280\mu m$ device to ~0.5pF for the smallest $80\mu m$ device, after correcting for bond pad capacitance. This makes the theoretical BW range from ~6MHz for the largest device to ~300MHz for the smallest device. Figure 7(b) shows the measured optical frequency response of the $1280 \mu m$ device that matches well with the theoretical BW predicted using the R and C values from the fitting. The fitted high-frequency rolloff fitted to 8MHz was close to that expected (4MHz from the limiting TIA or 5MHz from the device). We also found an unexpected low-frequency rolloff, which fitted to a pole-zero pair at 9Hz; this may be due to leakage in the AlN layer, and would be consistent with a leakage resistance of 600M$\Omega$. This is substantially higher (~6x) than measured, but also higher than the values our impedance meter could reliably measure, and corresponds to a resistivity of $2.6 * 10^{11} \Omega \cdot$cm, in the typical AlN resistivity range of $10^{11}$ - $10^{13} \Omega \cdot$cm [28].

 

Fig. 7. (a) Device impedance of representative 1280$\mu m$ diameter device. Dashed line represents lumped element parallel RC model, with a fitted R value of 1k$\Omega$ and C value of 28pF (zero at ~5MHz). Theoretical C value from literature permittivities and SEM thicknesses is 32pF. (b) Device normalized gain versus frequency for the 1280$\mu m$ device. Fitted single-pole corner frequencies of 9Hz and 8MHz using a TIA bandwidth (BW) of 4MHz. $\beta _{midband}$ refers to gain at 10kHz. Other device sizes have virtually identical frequency responses limited by the TIA BW.

Download Full Size | PDF

4. Conclusion and discussion

This work presents a high-$Q$ electro-optic modulator (EOM), formed by a piezoelectric AlN thin film sandwiched between two high-reflectance Bragg mirrors, for wireless medical implants. The EOM is fabricated using standard fabrication processes with the highest temperature of 350$^{\circ }$C, making it CMOS compatible. The device operates at a wavelength of ~1300 nm in the NIR window with low tissue absorption, enabling to establish an efficient optical wireless link to implants for high-bandwidth uplink data transmission. The device operation is successfully demonstrated ex vivo in chicken skin. We also demonstrated that the device is scalable down to micrometer in size (<100$\mu m$) and capable of operating at CMOS voltages down to 0.5V. Although there is extensive further work required to demonstrate the integration and operation of the EOM with CMOS electronics and to fully address biocompatibility and other challenges related to each targeted application, this work provides a path toward the realization of minimally invasive wireless implants for use in diagnostics and treatments.

Future versions of the device could trade off bandwidth (BW) for device modulation gain ($\beta$). In this work, we placed the device electrodes on the top of the Bragg mirror and the bottom of the Si substrate. This enables relatively easy fabrication, and reduces the total device capacitance, increasing the BW. However, the resulting capacitive division between the mirrors with a capacitance of 25pF/mm${^2}$ and the active layer with a capacitance of 240pF/mm${^2}$ (See table S5 and S6)) results in a smaller voltage acting on the active layer than the voltage applied to the device, reducing the maximum achievable $\beta$. The $\beta$ can be theoretically improved by a factor of ~9 if the electrodes are sandwiched between the mirror layers and the active layer, preventing this capacitive division (see Supplement 1). Since ITO is highly absorbing around the operating wavelength of 1300nm (Fig. S3), this also requires very thin ITO layers (<5nm) to form the device electrodes, which increases the device contact resistance and therefore further reduces the BW.

The choice of a different lower-absorption transparent conducting oxide to form the device top electrode could increase the device BW by two orders of magnitude. Here we used ITO - a well-characterized and readily available transparent conducting oxide - to form an electrode. However, the ITO absorption near the operating wavelength of interest (~1300nm) is very high ($\kappa \approx 0.2$). This forced us to use a thin ITO layer, leading to a relatively low BW due to its relatively high resistivity (~$10^{-3} \Omega *cm$). By using a different conducting oxide such as $ZnO$ with absorption and resistivity lower than those of ITO [29], the device BW can be improved by two orders of magnitude.

Modification of the fabrication process would enable piezoelectric materials requiring epitaxial growth to be used. We used aluminum nitride (AlN) as the active piezoelectric material for its relatively low permittivity ($\epsilon \approx 8.5$) and ability to deposit directly on top of thermal $SiO_2$, as shown in this work. However, AlN has a relatively small piezoelectric constant ($d_{33} \approx$ 5pm/V) [30] compared to other piezoelectric materials such as BTO ($d_{33} \approx$ 300pm/V) [31]. These materials with higher $d_{33}$ coefficients would allow for a much greater modulation depth ($\Delta R$) at a fixed applied voltage, leading to improved SNR and hence BER, especially if the device electrodes are placed between the mirror layers and the active layer. However, they also have much higher permittivities. BTO, for example, often has $\epsilon _r$ >1000 [32], which results in a smaller device BW.

System level improvements are possible. The system’s optical insertion loss from laser output to photodiode input was ~20dB as we used fiber-coupled equipment for ease of switching between different pieces of equipment. For example, direct coupling of light reflected off the device to a collecting photodiode without using an optical fiber could reduce the insertion loss and hence increase the received optical power by ~15dB, to as high as 17dB when the device is optically biased at a DC reflectance of 50%.

In this work, we post-processed the data from the device using a high-pass filter to reduce the effect of the laser 1/f noise on the system output. After high-pass filtering, our measured noise was typically within 1dB of the shot-noise limit. An another slight improvement can be achieved by using closed-loop noise reduction techniques [33] that could enable to minimize the laser 1/f noise and further improve the system noise performance.

Unlike traditional integrated photonic optical modulators, the electro-optic modulator (EOM) does not require a very precise alignment of the light beam. One of the most challenging aspects of existing high-$Q$ optical modulators - which typically make use of integrated photonics - is optical coupling into and out of the device. Traditionally, this is achieved by placing an optical fiber directly above an edge or grating coupler, and requires sub-micron to single-micron accuracy, depending on the method used [34]. It is necessary to dramatically relax this alignment requirement for a modulator intended to be used to build a wireless implant because such alignment precision is extremely difficult to achieve. In contrast to the existing photonic modulators, the EOM eliminates the need for high alignment accuracy as it only requires the light beam focused to within the diameter of the device. For the largest $1280\mu m$ diameter device, the transverse misalignment tolerance is roughly $\pm$0.5mm with the current setup.

Device operation is sensitive to angular misalignment. The EOM operation is sensitive to incident angle ($\theta$) between the beam direction and the device normal due to its high-$Q$. Any deviation in $\theta$ will cause a shift in the device resonant wavelength that can be approximately calculated using the formula: ${\Delta }{\lambda _{r}} = {\lambda _{r}}{\cdot } (1 - cos(\theta )^2)$ and a reduction in the power coupled to the device, which is also quadratic with $\theta$. Therefore, angular alignment of the device to the laser beam should be carefully performed during the surgical placement of an EOM-based wireless implant in tissue.

Overall, we have shown here that substantial improvement is possible at the device level and at the system level, with the potential for two orders of magnitude increase in bandwidth, and two orders of magnitude improvement in noise performance from improved optical coupling. This, combined with the results presented thus far, present a compelling case for use of this device as a platform for wireless readout from implantable devices. While we have demonstrated microscope-free alignment of devices down to $80 \mu m$ in size, further reductions in size are possible with more careful beam shaping and alignment procedures. Theoretically, devices with diameters as small as a single half-wavelength (600nm) are possible to fabricate, at the expense of increased difficulty with alignment and optical coupling.