An Enhanced Space-Mapping Neural Network Incorporating a Dynamic Scaling Layer for Parametric Modeling of Microwave Components

Abstract

Space-mapping neural network (SMNN) technology has been widely applied for parametric modeling of microwave components. However, existing SMNN technologies struggle to address the challenges posed by the unknown and unevenly distributed numerical outputs of the mapping neural network (MNN). This letter proposes an enhanced SMNN structure that incorporates a dynamic scaling layer to tackle with this challenge. In the proposed structure, the equivalent circuit model is used as a coarse model. The relationship between the geometrical parameters and circuit element values is learned by an MNN. The numerical distribution of the MNN’s outputs is adjusted by the dynamic scaling layer with additional scaling factors for the circuit element values. A two-stage modeling method is proposed to train the enhanced SMNN structure. Using the proposed enhanced SMNN structure allows us to integrate the regulation of the numerical distribution of the MNN’s outputs into an automated framework, avoiding the risk of gradient vanishing or explosion during the training process, consequently achieving a higher modeling accuracy. Two microwave modeling examples are used to demonstrate the advantages of the proposed method.