Does Optical Loss Set the Scaling Exponent of Neural Maxwell Surrogates?

Neural networks are increasingly used as fast stand-ins for electromagnetic solvers. Such a network, called a surrogate, maps a nanostructure's shape to the light field around it hundreds of times faster than a numerical simulation. The structures that matter most at the smallest scales are metallic, or plasmonic: light couples to oscillating electrons at the metal surface, and those electrons lose energy as heat by colliding with the lattice. This heating, called Ohmic dissipation, broadens optical resonances and shortens how far light travels along the metal. It cannot be designed away below about 100 nm, because dielectric cavities stop confining light there. Yet nobody knows how surrogate accuracy grows with model size and data on such lossy problems. The best physics-aware electromagnetic surrogate was trained only on lossless silicon, at one fixed size. Neural scaling laws, which predict how loss falls as a power of parameters and data, have been measured for text and fluid simulation, but never for wave physics with absorption.

We propose to treat the metal's electron damping rate as a single continuous dial and to measure how the scaling exponents of surrogate error move as the dial turns. Damping has two opposite effects on learnability. It smooths the map from geometry to field, because broad resonances no longer swing wildly under small shape changes, and that should make the problem easier to scale. It also concentrates field gradients and heating inside a skin depth of a few tens of nanometers, which should make the problem harder for any architecture with limited spatial resolution. Our claim is that these two effects can be measured before training large models. Together they should predict both the exponent and which architecture family is compute-optimal at a given loss level.

Core Claim

Write surrogate error as $L(N,D) = E + A N^{-\alpha} + B D^{-\beta}$, where $N$ is parameter count, $D$ is the number of simulated training layouts, $E$ is irreducible error, and $\alpha,\beta$ are the exponents. We claim that $\alpha$ and $\beta$ are systematic functions of the Drude damping rate $\gamma$ (the electron collision rate in the metal's permittivity model). We also claim that this dependence is predictable from the eigenvalue decay of the target fields.

The prediction comes from the solvable random-feature theory of A Dynamical Model of Neural Scaling Laws. That paper derives $L(N) \propto N^{-(a-1)}$, where $a$ is the rate at which the target function's power decays across the eigenmodes of the model's kernel. Resonance broadening, which lowers the quality factor $Q=\omega_0/\Delta\omega$, should raise $a$ and steepen scaling. Skin-depth localization should add slowly decaying high-frequency spatial content and flatten it. The net effect should be a non-monotonic $\alpha(\gamma)$ with a measurable turnover.

A second, architectural claim follows from physics. Damping sets the plasmon propagation length, the distance a surface wave travels before it is absorbed. That length ranges from about 100 μm in low-loss air-host geometries (Long-range air-host plasmonic propagation with subwavelength confinement) to sub-micron in lossy metals. Local architectures such as convolutional UNets should be compute-optimal when this length is shorter than their receptive field. Global ones, attention or Fourier operators, should win when it is longer. The crossover should therefore track $\gamma$ in a predictable way.

Research Program

There is no public plasmonic surrogate benchmark, so we will build one. We start from the 2D periodic metagrating setting of WaveY-Net, which trained a physics-augmented UNet on 30,000 silicon layouts solved with the Ceviche finite-difference frequency-domain solver. That solver handles a complex permittivity at one frequency directly, so dispersion costs nothing extra. We keep the layout distribution and replace silicon with a Drude metal whose damping is swept over roughly eight values. The sweep runs from near-lossless, through silver-like and gold-like damping, to strongly over-damped. We add a lossless dielectric anchor that reproduces WaveY-Net's setting, and a small set of illumination wavelengths so that resonances are actually crossed.

For each damping value, we will train three architecture families across a grid of widths, depths and dataset sizes:

For the Transformer, we use the depth parameterization of Don't be lazy: CompleteP. That method transfers tuned learning rates across depth, so deep and shallow models are not unfairly detuned. Scaling laws will be fit to the lowest-error frontier at each compute level, the convex-hull method of Gemstones. Gemstones showed that sloppy fitting alone moves exponents between 0.45 and 0.80. We will also fit the size-dependent-exponent form of Farseer to check whether a single power law is even adequate.

In parallel, and before any large training, we compute the predicted exponent for each damping value. We take the eigendecomposition of the empirical kernel of a small random-feature model on a few thousand samples, measure how the target's power decays across modes, and convert that decay rate $a$ into a predicted $\alpha$. We will report three error metrics: relative field error, diffraction-efficiency error, and the error in absorbed power $\int \omega\,\varepsilon_2 |E|^2$. The last is the quantity a plasmonic designer actually cares about and the one most sensitive to skin-depth errors.

