2024 年 21 巻 15 号 p. 20240289
An accurate and efficient electromagnetic (EM) modeling approach based on Neuro-space mapping (Neuro-SM) is proposed for substrate integrated suspension line (SISL) devices. A new coarse model containing 1 layer of dielectric plates, 2 layers of metal plates and 2 air cavities are developed to accelerate EM model simulation speed. To improve EM model accuracy, a parameter mapping network and a frequency mapping network are added into the structure. The EM responses of the SISL devices are taken as the fine model, while the proposed coarse model and two mapping networks are considered as the advanced Neuro-SM model. A three-step training process is developed to reduce errors between the fine and Neuro-SM model. The reasonable Neuro-SM model structure and efficient training algorithm greatly simplify the SISL device modeling process and further shorten the SISL device design cycle. The experimental results of a low-pass filter with SISL structure show that the training and test errors of the proposed model are 1.1% and 1.3%, which proves that the established model could accurately represent the characteristic of the SISL device.