Neural network inspired backpropagation algorithm for nanophotonic optimization

This is the authors' abstract. We don't add key points for this paper.

In recent years, optimizing the design of photonic structures has become a key area of nanophotonic research.Traditional methods like shape optimization and topology optimization each have their advantages, but also limitations.Shape optimization is fast and easy to manufacture but struggles with achieving the best possible performance, especially for complex designs.Topology optimization, on the other hand, can create more innovative designs but often results in structures that are hard to fabricate.In this work, we propose, to our knowledge, a new method that combines the best features of both shape and topology optimization.We apply this method to design photonic devices like gratings, which are used in many optical applications.To make the process more efficient, we construct backpropagation in nanophotonic structures inspired by neural networks.This allows us to quickly calculate the gradients of the parameters, speeding up the optimization process.We design a highly efficient blazed grating and a polarization beam splitter (PBS) using our method.Better results are achieved than the previous, including finding the structure with high first-order performance over a broader spectrum and improving the performance of the PBS from 96% to 98.8%.By reducing optimization time from over 11,000 h to just 38 h, our method opens up new possibilities for faster and more efficient optical designs.