Enhanced Nonlinear Optical Coefficient in Silicon for Efficient All-Optical Neural Networks

Researchers Yoshihiro Honda, Yuya Shoji, and Tomohiro Amemiya have demonstrated a significant enhancement in the nonlinear optical coefficient of silicon by incorporating a metamaterial structure into an existing silicon waveguide. This enhancement, quantified as a two-photon absorption coefficient of 424 cm/GW, is approximately 1,200 times higher than that of standard silicon. The study achieved a high inference accuracy of 98.36% in handwritten character recognition tasks using this metamaterial waveguide-based nonlinear optical activation function, comparable to the widely-used ReLU function. These findings suggest that this approach could lead to more power-efficient and compact all-optical neural networks, potentially advancing the development of faster and more efficient computing systems.

The paper, titled "Optical activation function using a metamaterial waveguide for an all-optical neural network," can be accessed on arXiv with the identifier arXiv:2408.13129.