Architecture class within ONNs.
Physics / mechanism
Origin: Lin et al. (Ozcan group, UCLA) demonstrated the first all-optical D2NN classifier in Science 2018, with cascaded 3D-printed phase plates performing MNIST-class classification at the speed of light through the stack. The architecture has since spread to USTC, Stanford, CMU, and many others.
Key parameters:
- Latency sub-ns per layer (light propagation through optical path)
- Energy dominated by source laser and detector array (not the diffraction itself, which is “free”)
- Precision historically the weak point — analog at every layer, error compounds with depth
- Reconfigurability is the major variable — fixed phase plates are fastest but task-locked; SLM-driven systems are slower (limited by SLM refresh) but reprogrammable per workload
Distinguished from sibling architectures
| Architecture | Compute primitive | Where it lives | Key trade-off |
|---|---|---|---|
| D2NN (this page) | Diffraction through phase masks | Free space | Massively parallel, but precision compounds; reconfigurability tied to SLM speed |
| MZI mesh ONN (Lightmatter, Lightelligence) | Guided-wave Mach-Zehnder phase shifters | Silicon photonic IC | Foundry-fabable; but O(N²) phase shifters limits scale |
| Photonic tensor cores | Wavelength-multiplexed multiply-accumulate | Silicon photonic IC | High throughput; precision still 4-6 bit |
| Photonic reservoir | Fixed nonlinear dynamical system | Free-space or fibre | Good for time-series / ODEs; limited generality, no commercial breakout |
| Analog optical (Salience) | PCM crossbar with optical readout | Hybrid PCM + photonic | Foundry-adjacent; precision-bound |
Why this matters now (2026)
Companies using
Connected ideas
Sources
Frontier (open questions)
See frontmatter frontier: block.
Reading list
- To be added — Ozcan UCLA originating Science 2018 paper + key follow-ups. Run
/kb-research diffractive-deep-neural-networksto populatesources/papers/.