Optical neural networks do their matrix multiplications almost for free. Getting them to compute anything a linear map cannot is the hard part, because light barely interacts with light. Most systems therefore send the signal back into electronics to apply the activation function, which is the nonlinear "squash" between layers. That round trip is costly: architecture-level modelling shows data converters and memory dominate photonic accelerator energy. Epsilon-near-zero (ENZ) materials offer a way out. These are thin films whose permittivity, the quantity that sets how strongly a material responds to an electric field, is tuned to nearly zero at some wavelength. Near that point, light changes the film's refractive index by order unity on femtosecond timescales. The catch is that the effect only appears at very high intensity, near the damage limit of the film. Every nonlinear layer in an all-optical network therefore consumes a real optical energy budget, and the light reaching each layer depends on everything upstream of it.
This proposal treats that budget as the central design variable. We will build a hardware-aware architecture search whose hardware model is an optical power-flow simulation. The search decides how many ENZ nonlinear layers to place, where in the optical path to place them, which film to use, and how much light to deliver to each. It enforces the film's damage threshold as a hard constraint rather than a soft penalty. The scientific question is whether a few well-placed optical nonlinearities recover most of the accuracy that linear optics loses, at a pump energy below what optoelectronic activations cost. Equally, we want to know whether the obvious design of "a nonlinearity after every layer" is infeasible once loss and damage are modelled honestly.
Core Claim
Existing optical architecture search treats nonlinearity as somebody else's problem. ADEPT differentiably searches the layout of linear photonic tensor cores (meshes of couplers and phase shifters) under foundry layout constraints. O-HAS searches micro-architecture and dataflow mappings for linear optical units, and assumes activations happen electronically. CiMLoop for photonics models the energy of such accelerators but does not search, and routes every activation through photodetectors and converters. These works supply our method and our cost-modelling scaffolding; none of them contains a searchable optical nonlinearity.
Our claim is that once the nonlinearity is an intensity-driven ENZ film, two things change in kind. First, the activation's shape is no longer fixed. Whether a film acts as a saturable absorber, a phase modulator, or effectively a linear element depends on the intensity that arrives at it, and that intensity is set by upstream fan-out, insertion loss, and placement. Second, hardware cost stops being additive across blocks. The per-block lookup tables used by MoGA and validated in HW-NAS-Bench assume each block's cost is independent of its neighbours, and that assumption fails when one layer's output power determines the next layer's function and damage risk.
We hypothesise that the accuracy-versus-pump-energy optimum uses sparse, non-uniform nonlinearity placement concentrated where light is naturally intense. A search that models power flow will find this optimum. Uniform placement and additive cost models will miss it.
Physical Model
The device model is built from the measurements in the file. Caspani et al. report three relevant behaviours in aluminium-doped zinc oxide films near 1300 nm:
- a six-fold enhancement of the Kerr index $n_2$, which measures how much the refractive index changes per unit intensity;
- index changes of order unity;
- a switch from positive nonlinear absorption to saturable absorption exactly at the ENZ wavelength.
Kaipurath et al. report four behaviours in silver/silica stacks:
- a permittivity change of about 0.05, linear in pump power on a 100 fs timescale;
- crossing from metallic to dielectric behaviour;
- a response dominated by absorption on the metallic side and by refraction on the dielectric side;
- a damage threshold near 30 GW/cm².
Ciattoni et al. explain why the response accumulates coherently across the film without phase mismatch.
We turn these into a differentiable transfer function. It maps local input intensity $I$ to a complex transmission $t(I;\theta)$, where $\theta$ collects the film type, thickness, operating detuning from the ENZ wavelength, and $n_2$ and loss parameters. One honest caveat must be stated up front. These are pump–probe measurements, while a network activation is a self-action effect in which the signal modulates itself. The self-action curve is therefore a model inference, and every result below is reported across a sweep of $n_2$ and loss spanning the reported values. Caspani et al. themselves note that the constant, non-dispersive $\chi^{(3)}$ assumption is an approximation, which is another reason for the sweep.
