Forward Prediction and Inverse Design of Nanophotonic Devices Based on Capsule Network

Abstract

Deep neural networks have been successfully applied to forward predicting optical response and inverse designing topological structure of nanophotonic devices. However, the existing deep learning based methods need sufficient simulated data to train the model effectively. For those devices with complex structures that containing many design variables, obtaining enough training data through numerical simulations will become extremely time-consuming. In order to reduce the requirement of large amounts of training data, we present a new deep learning approach based on the Capsule Network in this paper. By employing the proposed model, we have designed and verified a series of silicon-based wavelength demultiplexer with more than one thousand design variables. The numerical simulations validate that the trained model can both effectively predict the optical response with a fixed topological structure, and inverse design the approximate topological structure for a needed given optical response. Comparison with the classical convolutional neural networks show that our model can obtain nearly the same performance when using only 60% of the training data.