Optical neural networks perform the linear algebra of inference by letting light interfere, diffract or scatter through engineered media, and after two decades of small-scale demonstrations they now work well enough for MNIST-class and sensing tasks, while depth, nonlinearity and reconfigurability remain unresolved.
Summary
An optical neural network (ONN) encodes a neural network’s weights into an optical structure and its activations into light, so that matrix-vector multiplication happens as a side effect of propagation rather than as a sequence of clocked digital multiply-accumulates. Three hardware families dominate the literature. Programmable photonic integrated circuits build the matrix from interferometer meshes or directional-coupler stages on silicon; free-space diffractive networks stack phase masks so that propagation between planes performs the fan-out; and volumetric or scattering media compress the interconnect into a monolithic block of engineered disorder. A recent example of the last family packs a classifier into ~3.8×10⁻⁴ mm³ at a claimed record neuron density of 1.0×10⁹ per mm³, reaching 94.8% blind-test MNIST accuracy and 1 μm effective resolution as an optical compressed imager.
The linear part of an ONN has never been the problem. Three parameters decide whether the technology matters. Component scaling: conventional Mach-Zehnder meshes need O(N²) active phase shifters, which caps matrix size; a multiport-directional-coupler architecture has been demonstrated that covers the N-dimensional unitary group with 3N phase shifters and realises an N×N matrix-vector product with 7N, breaking that scaling. Nonlinearity: optical materials have weak intrinsic nonlinearity, so most working systems fall back on electronics between layers, paying conversion energy and latency. Routes around this include multiple scattering through repeated data-embedding planes, which synthesises linear and nonlinear transforms concurrently at milliwatt CW power using only linear components; foundry-made optical-electrical-optical converters that pair a germanium photodetector with a micro-ring modulator to give a reconfigurable nonlinear transfer function with on-chip RF gain above unity; and saturable quantum emitters embedded in inverse-designed nanophotonics, proposed to operate at nW/μm², seven orders of magnitude below the nonlinearity threshold of conventional materials.
Depth and noise: analogue optics accumulates amplified spontaneous emission and insertion loss with every layer. A semiconductor-optical-amplifier neuron using cross-gain modulation was modelled to show genuine noise compression: because M weighted WDM inputs are compressed onto one amplified output wavelength, when the input OSNR is below about 19 dB the output OSNR is higher than the input, so noise does not simply accumulate to death across cascaded layers 2022 Shi Soa Aonn Noise Depth Scaling. Where physical depth cannot be increased, architecture can substitute for it: a sequential reasoning framework decomposes a task into coarse-to-fine steps executed by a single diffractive layer with dynamic operator selection, reaching 94% accuracy on 100-class optical fibre speckle classification at a reported system energy efficiency of 23.3 TOPS/W.
Evidence base
- 2026-05-14: Fabricated volumetric optical scattering network, volume ~3.8×10⁻⁴ mm³, neuron density 1.0×10⁹/mm³, 94.8% blind-test MNIST accuracy, 1 μm effective imaging resolution, average FSIM 0.93
- 2026-06-02: A single-layer diffractive ONN under a sequential coarse-to-fine reasoning framework classified 100 classes of optical fibre speckle at 94% accuracy with 23.3 TOPS/W system energy efficiency
- 2026-03-07/2026-06-07: Multiport-directional-coupler optical unitary converter covers the N-dimensional unitary group with 3N phase shifters and realises an N×N matrix-vector product with 7N, versus O(N²) for MZI meshes
- 2022-09-02: SOA cross-gain-modulation neuron model with 7 WDM channels at 10 Gbit/s OOK and −17.5 dBm/channel shows output OSNR exceeding input OSNR when input OSNR is below ~19 dB, so ASE does not accumulate fatally with depth 2022 Shi Soa Aonn Noise Depth Scaling
- 2026-01-08/2026-06-07: First reported monolithically integrated, foundry-fabricated silicon-photonic load-resistor OEO converter with reconfigurable nonlinear transfer, on-chip RF gain above unity at ~10 mW (load-resistor) and ~1 mW (current-injection) bias, and open eyes to 4 Gb/s
- 2026-08-13: Diffractive ONN plus photon counting shown to saturate the Nagaoka-Hayashi Cramer-Rao bound for band-limited spatial-frequency estimation and to outperform direct imaging in reconstruction
Open questions
- Does the SOA noise-compression result (output OSNR exceeding input OSNR below ~19 dB input OSNR) survive in a fabricated multi-layer chip, or only in the OSNR-recursion emulation 2022 Shi Soa Aonn Noise Depth Scaling?
- Can the 7N phase-shifter matrix-vector architecture be scaled to matrix sizes relevant to real workloads while holding insertion loss, phase-shifter drive power and calibration burden constant?
- Is the volumetric scattering medium’s fixed, nanolithographically written weight set a fatal limitation, or can inverse-designed volumes be made reprogrammable, for example via nonvolatile ferroionic phase control?
- Do the quantum-emitter and atom-cavity nonlinearities operate at the claimed nW/μm² in a fabricated device with realistic detuning and photon loss, and at what repetition rate?
Assessment drafted 2026-08-31 from up to 14 KB sources using the technology-scorecard framework; scores are a draft read pending review.