Designing Anisotropic Metasurfaces for Few-Photon Measurements

Engineers usually design polarization-controlling metasurfaces by making them match a target behavior as closely as possible. A metasurface here means a thin sheet patterned with sub-wavelength metal or dielectric shapes. The targets are things like a high polarization conversion ratio, a deep extinction ratio, or a small error against a desired field pattern. These devices are increasingly placed at the front of detectors that see only a handful of photons or events. In that regime, what limits the science is counting noise: with $N$ counts, relative uncertainty falls only as $1/\sqrt{N}$. In such a detector the design that best matches a target is not necessarily the one that extracts the most information per detected count. Fabrication errors make this worse. They leave behind a response that is slightly unknown and slightly frequency-structured, and in a low-count search that structure can look like a real signal.

We propose to design anisotropic metasurfaces against a statistical objective instead of a target-matching one. The objective is the Fisher information that the detected counts carry about the physical quantity of interest. Fisher information measures how sharply the data pin down a parameter, and its inverse is the best variance any unbiased estimator can reach. We average it over fabrication and incidence-angle uncertainty, and we treat that uncertainty as a nuisance to be marginalized rather than ignored. The quantity we report is an exposure-equivalent gain: how many times longer a conventionally designed device would have to integrate to match the statistically designed one. That number speaks directly to rare-event experiments, where the only alternative is to scale up the detector. Telescope Array, for example, is quadrupling its array to push a 3.4σ hotspot toward discovery (Telescope Array anisotropy).

Core Claim

For a count-limited polarimetric measurement made through an anisotropic metasurface, the claim is that the design maximizing expected Fisher information beats designs chosen for high polarization conversion ratio or low response error. It should give a measurable exposure-equivalent gain, plausibly tens of percent, and a lower false-detection rate under fabrication uncertainty. A gain of this size is not guaranteed; the experiments below are built to find out whether it exists at all.

The mechanism is simple. Suppose the detector sorts counts into polarization or frequency channels $k$, with expected counts $\lambda_k(\theta, d, \xi)$. These depend on the signal parameter $\theta$ (for example, a weak degree of linear polarization), the metasurface geometry $d$, and unknown fabrication and angle perturbations $\xi$. The Poisson Fisher information is

$$I(\theta; d, \xi) = \sum_k \frac{\big(\partial \lambda_k / \partial \theta\big)^2}{\lambda_k},$$

where the denominator includes background counts. This weighs response errors very differently from a squared-error-to-target loss. An error matters only insofar as it shrinks the contrast that $\theta$ produces between channels, relative to the counts already present in those channels. So, for example, a small extinction-ratio improvement in a low-background channel can be worth more than a wide bandwidth.

Because $\xi$ is not perfectly known, the design objective is the expected efficient information. This is the information about $\theta$ that remains after projecting out the directions that $\xi$ can mimic:

$$J(d) = \mathbb{E}_{\xi \sim p(\xi)} \Big[ I_{\theta\theta} - I_{\theta\xi} I_{\xi\xi}^{-1} I_{\xi\theta} \Big].$$

Here $p(\xi)$ is the fabrication-tolerance distribution, and the $I$ blocks are the Fisher information matrix entries for $\theta$ and $\xi$. The second term is what makes fabrication-induced spectral "bumps" costly: a design whose response errors resemble a signal loses information it would otherwise keep. The exposure-equivalent gain of design $d$ over a baseline $d_0$ is then $G = J(d)/J(d_0)$. The form follows directly from Fisher information adding up linearly with exposure for Poisson counts.

Prior Work and How This Differs

The method comes from the deep-learning metasurface design line. Xu et al. train a small neural network that predicts each meta-atom's local near-field response from its five-layer neighborhood. It is 5,000 times faster than full-wave simulation, with under 3% energy-normalized error, and they invert it non-iteratively toward target field profiles. Neural Metamaterial Networks show that smooth neural surrogates using the Swish activation give reliable analytic derivatives. That property matters here, because Fisher information requires derivatives of the surrogate's output with respect to $\theta$ and $\xi$, and its gradient requires derivatives with respect to the geometry $d$. We reuse both ideas, a local surrogate and smooth, differentiable outputs.

Both papers are deterministic and target-matching. Neither models measurement statistics or tolerance distributions, and replacing their losses is exactly our contribution.

The devices come from anisotropic polarization engineering. The slanted I-dipole reflective rotator converts linear polarization across 15–45 GHz. Its measured polarization conversion ratio is above 0.92, its insertion loss is below 0.3 dB, its angular stability holds over 0–20°, and its fabricated dimensions are within 10 µm of the design. The hybrid plasmonic mercury cadmium telluride absorber reaches 74% absorption in a submicron film and an extinction ratio of about 12 with a grating backing. Both papers optimize and report device figures of merit. Neither asks what those figures buy in a count-limited estimate.

The motivation comes from the counting-limited experiments. The Telescope Array hotspot dropped from 5.1σ to 3.4σ after trials correction, which shows that a multi-bin search pays heavily for every spurious feature. The EXO-200 analysis gained about 25% in sensitivity from neural-network event classification. That benchmarks what software alone typically gives, and it is the scale of gain our hardware-side objective should be compared against. The cryogenic RF diagnostics work shows that detector responses drift with temperature and magnetic environment. This is why we do not assume that a one-time calibration removes $\xi$.

