ANNs for design of silicon interlayer coupled structures

Abstract

In silicon optical device design, traditional methods are often time-consuming and lack of efficient convergence when directly employing artificial neural networks for inverse design. To address this challenge, we propose an efficient inverse design approach rooted in migration learning for silicon interlayer coupled structures. This method employs Figure of Merit (FOM) screening to preprocess data, followed by training the first Backward Propagation (BP) neural network on the silicon interlayer coupled structure dataset. Subsequently, the learned hyperparameters from the first BP neural network are transferred to the second BP neural network, enhancing the neural network’s accuracy significantly. The results in the mean absolute percentage error (MAPE) for the single-layer and the two-layer coupled-structure neural network can be reduced to 1.2845 % and 7.3409 % in respectively. These findings demonstrate the practical utility of the method in the inverse design of silicon interlayer coupled structures and provide guidance for the design of silicon optical devices.