Inverse-Designed Nanophotonic Neural Network Accelerators for Ultra-Compact Optical Computing

Source type: Paper (peer-reviewed, Nature Communications) Citation: Nature Communications (2026). DOI: 10.1038/s41467-026-68648-1 Lead author / institution: Joel Sved et al. — University of Sydney, School of Electrical and Computer Engineering

fetch_status: summary-derived. The Nature Communications article redirects to an authentication gateway (IDP login wall); ResearchGate mirror returned 403. Figures below are grounded in the paper's abstract as surfaced via WebSearch plus the institution's companion write-up (techxplore.com/news/2026-03-ultra-compact-photonic-ai-chip.html). Canonical URL preserved.

Summary

The now-published primary paper for the University of Sydney inverse-designed nanophotonic neural-network accelerator. Rather than building a network from discrete photonic components (MZIs, microrings), the design treats each subwavelength voxel of a nanostructured silicon region as a trainable degree of freedom and uses inverse design to sculpt the structure so that light propagating through it directly performs the neural-network computation. The result is an extreme computational density — approximately 400 million parameters per mm² — in a footprint on the order of tens of micrometers (roughly a human-hair width), computing at the speed of light / picosecond timescale with effectively no resistive heat.

This entry is an authority upgrade: the KB already tracked this work via a news article (nanophotonic-neural-network-sydney, techxplore). This is the peer-reviewed Nature Communications source for the same chip, adding the quantitative compute-density figure and explicit on-chip benchmark accuracies that the news writeup lacked.

Key Claims & Figures (quoted)

• ~400 million parameters per mm² computational density — each subwavelength voxel a trainable degree of freedom. (Nature Communications abstract)
• On-chip classification accuracy: ~89% (MNIST), ~90% (MedNIST). (Nature Communications abstract)
• 90–99% classification accuracy on 10,000+ biomedical images (breast, chest, abdomen MRI scans). (techxplore companion)
• Picosecond timescale operation ("trillionths of a second"), at the speed of light, with no electrical-resistance heat. (techxplore companion)
• Footprint ≈ tens of micrometers, comparable to the width of a human hair. (techxplore companion)
• Computation embedded directly into nanoscale photonic structure via inverse design — no reconfiguration needed. (both)

Relevance to KB

Strengthens Photonic Neural Networks, Photonic Accelerators, and the University of Sydney entity with the peer-reviewed figures. The ~400M params/mm² density is a new headline metric. Note: inverse-designed passive nanostructures are fixed-function (trained once into the geometry) — a contrast to the programmable/trainable-on-chip direction of the Ashtiani et al. on-chip-backprop result, and worth framing as the "ultra-dense fixed accelerator vs. reconfigurable trainable processor" axis.

Notes

Same underlying work as nanophotonic-neural-network-sydney (news-tier); kept as a separate, higher-authority source per dedupe-by-authority. Evidence label strong (peer paper) when compiled.