Gate Sweeps as a Signal Fingerprint for Graphene Terahertz Dark-Matter Searches

Graphene plasmonic resonators are tiny antennas whose resonant frequency moves when you change a voltage on a nearby electrode. The communications community built them to pack thousands of terahertz antennas onto a chip. We propose using that voltage knob for a different purpose: searching for an invisible particle that would leave a faint, single-frequency deposit of power in the device. The usual reason to tune a detector is that a sharp resonance amplifies the signal. That reason does not apply here, because graphene resonators at 1 THz are too lossy to be sharp. The useful fact is a different one. As the gate voltage sweeps, a real single-frequency signal changes the detector's output along a curve that can be computed in advance. Broadband thermal backgrounds and electronic drifts follow different curves. A likelihood fit across the whole sweep can use this known shape to pick out the signal, much as a known initial state lets collider physicists recognize an invisible particle from what is missing.

The opportunity is concrete. Terahertz frequencies correspond to particle masses of roughly one to tens of milli-electron-volts. Metallic cavities are hard to use in this mass range because their size is tied to the wavelength. Graphene detectors already reach noise-equivalent powers (the signal power that equals the noise in one second of averaging) of 0.1–0.2 pW/Hz^1/2, as reported in Recent Progress in the Development of Graphene Detector. The missing piece is an analysis that turns a low-quality, tunable absorber into a background-robust search instrument. This note proposes that analysis and a laboratory test that could falsify it.

Core Claim

For a gate-tunable graphene absorber with a quality factor of order a few, electrostatic tuning improves rare-signal sensitivity mainly by supplying a signal-specific response template, not by resonant enhancement. The factor of a few follows from the plasmon damping time. The Kubo-model conductivity used across this literature assumes relaxation times of 0.1–1 ps, so at 1 THz the product of angular frequency and relaxation time is only about 0.6–6.

We define the template as follows. Let $A(\omega;E_F)$ be the device's absorptivity at angular frequency $\omega$ when the graphene Fermi level (set by the gate) is $E_F$. A monochromatic signal of power $P_s$ at frequency $\omega_s$ then produces an output proportional to $P_s\,A(\omega_s;E_F)$. Its shape in $E_F$ is a peak located where the resonance crosses $\omega_s$.

The device's own thermal emission, together with absorbed ambient radiation, produces an output proportional to $\int A(\omega;E_F)\,B(\omega,T)\,d\omega$, where $B$ is the thermal spectrum at temperature $T$. The factor $A$ appears in both absorption and emission because of Kirchhoff's law. The thermal curve is smooth and monotone-like in $E_F$.

We claim that an unbinned likelihood over (gate voltage, readout power) pairs will reach a lower minimum detectable line power than a static absorber of the same area and detector noise. It will do so under realistic slow drifts in background level. It will also do better than the same gated device analyzed step-by-step with power thresholds. The advantage should grow as background drifts become larger relative to white noise, because a static device has no "signal-off" state against which to measure drifts.

Relation to Prior Work

Three device papers supply the hardware and its limits.

We reuse their Kubo-conductivity modeling to build $A(\omega;E_F)$. None of them treats the gate sweep as a measurement axis, and none addresses weak-signal detection. We also take the −9.5 dB gain as a constraint: any design that relies on high radiation efficiency is ruled out.

The statistics direction supplies the method.

The missing-energy direction supplies the logic and the failure modes to guard against.

Research Program

The first component is a calibrated forward model. We will measure $A(\omega;E_F)$ on a fabricated stack-patch absorber across the 220–325 GHz band used in the stack-patch work. This band corresponds to masses of about 0.9–1.3 meV. The absorptivity will be measured by direct reflection and transmission, as $A = 1 - R - T$, at many gate voltages. We will also map gate voltage to $E_F$ repeatedly to quantify hysteresis.

The relaxation time and the voltage-to-$E_F$ map become nuisance parameters in the fit rather than fixed constants. The gate range will be capped below the dielectric-breakdown margin. The file flags breakdown and leakage as real risks at the fields near 1 V/nm needed for 1–2 eV swings. Part of the question is therefore how much sensitivity survives when the usable tuning range is modest.

The second component is the likelihood. Each measurement is a readout power at a gate setting and time. The model contains four terms: the signal template; a thermal-background term built from the same $A$; a slow drift term; and white detector noise sized by the measured noise-equivalent power. We scan the signal frequency and correct for trials across the scanned band, following the ABRACADABRA practice. Frequentist coverage will be checked with toy experiments.

