Symmetry-Certified Search for Isolated Photonic Weyl Points

Some crystals have energy-band crossings that cannot be removed by any small change to the structure, because a symmetry of the lattice forbids the two bands from mixing. The strongest such guarantees come from nonsymmorphic symmetries, meaning rotations or mirror reflections combined with a fractional shift of the lattice (a "screw axis" or a "glide plane"). A protected crossing alone is of limited use, though. In the low-index polymer photonic crystal that realized a charge-2 Weyl point pinned by a screw axis, other bulk bands sit at the same frequency, so the interesting surface states leak into the bulk as "leaky resonances" (Observation of a Charge-2 Photonic Weyl Point in the Infrared). The open design problem is to keep the protected crossing and also clear the frequency window around it of every other band. Today this is done by hand, using forward electromagnetic simulations that are expensive to run.

This note proposes treating that design problem as an architecture search and borrowing what the neural architecture search (NAS) community learned about controlling search cost. Group theory acts as an exact, zero-cost filter: any candidate geometry whose space group contains the right screw axis is guaranteed to host the crossing at the target point of momentum space, without a single simulation. The expensive solver is spent only on what symmetry cannot guarantee, namely how isolated the crossing is in frequency. It is called at the lowest numerical resolution that still ranks candidates correctly, an idea taken from budgeted performance estimation in NAS. The claim is that this pairing reaches a given isolation window with far fewer solver calls than black-box search over unconstrained geometries, and that it can settle whether a complete isolation window exists at all at refractive index 1.52.

Core Claim

In NAS, "zero-cost proxies" score an untrained network from one forward and backward pass. Their weakness is that they are empirical correlates with no theory explaining why they work (Zero-Cost Operation Scoring in Differentiable Architecture Search). In nonsymmorphic band theory the situation is reversed. Whether bands must stick together at a zone-boundary point or at $\Gamma$ follows exactly from the irreducible representations (irreps) of the space group, the symmetry "labels" that bands can carry. One example is the Berry phase of exactly $\pi$ around the nodal lines of BaTaS$_3$, enforced by the screw symmetry of $P6_3/mmc$ (Node-surface and node-line fermions from nonsymmorphic lattice symmetries). Without symmetry, a three-dimensional crossing generically gaps out (Dirac semimetal and topological phase transitions in A3Bi).

This exact certificate removes the hardest part of the search, which is finding a crossing that exists only when parameters are finely tuned. The search objective then reduces to one smooth-ish scalar, the isolation window

$$W = \frac{\min_{\mathbf{k},\, n \notin \{\text{Weyl bands}\}} |\omega_n(\mathbf{k}) - \omega_W|}{\omega_W},$$

where $\omega_W$ is the frequency of the protected Weyl point, $\omega_n(\mathbf{k})$ are the frequencies of all other bulk bands at wavevector $\mathbf{k}$, and $W>0$ means no other bulk state shares the Weyl frequency anywhere in the Brillouin zone (the cell of distinct wavevectors). The new contribution is not a symmetry-constrained parametrization alone, which physicists use implicitly when they hand-design. It is a search procedure that enumerates symmetry-compatible structural motifs and allocates simulation budget across fidelities, together with a measured accounting of how many solver calls the symmetry certificate saves.

Research Program

The search space is built in two tiers. The discrete tier enumerates structural motifs: rod or ellipsoid primitives placed at Wyckoff positions (the special symmetric sites of a space group) of space groups that contain a $4_1$ or $6_1$ screw axis or a glide. Each motif is kept only if the irreps at the target point force the degeneracy and give a nonzero topological charge. This check is pure linear algebra, done in microseconds per motif. The continuous tier optimizes the dimensions, orientations, and fill fraction of each surviving motif. Every parameter change is applied through the symmetry operations, so every intermediate geometry keeps the screw axis.

This design responds to a lesson from differentiable NAS. In DARTS, a continuous blend of candidates differs from the discrete architecture finally chosen, and performance can fall off a cliff at the point of discretization (Understanding and Robustifying Differentiable Architecture Search). In a photonic crystal, a naive blend of two geometries generally breaks the screw axis and destroys the very crossing being optimized. Our search therefore never relaxes across motifs; it relaxes only within the symmetric subspace of one motif.

Evaluation uses a plane-wave eigensolver at several resolutions, from coarse plane-wave cutoffs and sparse k-grids up to converged settings. Following Rethinking Performance Estimation in Neural Architecture Search, which finds the minimum training budget that still ranks networks reliably, we first measure the cheapest fidelity whose ranking of $W$ agrees with the converged ranking. A successive-halving or Bayesian optimization loop then promotes candidates through fidelities. The final candidates are checked independently. We compute the Chern number on a small sphere around the Weyl point to confirm charge 2, and run RCWA transmission simulations (rigorous coupled-wave analysis, the transmission solver used in the woodpile work) to confirm that surface arcs no longer overlap bulk states.

