A small vibrating structure can report an electric potential if one of its electrodes sits a fraction of a micron from a surface held at that potential. The electrostatic pull between them softens the structure's effective spring, so the vibration frequency drops as the voltage rises. That frequency can also be read wirelessly. A reader coil sends a short burst of drive, goes quiet, and then listens to the resonator "ring down," meaning it keeps vibrating freely after the drive stops, much as a bell keeps sounding after it is struck. Because the reader measures only the pitch of that ringing, and not its loudness, the result is supposed to be independent of how far away the reader is. Combining the two gives a battery-free, leadless electrometer tag that could sit inside a sealed vessel, on an energized insulator, or on a moving part. The difficulty is that everything else shifts the pitch too, including temperature, deposited mass, humidity, and, as argued below, the reader distance itself.
The proposal is to treat each ring-down as a waveform to be classified, in the way rare-event physics treats detector pulses, and not as a single frequency number. Within one ring-down we track how the frequency changes as the vibration dies away, and we compare several vibration patterns (modes) of the same structure at once. From this we recover three things: a distance-free zero-amplitude frequency for each mode, a separation of genuine voltage changes from temperature and mass drift, and a way to flag spurious events. The claim this rests on is specific and testable. Voltage-tunable resonators are intrinsically nonlinear, and that nonlinearity grows with the voltage. As a result, naive time-gated readout of them is not distance-independent, and the error it makes grows as the square of the voltage being measured.
Core Claim
Time-gated wireless readout has been shown to decouple resonance frequency from standoff for passive, fixed-coupling resonators. Analysis and Validation of Contactless Time-Gated Interrogation Technique for Quartz Resonator Sensors interrogated a 4.432 MHz quartz crystal through air-coupled coils and reported frequency variations that "did not exceed 3 Hz even at distances up to 2 cm." Technique and Circuit for Contactless Readout of Piezoelectric MEMS Resonator Sensors extended the method to a 6.3 MHz aluminum-nitride-on-silicon disk over 12 mm, but noted a "small residual variation" with distance that its equivalent-circuit model did not explain. These papers supply our readout method. They also motivate the proposal, because they show that distance independence is only approximate even when the resonator is passive.
For a resonator with an electrode across a gap $g$ at total potential $V$, the electrostatic force supplies a negative linear stiffness and, in addition, amplitude-dependent terms. To leading order, the instantaneous angular frequency during a ring-down of mechanical amplitude $a$ is
$$\omega(a) \approx \omega_{\mathrm{lin}}(V,T) + \kappa(V)\,a^2, \qquad \frac{\omega_{\mathrm{lin}}(V,T)-\omega_0(T)}{\omega_0(T)} \approx -\frac{\varepsilon A V^2}{2 k g^3}.$$
Here $\omega_0(T)$ is the unbiased frequency at temperature $T$, $\varepsilon$ is the gap permittivity, $A$ is the electrode area, and $k$ is the mechanical stiffness. The coefficient $\kappa(V)$ is an intrinsic mechanical nonlinearity plus an electrostatic contribution that grows with $V^2$. That contribution comes from the higher-order terms of the force law $1/(g-x)^2$. The initial ring-down amplitude is set by how strongly the reader couples to the tag, and that coupling falls with distance. A reader that averages frequency over its listening window therefore reports $\omega_{\mathrm{lin}} + \kappa(V)\langle a^2\rangle$, which depends on distance with a slope proportional to $\kappa(V)$.
The fix uses only the received waveform. The received envelope is $u(t)=\beta(d)\,a(t)$, where $\beta(d)$ is an unknown, distance-dependent scale factor. Fitting instantaneous frequency against $u^2$ within a single ring-down gives an intercept $\omega_{\mathrm{lin}}$ that does not depend on $\beta$ at all.
Separating voltage from temperature and mass then relies on multiple modes. Temperature shifts every mode's fractional frequency by nearly the same material coefficient. The file reports −41.6 ppm/°C for the aluminum-nitride disk, so a 0.1 °C drift already amounts to roughly 4 ppm. Voltage shifts each mode according to how much that mode moves the gated electrode. Mass shifts each mode according to where the mass lands, and that spatial non-uniformity was noted in the MEMS readout paper. Three modes with calibrated sensitivities give a well-posed separation of three perturbations. Any leftover mismatch, together with the within-ring-down decay rate and frequency glide, serves as a consistency check that flags events outside the model.