Research Program

We will build two differentiable surrogates. The first maps the rotator's unit-cell parameters, incidence angle, and frequency to its reflection Jones matrix, the 2×2 complex matrix that describes how the surface transforms each polarization. The second maps the plasmonic absorber's cuboid and grating geometry to its polarization-resolved absorption. Each surrogate will be trained on full-wave simulations with a Swish-activated network, and its error checked against held-out simulations. Following Xu et al., we will keep it below a few percent.

On top of each surrogate sits a count model. It turns the device response, a source with unknown weak polarization $\theta$, a background rate, and an exposure into expected channel counts. From the count model we compute $J(d)$ by automatic differentiation and Monte Carlo over $p(\xi)$. That distribution includes about 10 µm dimensional jitter and substrate permittivity spread, matching the rotator's reported tolerance, along with incidence-angle spread.

We will run three versions of $p(\xi)$ that represent three calibration states. In the "fully uncalibrated" state, $\xi$ is unknown at analysis time. In the "calibrated once" state, $\xi$ is measured, but slow drift is added afterward. In the "perfectly calibrated" state, $\xi$ is known exactly. Designs are optimized by gradient ascent on $J(d)$ from multiple starting points. Every finalist is then re-verified in full-wave simulation, so that surrogate exploitation does not inflate the gains.

Baselines

The naive reference point is the published geometry as is: the rotator at its reported dimensions, and the absorber with the gold-grating backing. Each is evaluated through the same count model. The file reports their device metrics, a measured polarization conversion ratio above 0.92 for the rotator and an extinction ratio of about 12 for the absorber. There is no reported result found on Lacuna for the Fisher information or count-limited estimation performance of either device, so we compute that ourselves.

The strongest existing method is the target-matching approach of Xu et al., run on the same surrogate. It is the strongest because it is what practitioners actually use, and its full-wave fidelity is established at under 3% error. For the comparison to be fair, both pipelines use the same surrogate, the same geometry bounds, the same optimizer, and the same number of surrogate evaluations and random restarts. The target-matching loss is set to an ideal 90° rotator for the first testbed and an ideal polarizer for the second.

A third baseline is robust target-matching, which minimizes response error averaged over the same $p(\xi)$. It isolates whether any gain comes from knowing the tolerance distribution or from the statistical objective itself.

Two ablations isolate the key ingredients. The first optimizes plain $I_{\theta\theta}$, without the nuisance-projection term, which isolates what marginalizing fabrication uncertainty adds. The second runs the statistical objective with the background rate set to zero, which tests whether the advantage depends on the count regime.

Beating the strongest baseline means a median exposure-equivalent gain $G$ clearly above 1 across the fabrication draws, re-verified in full-wave simulation. The target is at least the roughly 25% that EXO-200 got from software. It also means a lower false-detection rate in the null tests described below.

Why not just…?

Why not just collect more counts? Exposure is the denominator of our metric, so this alternative is measured, not dismissed. A gain $G$ is exactly the exposure the conventional device would need to add. The proposal is worthwhile only where hardware redesign is cheaper than a fractional increase in exposure. For a fixed cryostat or array, the Telescope Array upgrade shows how expensive that increase is.

Why not optimize polarization conversion ratio and then calibrate the fabricated device? This is the most serious threat. If calibration were perfect and stable, the nuisance term would vanish, and any remaining advantage would have to come from the mismatch between the Fisher objective and the target-matching loss. The "perfectly calibrated" arm tests exactly this. If $G$ falls to about 1 there, and the drift arm shows that drift is negligible, we will report that the contribution reduces to tolerance-aware calibration, not design. The cryogenic drift results in the RF diagnostics paper make that outcome less likely in practice, but not impossible.

Why not keep the standard design and put a better machine-learning estimator on the counts? Fisher information bounds every unbiased estimator. If the hardware does not encode the parameter in the counts, no analysis can recover it, much as no software recovers light that the bare absorber film reflected. We will check this directly by running a learned estimator on counts from the baseline design. If it reaches the baseline's Cramér–Rao bound but not the statistically designed device's bound, the gap is a hardware gap.

Why not just use robust optimization of the usual loss? It is included as a baseline. If it matches $J$-optimization, the lesson is that tolerance awareness, not the statistical objective, is what matters. We would then report it that way.

Proposed Experiments

The first check is information. For each testbed, calibration state, and background level, we report the distribution of $G$ over 1,000 fabrication draws, for all designs, after full-wave re-verification. We also report the surrogate's error on the specific finalist designs, so that a gain created by surrogate error is visible.

The second check is a null test for false discoveries, modeled on the Telescope Array trials correction. We simulate count spectra with no polarized signal through fabricated-but-misreported responses. We then run a sliding search for a polarized spectral line, and estimate the post-trial false-detection rate with 10⁶ Monte Carlo null realizations. The prediction is that target-matched designs show excess false positives from fabrication-induced bumps, and that $J$-optimized designs suppress them. The no-nuisance ablation should sit in between.

The third check is discovery power. We inject weak polarized lines at a few count levels, and compare post-trial detection power at fixed false-positive rate across designs. The expected ordering should be consistent with $G$.

As a stretch goal, we fabricate the best and baseline rotator geometries on the same PTFE process and measure them in the anechoic-chamber setup the rotator paper used. The measured Jones matrices go into the count model to confirm that the predicted gain survives real fabrication.

The idea fails cleanly if $G$ stays within noise of 1 in the uncalibrated and drift arms, or if false-positive rates do not separate. Either outcome would tell metasurface designers that their usual target-matching figures are already adequate for count-limited use.