Parameter extraction method for assisting inverse design of DFB lasers based on rate equations and generative adversarial network

Abstract

Effective parameter extraction without redundant experiments is a hot topic in the inverse design of lasers. A tandem model based on a generative adversarial network (GAN) has been proposed to accurately extract theoretical rate equation parameters of distributed feedback (DFB) lasers, avoiding the erroneous tendency caused by measurement accuracy and optimization algorithms. Cascading the deep neural network (DNN) with the GAN, the trained model can effectively extract parameters and construct forward model by solving rate equations numerically, which is more generalized and robust than the traditional network. The solution can breakthrough the potential boundaries and local optima encountered from the structure consisting of forward convolutional neural network (CNN) and particle swarm optimization (PSO) algorithms. Results demonstrate the presented model obviously addresses the demand of the substantial time and computational resources of traditional numerical solving methods, exhibits powerful convergence and its relative average error (RAE) is significantly improved with comparison to PSO–CNN under similar conditions. The extracted rate equation parameters can offer a more precise reference theoretically for the previous initialization of other extraction approaches, even the suggested neural network serves as a guidance for other optoelectronic devices design inversed.