Related Experiment Video
Updated: Sep 26, 2025
04:47
Multimodal Optical Imaging Platform for Studying Cellular Metabolism
Published on: June 6, 2025
653
Integrated photonic metasystem for image classifications at telecommunication wavelength
Zi Wang1, Lorry Chang1, Feifan Wang1
1Department of Electrical and Computer Engineering, University of Delaware, Newark, DE, 19711, USA.
Nature Communications
|April 20, 2022
Summary
Miniaturized diffractive metasurfaces enable wavelength-selective image recognition for optical communication and autonomous vehicles. These silicon photonic chips perform high-throughput computations, handling input uncertainties without post-tuning.
Area of Science:
- Photonics
- Machine Learning
- Nanotechnology
Background:
- Miniaturized image classifiers are crucial for advancements in optical communication, autonomous vehicles, and healthcare.
- Metasurfaces offer unique optical properties through subwavelength structures for advanced light manipulation.
Purpose of the Study:
- To demonstrate wavelength-selective image recognition using diffractive metasurfaces on a silicon photonic platform.
- To develop high-throughput, compact machine learning hardware for photonic integrated circuits.
Main Methods:
- Engineered subwavelength structures in multi-layer metasurfaces for directional diffraction and dispersion control.
- Implemented near 10^3 nanoscale phase shifters within 0.135 mm^2 for vector-by-matrix multiplication.
- Developed pre-trained metasystems capable of handling fabrication and measurement uncertainties.
Main Results:
- Demonstrated a 15-pixel spatial pattern classifier with near 90% accuracy for femtosecond inputs.
- Showcased a multi-channel wavelength demultiplexer functionality.
- Presented a hyperspectral image classifier based on diffractive computing.
Conclusions:
- Diffractive metasystems offer a novel machine learning architecture for photonic integrated circuits.
- The developed metasystems exhibit dense integration of phase shifters and high spatial multiplexing capabilities.
- These photonic machine learning architectures are robust to input uncertainties, paving the way for practical applications.