The power budget enters through one constraint per nonlinear site $k$:
$$\max_x I_k(x;\alpha,\theta) \le I_{\mathrm{dmg}}, \qquad E_{\mathrm{pump}}(\alpha,\theta) = \sum_k A_k \tau \, \bar I_k(\alpha,\theta) \le E_{\mathrm{budget}},$$
The symbols are as follows:
- $x$ ranges over training inputs;
- $\alpha$ are the architecture choices;
- $I_k$ is the peak intensity at site $k$, obtained by propagating the input through the actual upstream optics, including loss;
- $I_{\mathrm{dmg}}$ is the damage intensity;
- $A_k$ is the illuminated area at site $k$;
- $\tau$ is the pulse duration;
- $\bar I_k$ is the intensity the film must be driven at to be nonlinear.
The coupling across layers lives inside $I_k$. For scale, 30 GW/cm² over a 0.1 µm² mode is about 30 W peak, or about 3 pJ in a 100 fs pulse. Each optical activation therefore plausibly costs picojoules, and that is the quantity the search trades against accuracy.
Research Program
The primary setting is free-space optics, because the ENZ data were measured on free-space thin films and this keeps the device model closest to experiment. The base system is the spectral-domain linear optical network of the knowledge-distillation paper. In that system, a 4f lens setup performs convolutions as element-wise multiplications in the Fourier plane, and pooling is replaced by frequency cropping. We stack several such stages. At each stage the search chooses among:
- no film;
- a film in the Fourier plane, where low-frequency energy is concentrated and intensity is high;
- a film in the image plane, where intensity is spread out.
It also chooses the film type and detuning, and a per-stage optical attenuation or gain factor that sets the operating point.
A secondary setting is integrated photonics. Here ADEPT-style linear meshes are interleaved with optional ENZ sections of chosen length. A photodetector–amplifier–modulator activation is available as an alternative choice at every site, with its energy taken from the CiMLoop photonic model. Including the electronic option is deliberate: the search itself tells us where all-optical nonlinearity is worth it and where conversion back to electronics wins.
The search runs over a one-shot supernet with relaxed discrete choices, following the general approach of HAT and MoGA. The difference is that the cost signal comes from forward power-flow simulation rather than a lookup table. Constraints are enforced by projecting architecture gradients toward feasibility, following ConNAS. We extend it in two ways: the "direction to feasibility" is computed from gradients of the simulated peak intensity, and the projection also applies to the continuous operating-point variables. This hard handling matters because the damage threshold is a cliff. A soft-penalty search can spend intermediate epochs in regimes that would destroy the film, and it ends with designs that narrowly miss the limit.
Three analyses come out of the program. The first is the accuracy-versus-pump-energy Pareto front. The second is the distribution of searched placements: how many nonlinear sites, and in which plane. The third is the error of an additive per-block cost model against the coupled simulation over the searched population.
Baselines
The naive reference point is a purely linear optical network trained normally. The knowledge-distillation paper reports 70.12% on HIGH-10 for this linear model. The strongest no-nonlinearity competitor is the same linear network distilled from a nonlinear electronic teacher, reported at 81.40% on HIGH-10 against 93.95% for the AlexNet teacher. We reimplement both on the same optical stages and data, with the same training epochs. The proposal beats this baseline if it exceeds 81.40% on HIGH-10 at a pump energy within the budget. Closing a meaningful share of the 12.55-point gap would be the headline result. We also combine distillation with our searched networks, so that any gain is attributable to nonlinearity and not to the training signal.
The strongest competitor that does have nonlinearity is a hybrid optoelectronic network: the same linear optics with photodetector–amplifier–modulator activations at every layer. Its energy comes from the CiMLoop photonic model, which reports 0.4% average energy error against Albireo and identifies converters and memory as the dominant costs. Device behaviour is taken from the graphene/silicon microring activation, which reports 7% and 8% gains over an arctan activation on MNIST and CIFAR-10. No absolute accuracy for that device was found on Lacuna. This baseline gets the same search budget, so it is a searched hybrid rather than a hand-placed one. Beating it means matching its accuracy at lower total energy per inference, or exceeding its accuracy at equal energy.
