University of Sydney

Type: Research Institution (Nanophotonics / Photonic Computing)

The University of Sydney's photonic computing group (lead author Joel Sved, School of Electrical and Computer Engineering) produced one of the most compelling 2026 demonstrations of photonic neural network viability: an inverse-designed nanophotonic neural network achieving 90-99% classification accuracy on 10,000+ biomedical images (breast, chest, and abdomen MRI scans) at picosecond timescales. The work is now peer-reviewed in Nature Communications (s41467-026-68648-1), which adds the headline quantitative result: a computational density of ~400 million trainable parameters per mm², with ~89% on-chip MNIST and ~90% MedNIST accuracy.

The key innovation is the inverse-design approach. Rather than manually designing photonic structures that approximate desired neural network operations, inverse design uses optimization algorithms to find the nanostructure geometry that directly implements the target function, treating each subwavelength voxel as a trainable degree of freedom. The resulting structures are counterintuitive by human standards but highly efficient — operating at the speed of light with no electrical resistance and therefore no heat generation during the computation itself. The trade-off: these accelerators are fixed-function once fabricated (the network is trained into the geometry), in contrast to reconfigurable, on-chip-trainable processors like Nokia Bell Labs' backprop chip.

Key Contributions

• Inverse-designed nanophotonic neural network: 90-99% accuracy, 10K+ biomedical images, picosecond timescale (Nanophotonic Neural Network Sydney)
• ~400 million trainable parameters/mm² computational density; ~89% MNIST, ~90% MedNIST on-chip (peer-reviewed Nature Communications) (Inverse-Designed Nanophotonic NN (Nat. Comm.))
• Subwavelength voxels as trainable degrees of freedom via inverse design (Inverse-Designed Nanophotonic NN (Nat. Comm.))
• Zero heat generation during computation — energy cost is at generation/detection, not inference (Nanophotonic Neural Network Sydney)
• Nanostructures at tens-of-micrometers scale — orders of magnitude smaller than electronic equivalents (Nanophotonic Neural Network Sydney)
• Published in Nature Communications 2026 (peer-reviewed; the prior KB entry was the news-tier companion) (Inverse-Designed Nanophotonic NN (Nat. Comm.))

Limitations of Demonstrated Work

• Fixed-function: the network is trained into the nanostructure geometry, not reconfigurable on-chip (contrast Nokia Bell Labs' on-chip-trainable processor)
• Limited to specific task types (classification)
• Scaling to larger networks is ongoing

Mentioned In

• Photonic Neural Networks — Demonstrating practical accuracy at nanophotonic scale
• Photonic Accelerators — Real-accuracy benchmark for inference hardware

Related Entities

• Nokia Bell Labs — Opposite pole of the density-vs-trainability axis (reconfigurable, on-chip-trained)
• UT Austin / ASU (SimPhony Team) — System-level benchmarking complementing hardware demonstrations

Changelog

• 2026-06-24 — Upgraded to the peer-reviewed Nature Communications source (s41467-026-68648-1); added the ~400M params/mm² density and explicit on-chip MNIST/MedNIST figures; reframed limitation from "inference only" to "fixed-function" (the relevant contrast is reconfigurability, not training-capability, now that Bell Labs has shown on-chip training elsewhere).