EUVIMEDEUVIMED entdecken
Crossref

Matrix–vector multiplication for photonic tensor core in optical neural network

Aijun Zhu, Tailin Xu, Wenrui Wei, Weigang Hou

Physica Scripta · 10.07.2026 · 101 (27) · 275507

DOI: 10.1088/1402-4896/ae8345 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 09.07.2026. Verfasst von Aijun Zhu und 3 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Guangxi Key Laboratory of Automatic Detecting Technology and Instruments, Guangxi Natural Science Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Matrix-vector multiplication (MVM) is the basic operation for photonic tensor core, which lies at the core of neural network inference and training, forming the dominant computational primitive underlying every layer of modern deep architectures. Although optical neural networks promise massive acceleration of these operations through parallel electromagnetic propagation, most existing implementations suffer from limited reconfigurability, off-chip footprint, or complex integration requirements. In this work, we introduce an on-chip computing element specifically tailored for neural-network MVM, constructed from a cascaded arrangement of metasurfaces. The metasurfaces incorporate vanadium dioxide and photoactivated GaAs, whose distinct switching mechanisms enable each layer to encode matrix weights and input vector elements through external electrical, thermal, and optical stimuli. As terahertz waves sequentially traverse the cascade, the system physically accumulates weighted contributions, yielding the equivalent multiplication between a 2 × 2 matrix and a 2 × 1 input vector—executed at wave speed. Finite integration time-domain simulations validate all logical combinations of matrix elements and inputs and confirm the correct arithmetic mapping for two independent computing channels. Moreover, it supports general weight programmability and maintains stable responses with respect to incident and polarization angles. By embedding the most fundamental neural-network operation directly into a planar metasurface stack, this work provides a compact pathway toward integrated, low-power optical accelerators and lays the groundwork for future metasurface-based deep learning hardware.

Förderung

  • National Natural Science Foundation of China
  • Guangxi Key Laboratory of Automatic Detecting Technology and Instruments
  • Guangxi Natural Science Foundation
  • Guangxi Key Laboratory of Brain-inspired Computing and Intelligent Chips

Zitieren

APA

Zhu, A., Xu, T., Wei, W., & Hou, W. (2026). Matrix–vector multiplication for photonic tensor core in optical neural network. Physica Scripta, 101(27), 275507. https://doi.org/10.1088/1402-4896/ae8345

Vancouver

Zhu A, Xu T, Wei W, Hou W. Matrix–vector multiplication for photonic tensor core in optical neural network. Physica Scripta. 2026;101(27):275507. doi:10.1088/1402-4896/ae8345

BibTeX
@article{Zhu2026,
  author = {Zhu, Aijun and Xu, Tailin and Wei, Wenrui and Hou, Weigang},
  title = {Matrix–vector multiplication for photonic tensor core in optical neural network},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {27},
  pages = {275507},
  doi = {10.1088/1402-4896/ae8345}
}
Kostenlos starten →