JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

On-chip diffractive optical neural network based on binary metasurfaces

Optics Express +

|

July 30, 2025

|

July 30, 2025

View abstract on PubMed

Summary

This summary is machine-generated.

This study introduces an on-chip diffractive optical neural network using binary metasurfaces. This compact device achieves 90% accuracy on the Iris dataset in simulations, demonstrating potential for efficient optical computing.

Area Of Science

  • Optoelectronics
  • Nanophotonics
  • Artificial Intelligence

Background

  • Diffractive optical neural networks (ONNs) offer a promising avenue for compact and efficient optical computing.
  • Metasurfaces provide a versatile platform for manipulating light at the nanoscale, enabling novel optical functionalities.
  • Integrating ONNs with metasurfaces can lead to miniaturized devices for complex computational tasks.

Purpose Of The Study

  • To propose and investigate an on-chip diffractive optical neural network (d-ONN) implemented using binary metasurfaces.
  • To optimize binary metasurfaces for compact size and high performance using advanced computational methods.
  • To demonstrate the d-ONN's capability in executing a real-world classification task.

Main Methods

  • Utilized a genetic algorithm (GA) for optimizing the binary metasurface design.
  • Employed the finite-difference time-domain (FDTD) method for accurate optical simulations.
  • Designed and fabricated a single-layer d-ONN on a silicon-on-insulator (SOI) platform.

Main Results

  • Achieved a compact single-layer d-ONN with an area of 16.5 µm × 23 µm.
  • Attained a simulation validation accuracy of 90.0% for the Iris dataset classification task.
  • Fabricated device demonstrated a measurement validation accuracy of 78.3%.

Conclusions

  • The proposed on-chip diffractive optical neural network based on binary metasurfaces is effective for classification tasks.
  • The optimization methods (GA and FDTD) enable the design of compact and high-performance d-ONNs.
  • While simulation results are promising, further improvements are needed to bridge the gap between simulated and measured performance.