现代应用物理2025，Vol.16Issue(1)：165-169,5.DOI:10.12061/j.issn.2095-6223.202412033
基于特征提取的CNN-LSTM高效时域电磁算法
Efficient Time-Domain Electromagnetic Algorithm Based on CNN-LSTM Network With Feature Extraction
摘要
Abstract
When solving time-domain electromagnetic response problem,accurate results can be obtained by using full-wave numerical methods such as finite difference time domain(FDTD).However,in complex scenes,it is often difficult to reduce the amount of computation and improve computational efficiency while ensuring computational accuracy.In addressing the challenges posed by the limited computational efficiency and substantial computational demands of conventional FDTD mothed,this paper employs a convolutional neural network(CNN)combined with a long short-term memory networks(LSTM)to effectively solve the time-domain electromagnetic response of the targets.First,a dataset of target and time-domain electromagnetic responses is constructed using the traditional FDTD methods.Then the geometric and electrical parameter distribution features of the target are extracted by constructed CNN.Subsequently,the nonlinear mapping model between target features and time-domain electromagnetic response is then constructed by using LSTM.Finally,a dedicated time-domain training algorithm is designed to train the CNN-LSTM model.The proposed CNN-LSTM model has the capacity to accurately predict the time-domain electromagnetic response of targets in specific scenes.Numerical simulation results of the 3D model demonstrate that the computational efficiency of the CNN-LSTM model is approximately 500 times that of the FDTD mothed,with a relative mean deviation of less than 0.03,validating the accuracy of the proposed method.关键词
时域有限差分/时域电磁散射计算/卷积神经网络/长短期记忆网格/深度学习/电磁正演Key words
FDTD/time-domain electromagnetic scattering computation/CNN/LSTM/deep learning/electromagnetic forward modeling分类
数理科学引用本文复制引用
彭傲,陈娟,殷岳萌,李少龙,王成浩,赵翠荣..基于特征提取的CNN-LSTM高效时域电磁算法[J].现代应用物理,2025,16(1):165-169,5.基金项目
国家重点研发计划资助项目(2020YFA0709800) （2020YFA0709800）
中国电波传播研究所稳定支持科研经费资助项目(A132303219) （A132303219）