The fabrication constraint stays fixed at $n=1.52$ with two-photon-polymerization-compatible feature sizes. A second run at a higher, generic dielectric index tests whether any isolation obstruction comes from the index or from the motif.

Baselines

The naive reference point is the published chiral woodpile (Observation of a Charge-2 Photonic Weyl Point in the Infrared), whose 45° layer rotation enforces a $4_1$ screw axis, with a grid sweep over rod width and height. It is the natural starting point because it is the only realized device of this kind in the file. That paper reports no isolation window; it reports qualitatively that no complete gap exists and that the arcs are leaky. There is no reported result found on Lacuna for $W$ of any screw-symmetric photonic crystal, so we will compute the woodpile's $W$ ourselves at converged fidelity, and that number becomes the bar. The sweep gets the same total solver budget as every other method.

The strongest competitor is unconstrained black-box search: Bayesian optimization, plus a genetic algorithm, over voxelized unit cells with no symmetry imposed. Crossings are detected numerically by dense k-sampling, and topological charge is checked by a Chern-number computation. This is the approach the file's reviewer objections name as sufficient. It is also the photonic analogue of early black-box NAS, whose cost crisis motivated efficient search (A Survey of Methods for Low-Power Deep Learning and Computer Vision). It runs with the same solver, the same fidelity ladder, the same wall-clock budget, and its own hyperparameter tuning budget matched to ours.

Because NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search found that random search often matches sophisticated optimizers, we also run random search inside the symmetry-certified space. Two ablations isolate the key ingredients. The first removes the symmetry tier but keeps multi-fidelity evaluation. The second keeps the symmetry tier but uses only converged single-fidelity evaluation.

Beating the strongest baseline means one of two outcomes at matched solver budget. Either we reach a strictly larger converged $W$, or we reach the woodpile's $W$ (or the first $W>0$) with at least an order of magnitude fewer full-fidelity solver calls. Every reported design must have a verified charge-2 crossing.

Why not just…?

Why not just hand-pick a space group and sweep its parameters, as physicists already do? This is the most serious threat. If the woodpile family already contains the best design, a careful sweep will find it and the search adds little. The motif-enumeration tier exists to test exactly this. If every motif with top $W$ falls in the woodpile family, we will report that the hand design was near-optimal, which is a useful negative result. If motifs at other Wyckoff positions or with glide-plus-screw combinations dominate, the enumeration has earned its place.

Why not just run Bayesian optimization over free-form geometry and let it find crossings? A crossing without symmetry protection is a finely tuned event, so a black-box optimizer must first hit a set of measure zero in parameter space before $W$ is even defined, and every evaluation needs dense k-sampling. The first ablation (no symmetry tier) measures this cost directly as solver calls to the first verified crossing. If unconstrained search turns out to rediscover screw-symmetric structures cheaply, the efficiency claim fails, and the experiment will show it.

Why not just train a neural surrogate that predicts band structures and search over that? A learned surrogate is a zero-cost proxy with the same weakness Zero-Cost Operation Scoring in Differentiable Architecture Search identifies: no guarantee, and degeneracies are exactly where smooth regressors fail. A surrogate could still help in the continuous tier. We will test one as an extra fidelity level, judged only by rank agreement with converged $W$, rather than letting it certify crossings.

Why not just use a DARTS-style differentiable relaxation over motifs? Blending motifs breaks the symmetry that protects the crossing. This is the photonic counterpart of the discretization gap and skip-connection collapse documented in Fair DARTS: Eliminating Unfair Advantages in Differentiable Architecture Search and DARTS-: Robustly Stepping out of Performance Collapse Without Indicators. We will run one relaxed variant and report how often its discretized output loses the certified crossing.

Decisive Checks

The first check comes before any search. Over a few hundred random symmetric candidates, we measure the Spearman rank correlation between low-fidelity and converged $W$. If no affordable fidelity ranks reliably, the cost argument collapses, and we will say so.

The second check is the matched-budget comparison against unconstrained Bayesian optimization and against in-space random search, reported as curves of best converged $W$ against solver calls.

The third check is physical. For the best design at $n=1.52$, a positive $W$ together with RCWA spectra showing surface arcs outside the bulk continuum would directly answer the leaky-resonance problem left open by the woodpile experiment. A provable $W\le 0$ across all certified motifs, with $W>0$ appearing only at higher index, would instead establish an index threshold that the community currently lacks.