Deep Neural Networks: Efficient Method To Design Photonic Device

A research team led by Prof. ZHANG Wenfu from the Xi'an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Sciences (CAS) recently reported a novel genetic-algorithm-based deep neural network (GDNN) to efficiently design photonic devices. Their up-to-date result was published on Photonics Research.

Novel photonic devices, especially photonic integrated circuits (PICs), are widely applied in transformative technologies including high-speed optical communication and computing, ultrasensitive biochemical detection, efficient solar energy harvesting and super-resolution imaging.

These photonic devices are normally designed by numerical simulations such as finite element method (FEM) and finite difference time-domain (FDTD). However, the algorithms are computation intensive and time consuming. A highly efficient design method is thus important to accelerate the development of next-generation optical applications.

According to the researchers, the rapid development of deep neutral networks (DNNs) has greatly drawn their attention, and they proposed a GDNN method, which requires an order of magnitude less simulation data for training. With this approach, they designed several silicon photonics devices.

The results indicate that the algorithm not only has high efficiency, but also has strong flexibility and the ability to deal with various design constraints.

The algorithm can be widely implemented in designing many complex micro- and nanophotonic structures that was not easily realized before.

This work was supported by the Strategic Priority Research Program of CAS, the National Natural Science Foundation of China, the Youth Innovation Promotion Association of CAS, the West Light Foundation of the Chinese Academy of Sciences, among others.