The ablation that isolates our key ingredient is uniform ENZ placement. It puts a film after every linear stage, uses a single global laser power tuned by grid search to the largest value that respects damage, and trains with the same physics-aware model. A second ablation keeps the search but replaces the power-flow cost with an additive per-block table, as in MoGA and HW-NAS-Bench. A third replaces gradient projection with a penalty schedule in the style of DUCCIO. All variants get the same number of supernet training steps and the same hyperparameter-tuning trials. The comparison also reports how many of each method's final designs violate the damage threshold.
Why not just…?
Why not just put an ENZ film after every layer and tune the laser? This is the real threat. If uniform placement matches the searched design, the contribution shrinks to physics-aware training of an optical activation, and we will report that. Our expectation runs the other way. Insertion loss compounds through stages, so the laser power that brings deep stages above their nonlinear threshold would push early stages past damage. The search can instead use attenuation, gain, and placement to equalise intensity. The uniform-placement ablation, swept across the loss and $n_2$ range, settles this directly. We will also test a cheap heuristic that places films only where simulated intensity is highest, since a heuristic that wins would make the search unnecessary.
Why not just use linear optics with distillation? The file's own numbers show the ceiling: 81.40% versus 93.95% on HIGH-10. The paper's suggested fix of oversampling the input works in free space but grows mesh size on chip. Distillation remains a baseline and an add-on to our method. What we cannot assume is that ENZ nonlinearity at realistic pump energy beats it. That is precisely what the Pareto comparison measures.
Why not just use electronic activations and an existing optical search like O-HAS? Its cost predictor is accurate to roughly 8.5% for energy and 12.5% for latency, but it has no optical nonlinearity in its design space. Converter energy is a known dominant cost. Still, picojoule-scale ENZ pumping could cost as much as conversion. Rather than argue this, we put both options in the same search space and report which the search picks at each site.
Why not just a soft penalty on power? ConNAS reports that penalty-based FBNet and TAS frequently fail strict constraints, while ConNAS satisfied them in 100% of runs and came within 0.14–0.65% of the best feasible NATS-Bench architecture. Our penalty ablation tests whether this advantage carries over to a physical damage threshold.
Proposed Experiments
On HIGH-10 in the free-space setting, we trace accuracy against pump energy per inference for every method above. The anchors are the reported linear (70.12%), distilled (81.40%), and teacher (93.95%) figures.
On MNIST, FashionMNIST, SVHN, and CIFAR-10, the benchmarks ADEPT used, we run the integrated setting, where ADEPT reports no absolute accuracies on Lacuna. Here we compare searched ENZ placement, searched hybrid placement, and linear meshes at matched footprint, measuring accuracy, simulated energy, and damage-violation rate.
Three checks are decisive for the core claim:
- Searched non-uniform placement must beat uniform placement at matched pump energy, and the advantage must survive the device-parameter sweep.
- An additive cost model must show substantial rank-order error against the coupled simulation over the searched population. Kendall correlation is the natural metric, following HW-NAS-Bench's FLOPs-versus-latency analysis.
- Some region of the Pareto front must exist where all-optical ENZ designs dominate both distillation and the searched hybrid.
If none of these hold, we will have shown, with a calibrated model, that intensity-driven ENZ activations are not yet competitive. That negative result would still steer both communities.
References
- ADEPT: automatic differentiable DEsign of photonic tensor cores
- O-HAS: Optical Hardware Accelerator Search for Boosting Both Acceleration Performance and Development Speed
- Architecture-Level Modeling of Photonic Deep Neural Network Accelerators
- MoGA: Searching Beyond Mobilenetv3
- HW-NAS-Bench: Hardware-Aware Neural Architecture Search Benchmark
- Enhanced Nonlinear Refractive Index in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ε</mml:mi></mml:math>-Near-Zero Materials
- Optically induced metal-to-dielectric transition in Epsilon-Near-Zero metamaterials
- Enhanced nonlinear effects in pulse propagation through epsilon‐near‐zero media
- Knowledge Distillation Circumvents Nonlinearity for Optical Convolutional Neural Networks
- HAT: Hardware-Aware Transformers for Efficient Natural Language Processing
- Constraint-guided Hardware-aware NAS through Gradient Modification
- Graphene/silicon heterojunction for reconfigurable phase-relevant activation function in coherent optical neural networks
- Enhancing Neural Architecture Search With Multiple Hardware Constraints for Deep Learning Model Deployment on Tiny IoT Devices