A Novel Convolutional-Autoencoder Based Surrogate Model for Fast S-parameter Calculation of Planar BPFs
A Novel Convolutional-Autoencoder Based Surrogate Model for Fast S-parameter Calculation of Planar BPFs
复制标题
DOI：
10.1109/ims37962.2022.9865285
复制
发表时间：
2022-06
期刊：
影响因子：
--
通讯作者：
Ren Shibata;M. Ohira;Zhewang Ma
中科院分区：
文献类型：
--
作者：
Ren Shibata;M. Ohira;Zhewang Ma
近年来，基于深度学习的代理模型被引入来加速电磁分析。例如，利用卷积神经网络（CNN）构建了一个代理模型，该模型的输入是被分析的几何图像，输出是其特征。然而，建立这样的代理模型需要大量的电磁仿真结果。为了解决这一问题，本文提出了一种新的基于卷积自编码器（CAE）的代理模型，该模型由CAE中的编码器和密集层组成，用于从平面带通滤波器（bpf）的电路图案图像中计算$S$参数。作为一个例子，通过无监督学习和迁移学习构建了典型三阶BPF的代理模型。通过与传统的基于cnn的代理模型进行比较，评价了基于cae的代理模型的泛化性能。
Recently, surrogate models based on deep learning are introduced to speed up electromagnetic (EM) analysis. For instance, a surrogate model, ofwhich the input is image of geometry under analysis and the output is its characteristics, has been constructed by convolutional neural network (CNN). However, a large amount of EM simulation results is required to build such surrogate models. To solve this problem, this paper proposes a novel convolutional-autoencoder (CAE) based surrogate model, which consists of an encoder in the CAE and dense layers, to calculate $S$-parameters from circuit-pattern images of planar bandpass filters (BPFs). As an example, the surrogate model for a typical third-order BPF is constructed through unsupervised and transfer learnings. The generalization performance of the CAE-based surrogate model is evaluated by comparingwith that of conventional CNN-based one.