On-chip reconfigurable diffractive optical neural network based on Sb₂S₃.

Area of Science:

• Photonics and Machine Learning
• Optical Computing and Integrated Photonics

Background:

• Implementing machine learning functions on-chip requires compact, efficient solutions.
• Traditional hardware modifications or re-fabrication limit the reconfigurability of optical neural networks.

Purpose of the Study:

• To propose a Sb2S3-based reconfigurable diffractive optical neural network (RDONN) for on-chip integration.
• To demonstrate a passive, all-optical solution for reconfigurable machine learning functions.

Main Methods:

• Utilized a two-dimensional electromagnetic propagation model to construct the RDONN architecture.
• Employed multilayer metalines made from Sb2S3, a low-loss phase change material, for reconfigurable weights.
• Tested the RDONN on the Iris dataset using both intensity and phase modulation inputs.

Main Results:

• Achieved high classification accuracies of 95.0% with intensity modulation and 98.3% with phase modulation.
• Demonstrated the feasibility of reconfigurable weight manipulation without hardware modification.
• Confirmed the integration capability with standard silicon-on-insulator systems.

Conclusions:

• The proposed Sb2S3-based RDONN offers a viable, compact, and passive all-optical solution for on-chip machine learning.
• This reconfigurable optical chip design facilitates the development of future all-optical in situ learning systems.
• The technology holds significant promise for the design and fabrication of real-world optical computing chips.