Baselines

Naive reference. A nearest-neighbor lookup that returns the simulated field of the closest training geometry. This is the field-prediction analogue of the "Train Best" baseline that IDToolkit found frequently beats generative inverse designers. It uses the same training sets at every data size. A trained surrogate has to beat it at matched $D$, or its scaling curve means nothing. We expect the lookup to be strongest at high damping, where fields vary smoothly with geometry, and that is itself a test of the smoothing hypothesis.

Strongest existing method. WaveY-Net. Its reported number is 0.67% diffraction-efficiency error, against 3.7% for a purely data-driven UNet, on the silicon metagrating dataset (WaveY-Net). We will retrain it ourselves on our lossless anchor to confirm we reproduce that regime, then on every damping value with the same data, the same tuning budget per run, and the same compute grid as the other families. For lossy metals there is no reported result on Lacuna for any surrogate. Beating WaveY-Net therefore means one of two things. Either a family reaches lower absorbed-power error at matched compute on lossy data, or the exponent prediction correctly identifies in advance where WaveY-Net's architecture stops being compute-optimal.

Scaling-law reference. The only physics-surrogate scaling result in the file is $N \propto C^{0.75}$ for compute-optimal mesh Transformers on fluid flow (Training Transformers for Mesh-Based Simulations). We test whether lossless Maxwell data reproduces something near it and whether damping moves it.

Key ablation. The ablation separates the two mechanisms. We build a lossless-geometry dataset in which absorption is added uniformly to the dielectric background. This shortens propagation length and broadens resonances without creating metallic skin-depth gradients. If this control shows only the steepening branch of $\alpha(\gamma)$, that supports the decomposition. A second ablation normalizes each target field to unit energy. This removes the trivial effect that high-Q fields simply have larger amplitudes.

Why not just…?

Why not just fit one scaling law per dataset and skip the theory? That is what we do as the measurement step. The difference is the forecast. The useful output is predicting the exponent at an untested damping value from a cheap spectral measurement. A per-dataset fit is the comparison condition for that forecast. If kernel-spectrum predictions are no better than interpolating fitted exponents between neighboring damping values, the theoretical component fails, and we will report it.

Why not just stratify a single model's error by Q factor? This is a real threat. If error depends only on each sample's linewidth, then damping matters only through the per-sample Q distribution, and a Q-reweighted lossless dataset would reproduce everything. We test this directly. We construct a lossless dataset resampled to match the Q histogram of a lossy one and compare its scaling exponents. Matching exponents would mean the skin-depth branch is negligible and the claim reduces to "Q sets difficulty." Diverging exponents isolate the metallic contribution.

Why not just run the solver, since it is parallel? IDToolkit obtained an 11.6× speedup on 16 cores with MEEP. That is two orders of magnitude short of WaveY-Net's roughly 700× single-run and 10,000× batched speedups. Gradient-based freeform design needs millions of differentiable forward passes. The open question is not whether surrogates are faster but how large they must be to be trusted on lossy metals, which is exactly what the exponents answer.

Why not just use the largest attention model everywhere? If global models dominated at all damping values and compute levels, the architectural crossover claim would be false. The cross-family compute frontiers will show this directly. The prior expectation is mixed: masked Transformers beat message-passing networks on fluids, but local convolution is a natural fit for short-range fields.

Decisive Checks

The proposal succeeds if three results hold.

  1. The fitted $\alpha$ varies with damping by more than the fitting variance Gemstones documents, measured with bootstrap intervals over frontier runs.
  2. Exponents predicted from the kernel spectrum of small models forecast the large-model exponents at two held-out damping values better than interpolation does.
  3. The compute crossover between UNet and attention families moves with propagation length in the predicted direction.

It fails cleanly if exponents are flat in damping after field normalization, or if the Q-matched lossless control reproduces them. In either case, damping would be an amplitude effect rather than a change in the learning problem.

A secondary check addresses the out-of-distribution "surrogate trap" that IDToolkit identified. We hold out geometries with gaps below 20 nm and test whether scaling closes the gap on them or whether their error plateaus at the irreducible term $E$. A plateau would mark where classical Drude physics, and not model size, becomes the limit. Fundamental Scaling Laws in Nanophotonics places that limit for classical Drude models below 2 nm.