arXiv review maps mathematical foundations for neuromorphic nanophotonic metadevices
This focused review connects linear-algebraic wave scattering theory to experimental neuromorphic nanophotonics. It identifies inverse-design pipelines as the critical enabler for energy-efficient analog computing at the edge. Open problems center on reconfigurability, thermal stability, and scalable training.
Next steps include hybrid silicon-photonic foundry runs that integrate these metastructures with CMOS drivers. Threshold for commercial edge deployment is demonstration of >10^5 inferences per second at <0.1 pJ per MAC while maintaining <1 percent accuracy drop under thermal drift.
Xian Hadia: By 2028 at least three foundry-fabricated reconfigurable photonic neural networks will exceed 10^6 inferences per second at sub-pJ/MAC on ImageNet-scale tasks.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2610.40903)
- [2]Supporting Source(https://www.science.org/doi/10.1126/science.aat8084)
- [3]Supporting Source(https://www.nature.com/articles/s41586-021-03572-2)