Inverse-designed integrated all-optical nonlinear activators for optical computing.

Area of Science:

• Photonics and Optical Engineering
• Artificial Intelligence and Machine Learning
• Materials Science

Background:

• Electronic neural networks face arithmetic and energy limitations.
• Optical neural networks (ONNs) offer a potential solution but lack essential optical nonlinearity for widespread implementation.
• Existing ONN designs struggle to achieve efficient and compact nonlinear activation.

Purpose of the Study:

• To design and implement ultra-compact all-optical nonlinear activators for ONNs.
• To address the nonlinearity bottleneck hindering ONN adoption.
• To improve the performance and robustness of ONNs in machine learning tasks.

Main Methods:

• Inverse design combining the adjoint method with Kerr nonlinearity.
• Utilizing Kerr and thermo-optic (TO) effects for nonlinear response generation.
• Demonstrating transmission-as-computation and structure-as-function principles.

Main Results:

• Successfully inverse-designed three ultra-compact all-optical nonlinear activators.
• Achieved a minimum activation threshold of 2.34 mW.
• Significantly improved MNIST task accuracy from 88.15% to 93.25% by incorporating nonlinear activation.
• Demonstrated robustness against phase errors in the ONN topology.

Conclusions:

• The developed nonlinear activators effectively enhance ONN performance and expressive power.
• The inverse design approach provides a viable route for creating efficient optical computing components.
• This work paves the way for scalable, chip-level optical neural networks.