The third component is a device-design study. Using the forward model, we will ask how sensitivity depends on relaxation time, tuning range, and the ratio of drift to white noise. The aim is to state which improvements to graphene devices matter for this use.

Baselines

The naive reference point is a static broadband absorber read out by cut-and-count. It uses the same absorber area and detector noise with the gate fixed at its highest-absorption point, and it looks for excess power above a threshold. It shows what the gate electrodes, sweeps, and modeling must beat to justify their cost.

The strongest competitor is a static broadband absorber with the best reported graphene detector noise. We will set its noise-equivalent power to 0.1 pW/Hz^1/2, the lowest value in Recent Progress in the Development of Graphene Detector. It will be analyzed with the binned power-spectrum and look-elsewhere pipeline of Design and implementation of the ABRACADABRA-10 cm axion dark matter search.

This baseline is the real threat. With a quality factor of order unity, a broadband absorber collects nearly the same signal power from a surface-converting dark-matter field, and it integrates the whole band all the time. Its high-efficiency form is represented by the simulated 99.3% average absorption of the metasurface absorber paper. We found no reported result on Lacuna for a terahertz dark-matter limit from any graphene device, so the comparison must be made in our own simulation and injection tests.

For fairness, every method gets identical total observation time, absorber area, detector noise, drift spectra, and injected signals. Every method also gets the same budget for tuning its analysis choices, fixed before unblinding.

The third baseline is the gated device analyzed step-by-step. Each gate setting is thresholded independently and the results are combined afterward. Comparing it with the full likelihood isolates the gain from fitting the template shape jointly rather than merely owning a tuning knob.

The key ablation replaces the full template with a two-state on/off contrast between high and low absorptivity. This removes the frequency-dependent peak shape and keeps only amplitude modulation. A second ablation fits with the relaxation time fixed at a deliberately wrong value, to measure how much sensitivity depends on model accuracy.

Beating the strongest baseline means a factor of two or more lower minimum detectable line power, at a 5σ post-trial false-positive rate, under drift levels matching those measured on the device. The gain must also hold with coverage verified.

Why not just…?

Why not just use a cryogenic metallic cavity? A cavity's size is set by the wavelength, which is 1 mm or less here, so its volume and conversion power shrink with it. The corpus argues this through Ultramassive MIMO Systems at Terahertz Bands and the ABRACADABRA scaling discussion. We do not test cavities directly. We do not claim graphene beats a high-Q cavity where a cavity is practical.

Why not just use a static broadband absorber and integrate longer? This may well win, and the proposal is built to find out. If background drifts are negligible compared with white noise, the broadband device should match or beat the gated one, because it never spends time off-resonance. A static device cannot switch the dark-matter signal off, so drifts it cannot separate from signal set its floor. The gate sweep supplies a partial "off" state with a predicted shape. The design study will map the crossover drift level. The injection test will then tell us which side of the crossover real devices sit on. If the crossover lies above realistic drift, we will report that tuning is not worth it.

Why not just threshold each gate step, or use a binned power spectrum? The τ→μγ search showed that unbinned joint fits recover correlation information that cuts discard. The stepwise baseline and the on/off ablation measure exactly how much that gain amounts to here.

Why not just use a model-free anomaly detector like the normalizing-flow method from the CMS study? That method interpolates backgrounds from sidebands of a data axis. In our data, the gate axis carries a known physical signal shape, so discarding it wastes information. We will still run a flow-based sideband estimator as a secondary check. If it matches the template fit without any model, that finding would argue against investing in the calibrated forward model.

Proposed Experiments and Decisive Checks

The laboratory test injects signals into the fabricated absorber. A weak, attenuated line from a tunable source in the 220–325 GHz band is set near the detector's noise floor. At the same time, controlled broadband thermal loads and slow temperature drifts are imposed. Injections are blinded: frequency and amplitude are chosen by someone outside the analysis. We will report recovered significance, bias in the fitted frequency and power, and false-positive rate.

The decisive null test is modeled on the same-sign check of the CMS study. We will run long sweeps with no line but with the broadband load deliberately varied. The template fit must not produce fake peaks above the trial-corrected expectation. If the Kirchhoff-coupled thermal term imitates a line peak, the central claim fails.

The second decisive check is the comparison with the strongest baseline at matched exposure. It must show the factor-of-two gain at measured drift levels, or the proposal concludes that static absorbers suffice.

The third check reruns the analysis with the gate range capped at progressively safer voltages. This shows whether the advantage survives the tuning range that real dielectrics can hold.