Optical neural network chips do matrix multiplication by sending light through grids of tunable interferometers. Each interferometer is a Mach–Zehnder interferometer (MZI), a small device that mixes two waveguides by an amount set by a phase shifter. These devices are micrometers across, while transistors are nanometers across, so chip area runs out quickly. A mesh that can realize any matrix needs a number of MZIs that grows with the square of the matrix size. Cheaper meshes with roughly $N \log N$ MZIs, such as butterfly-shaped layouts, can still train networks well. Among meshes with the same MZI count, though, some layouts work and others fail, and at present the only way to tell them apart is to train each one. We propose a score that can be computed before any training, using one backward pass. It measures how much of the gradient an unconstrained dense layer would want to follow can actually be followed by turning the phase knobs of a given mesh. Candidate layouts are ranked by this "gradient capture ratio," optionally averaged over simulated fabrication and phase errors.
This is exactly where training-free performance estimation, which scores neural architectures from their state at random initialization, should matter most, and where its standard tools break down. The strongest critique of such scores is that they often rank architectures no better than counting parameters does (ZiCo; Evaluating Efficient Performance Estimators of Neural Architectures). Under a fixed chip-area budget, every candidate has the same component count, so counting parameters gives no ranking at all. The most robust existing scores, however, count ReLU activation patterns. Photonic layers are interference transforms read out by homodyne detection (mixing the output light with a reference beam to recover the field's real part), and they contain no ReLUs. We want a score built from the physics of the mesh itself.
Core Claim
Take a mesh with phase settings $\phi$ that realizes the weight matrix $W(\phi) = \mathrm{Re}(U(\phi))$. This is the pseudo-real readout from Scalable and compact photonic neural chip with low learning-capability-loss, which introduced $O(N\log_2 N)$ FFTUnitary and P-GridUnitary meshes and showed that universal meshes are over-provisioned for neural networks. At a random initialization $\phi_0$, let $g = \mathrm{vec}(\partial L/\partial W)$ be the gradient of the task loss $L$ with respect to the realized weights, computed on one minibatch. Let $J = \partial\, \mathrm{vec}(W)/\partial \phi$ be the mesh Jacobian, whose columns are the directions in weight space that the phase knobs can move. We define the capture ratio
$$C = \frac{\lVert P_J\, g \rVert^2}{\lVert g \rVert^2},$$
where $P_J$ is the orthogonal projection onto the column span of $J$. $C$ is the fraction of the dense layer's desired first-order update that the mesh can reproduce. For multi-layer networks, the per-layer ratios are combined with weights proportional to each layer's $\lVert g_l\rVert^2$. The noise-aware version averages $C$ over draws of phase error, per-MZI insertion loss, and neighbor thermal crosstalk applied to $\phi_0$.
The claim is that $C$ ranks iso-component mesh layouts by their trained accuracy substantially better than two alternatives. The first is generic proxies transplanted from electronic networks. The second is a task-agnostic measure of how many weight-space directions the mesh can reach. The second comparison is what makes this more than an expressivity count. Two meshes can reach subspaces of the same dimension while differing in whether those subspaces contain the directions a particular task needs.
Research Program
The first component is a benchmark, because none exists. We will enumerate families of iso-MZI mesh layouts at $N \in \{8, 16, 64, 256\}$. These include FFTUnitary, P-GridUnitary, truncated Clements grids, butterflies with permuted stage orderings, and randomly placed MZI sets at matched count and matched depth. Each layout is used as the linear layer of small photonic networks with detector-style nonlinearities. We train them in simulation on MNIST, Fashion-MNIST, and CIFAR-10, under both ideal and noisy device models. This gives ground-truth accuracy for every candidate.
The second component is the proxy itself. We compute $J$ by automatic differentiation through the mesh. At large $N$, where $J$ has on the order of $N^2 \times N\log N$ entries, we use a randomized sketch of its column space. We also ask whether $C$ at initialization is stable across seeds and minibatches. Evaluating Efficient Performance Estimators of Neural Architectures found that proxies are only reliable at true initialization, and our score inherits that constraint.
The third component combines $C$ with area and loss. We build a Pareto search that takes MZI count, depth (which drives insertion loss), and waveguide crossings as layout costs, in the spirit of the placement and routing tools Apollo and LiDAR, and uses $C$ as the accuracy predictor. HaeNAS pairs zero-cost proxies with digital latency models. We replace both halves: a photonic cost in place of the latency model, and a mesh-native score in place of generic proxies. System energy from converters and memory, which Architecture-Level Modeling of Photonic Deep Neural Network Accelerators shows can dominate, is out of scope beyond reporting it for the final selected designs.
Baselines
The naive reference points are random ranking and component count. Within each iso-MZI family, component count is constant by construction and scores exactly zero. Across families it is a fair baseline, and we report it there. For context, ZiCo reports that parameter count reaches Kendall's $\tau = 0.57$ on NATSBench-TSS, just below ZiCo's 0.61 and above Synflow's 0.54. No reported result was found on Lacuna for any proxy's rank correlation on a photonic mesh search space.