The last piece addresses a tension the file identifies: a wireless tag has no leads with which to supply the large DC bias that makes electrostatic coupling useful. Resonant Optomechanics with a Vibrating Carbon Nanotube and a Radio-Frequency Cavity boosted coupling with a wired ~7 V gate. We instead build the bias into the tag as a charged dielectric layer (an electret) of potential $V_0$ in series with the measured node. The frequency response then becomes approximately linear in the measured potential $V_m$, through the cross term $2V_0V_m$. This is the same cross-term logic as DC-biased electrostatic actuation, and it avoids the quadratic dead zone around $V_m = 0$.
Research Program
The sensor tag is a piezoelectric-on-silicon flexural plate resonator in the low-MHz range, vacuum-sealed. It is piezoelectrically coupled to a planar coil for readout, as in the MEMS readout paper. It also has a sub-micron lateral gap to a floating electrode that carries an electret layer and is wired, on the tag side only, to the node being measured.
The resonator is deliberately more compliant than a radial contour disk. Stiff contour modes show only weak voltage tuning, which is Gap 3 in the file. The first design task is a finite-element and circuit calculation that sets gap, area, and stiffness. Its goals are that electrostatic tuning per volt exceeds the ~1 ppm frequency resolution implied by the quartz result, that the electret-plus-signal voltage stays below the voltage at which the structure snaps shut across the gap, and that at least three modes with clearly different electrode participation fall inside the reader's excitation band. We do not assume these targets are reachable. The calculation decides it, and a negative answer would redirect the work toward a softer structure.
The reader reuses the direct-digital-synthesizer burst-and-listen architecture of the MEMS readout paper. It is extended to excite three modes, either with sequential bursts or a multitone burst, and to digitize the full ring-down.
On the analysis side, we extract instantaneous frequency and envelope for each mode from the ring-down. We then fit the backbone relation above to obtain each mode's zero-amplitude frequency and apparent nonlinear slope, and invert a calibrated 3×3 sensitivity matrix for voltage, temperature, and mass.
Background identification follows the logic of α-event characterization and rejection in point-contact HPGe detectors. That work separated surface alpha events from bulk events using a slow-tail parameter that simple amplitude cuts could not see. Our analogue is a small set of physically meaningful waveform descriptors: the fit residual, the ratio of apparent nonlinear slopes across modes, the decay rate, and the long-timescale leakage of the electret potential across successive readings.
The file contains no datasets from any of the wireless or nanomechanical platforms; each paper notes that its data are private bench measurements. We will therefore release the full ring-down waveform corpus with ground-truth perturbation labels.
Baselines
The naive reference is the published time-gated protocol applied unchanged. It uses a single mode, fixed drive, and a window-averaged frequency, and it interprets the shift as voltage through a single calibration made at one distance, as in Technique and Circuit for Contactless Readout of Piezoelectric MEMS Resonator Sensors. It runs on the same tag, reader, and ring-down recordings as our method, so the only difference is the analysis. Beating it means a smaller apparent-voltage error across the 0–12 mm standoff sweep, and the proposal predicts that this error grows as the square of the bias for the naive method.
The continuous-drive, amplitude-based competitor is the vibrating-electrode electrostatic sensor of A Non-Contact Electrostatic Potential Sensor Based on Cantilever Micro-Vibration for Surface Potential Measurement of Insulating Components. It reports a maximum measurement error of 0.92% and a linearity error of 0.47% at a 10 mm calibration distance over 1–10 kV, and relies on an exponential distance-correction model over 5–70 mm. Its absolute voltage scale is very different from ours. We will therefore compare in its own terms, as percent-of-range error at the calibration distance and away from it, with standoff varied without the true distance being supplied to either method. Beating it means matching its error at the calibration distance while keeping that error flat as standoff changes by unknown amounts.
The strongest simple competitor is a differential tag. It carries a second, identical resonator with no gated electrode as a temperature reference and reports the difference of the two single-mode frequencies. We found no reported result on Lacuna for a differential reference resonator in wireless time-gated readout. We will build it on the same die, read it with the same reader and dwell time, and calibrate it with the same number of reference points. Beating it means a lower voltage error under combined temperature and standoff variation and a lower false-alarm rate under mass-loading events.
The learned competitor is the two-stage approach of Deep learning based pulse shape discrimination for germanium detectors. It compresses waveforms with a convolutional autoencoder and classifies the latent code with a small network, and it reached 90% background rejection at 90% signal acceptance on germanium calibration data, a different physical system. We will train it on raw ring-downs from the same corpus, with the same labeled calibration budget as our sensitivity-matrix fit, and tune it across a matched grid. It is evaluated on held-out standoff distances and temperatures.
Two ablations isolate the key ingredients. The first keeps the multimode separation but replaces the zero-amplitude intercept with the window-averaged frequency. The second keeps the intercept fit but uses a single mode with a separate temperature reading.
Why not just…?
Why not just drive the tag weakly so the nonlinearity never matters? Reducing drive does shrink the amplitude-dependent pull, but at a given standoff it also shrinks the received ring-down. The quartz study already reports a severe loss of signal-to-noise with distance. The drive-sweep experiment below will map the trade-off directly. If a drive level exists that keeps the nonlinear bias below the noise across the full 12 mm while preserving ppm resolution, the intercept fit is unnecessary, and we will report that outcome.
Why not just use a reference resonator to cancel temperature? This is the real threat for temperature alone, and it may perform as well as our method in a stable, dry, fixed-distance setting. Its weakness is structural. The reference lacks the voltage-dependent nonlinearity, so the differential signal still carries a distance-dependent bias proportional to $\kappa(V)$. A second, physically separate resonator also accumulates mass and moisture differently from the sensing one. The combined temperature, standoff, and droplet experiments are designed to settle this. If the differential tag matches us on all three, the multimode design adds only complexity and we will say so.
Why not just let a neural network learn the whole mapping from ring-down to voltage? A classifier on raw waveforms could discover the same features, and on in-distribution data it may match or beat hand-built descriptors. What the autoencoder approach has not shown is invariance to a nuisance variable, such as the unknown scale factor $\beta(d)$, when that variable falls outside the training range. Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net shows how badly waveform-level statistics can shift between domains. Its tail-slope histogram overlap between simulation and real detector pulses was 0% before domain translation. We test the learned model by training on standoffs of 2–6 mm and evaluating at 8–12 mm. We also train a hybrid that feeds the network our intercepts, so that we learn whether the physics features help learning or merely replace it.
Proposed Experiments
The first decisive check needs no new device. We rerun the published protocol on a plain aluminum-nitride disk while sweeping drive amplitude and standoff. If the "small residual variation" reported by the MEMS readout paper tracks the initial ring-down amplitude and disappears after the zero-amplitude intercept fit, the residual has a mechanical, not electromagnetic, origin. That finding would be useful on its own. If the residual persists after the fit, the claimed mechanism is wrong for passive resonators and we re-examine it before building gated tags.
The second check uses the gated tag with the electret bias. We apply known potential steps to the floating node while sweeping standoff from contact to 12 mm and laterally misaligning the reader. The main result is the apparent-voltage error versus standoff at several bias levels, for every baseline. The proposal predicts that the naive method's error grows as the square of the bias and that the intercept-based method's error stays flat. If both stay flat, the core claim fails.
The third check tests discrimination. Temperature ramps, humidity steps, water-microdroplet deposition and evaporation (reusing the MEMS readout paper's protocol), reader-side electromagnetic bursts, and mechanical taps are interleaved randomly with true potential steps. Performance is reported as detection efficiency against false-alarm rate for voltage steps of a fixed size, together with voltage error while each background is active. Following the germanium literature, we quote the operating point at 90% signal acceptance.
The final check runs for weeks. It measures electret leakage under dry and humid conditions and asks whether the slow decay of the built-in bias can be tracked and separated from genuine drift in the measured node, using the multimode separation and the leakage-time signature.
References
- Analysis and Validation of Contactless Time-Gated Interrogation Technique for Quartz Resonator Sensors
- Technique and Circuit for Contactless Readout of Piezoelectric MEMS Resonator Sensors
- Resonant Optomechanics with a Vibrating Carbon Nanotube and a Radio-Frequency Cavity
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- A Non-Contact Electrostatic Potential Sensor Based on Cantilever Micro-Vibration for Surface Potential Measurement of Insulating Components
- Deep learning based pulse shape discrimination for germanium detectors
- Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net