The strongest existing methods are ZiCo and the empirical neural tangent kernel condition number from TEG-NAS. ZiCo scores a network by the ratio of gradient mean to gradient spread across batches. The neural tangent kernel is the matrix of inner products between per-example parameter gradients, and its condition number measures how evenly training can reduce error across examples. Both use gradients with respect to the trainable parameters, which here are the phases, so both can be applied directly. Both are the natural competitors because each claims to capture trainability without training. We will compute them on the same minibatches, with the same initialization distribution and the same number of batches as $C$. The one exception is ReLU linear-region counting, which cannot be defined for these networks.
The key ablation is a task-agnostic reachability score: the normalized effective rank of $J$, which counts the independent directions the knobs can move with no task data involved. It isolates the ingredient we claim matters, which is alignment with the task gradient. A second ablation compares noise-free $C$ with noise-averaged $C$ when predicting accuracy under the noisy device model. We count the proposal as successful if $C$ beats ZiCo, the kernel condition number, and effective rank on Kendall's $\tau$ within iso-MZI families, with non-overlapping bootstrap intervals, on at least two of the three datasets.
Why not just…?
Why not just train every candidate mesh in simulation? At $N = 8$ this is cheap and correct. We use it as ground truth and do not claim the proxy adds value there. The case for a proxy starts at $N \geq 64$ with multi-layer networks. Here, candidate counts reach thousands, noise-robust evaluation multiplies training cost by the number of fabrication draws, and on-chip evaluation runs into millisecond-scale thermal tuning (Reconfigurable versatile integrated photonic computing chip). We will report the wall-clock crossover point beyond which a proxy followed by training a short list outperforms training everything. If that crossover never arrives at realistic sizes, the proxy is not worth using, and we will say so.
Why not just count reachable directions (rank of $J$) or check that every input reaches every output? This is the most serious threat, because such a check is simpler and requires no data. If effective rank matches $C$ on Kendall's $\tau$, the task gradient adds nothing and the contribution reduces to a mesh expressivity measure. That would still be useful, but it would be a different paper. The ablation settles this directly. We expect the two to diverge most on CIFAR-10, where the useful weight directions are more structured than on MNIST.
Why not just use ZiCo or the neural tangent kernel condition number as they are? These scores were validated on ReLU image networks, and NAS-Bench-360 shows that proxy quality collapses outside that setting. Their theory assumes independent weights. Phase parameters are coupled through products of MZI transfer matrices, so gradient statistics per phase need not reflect what the realized matrix can do. This is a hypothesis, and the head-to-head comparison tests it.
Why not just fabricate a universal mesh and prune it? At scale the universal starting point does not fit. Scalable and compact photonic neural chip with low learning-capability-loss estimates over a million MZIs and 384 cm² for $N = 1024$, so the choice of layout has to happen before fabrication.
Proposed Experiments
The anchor experiment reproduces the regime of Scalable and compact photonic neural chip with low learning-capability-loss. That work reports that an 8×8 FFTUnitary mesh with 24 MZIs reaches about 97% on MNIST versus 64 MZIs for the universal design, and that its fabricated 4×4 chip reaches 99.40% on MNIST. We first check that our simulator recovers FFTUnitary's advantage over random 24-MZI layouts. We then check whether $C$ ranks FFTUnitary near the top without training.
The decisive checks come in three parts. First, within iso-MZI families at $N = 64$ and $256$, we measure Kendall's $\tau$ and the true rank of the proxy's top-1% candidates, using the BestRanking metric of Evaluating Efficient Performance Estimators of Neural Architectures, for $C$ against every baseline. Second, under the noisy device model, we test whether noise-averaged $C$ predicts post-noise accuracy better than noise-free $C$. A positive result would show that noise-awareness belongs inside the score rather than being applied afterward. Third, we test transfer: rankings computed on MNIST minibatches are checked against accuracy on CIFAR-10 and on one NAS-Bench-360 task such as ECG classification. The aim is to see whether, as NAS-Bench-360 found for electronic networks, the best mesh layout depends on the task. If it does, a task-aware score is warranted.
A negative result would also be informative. If effective rank matches $C$, or ZiCo matches both, the practical recommendation becomes the simpler score.
References
- ZiCo: Zero-shot NAS via inverse Coefficient of Variation on Gradients
- Evaluating Efficient Performance Estimators of Neural Architectures
- Scalable and compact photonic neural chip with low learning-capability-loss
- Apollo: Automated Routing-Informed Placement for Large-Scale Photonic Integrated Circuits
- LiDAR: Automated Curvy Waveguide Detailed Routing for Large-Scale Photonic Integrated Circuits
- HaeNAS: Hardware-Aware Efficient Neural Architecture Search via Zero-Cost Proxy
- Architecture-Level Modeling of Photonic Deep Neural Network Accelerators
- Understanding and Accelerating Neural Architecture Search With Training-Free and Theory-Grounded Metrics
- Reconfigurable versatile integrated photonic computing chip
- NAS-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks