Research On Electromagnetic Modeling And Application Based On Artificial Intelligence

Recently,with the rapid development of graphics processing units,deep learning has been applied to widespread applications,and thus bringing vast resources from software algorithms to hardware structures.Hence,it is convincing that deep learning can be beneficial to traditional electromagnetic field.On the one hand,the everaccelerating iteration cycle of wireless communication systems requires lower costs and faster design process,making it crucial to speed up the design process of microwave devices.On the other hand,considering the increasing amount of large-capacity transmission,it is essential to introduce reliable methods to increase the channel capacity,which makes the development of new methods indispensable.Hence,under the background of wireless communication system,this dissertation mainly focuses on the application technology of deep learning in the microwave field,while taking the two problems,i.e.,optimizing the design process of traditional devices and developing new methods to expand the channel capacity,as key issues.Concretely,the possibility and limitations of applying deep neural network to the two key issues are thoroughly investigated.The main content of this dissertation is as follows:1.The reliability of data-driven artificial neural network(ANN)model and convolutional neural network(CNN)model in microwave field is studied and compared.To speed up the parameter extraction process of microwave device filters,the parameter extraction models based on two typical network architectures are designed,respectively.The performance of the two proposed architectures in electromagnetic regression problem is then analyzed.Further,considering the speed limitation of in filter tuning process,a method incorporated with the designed CNN is proposed.Finally,the designed filter model verifies the efficiency of the data-driven CNN method.2.The reliability of another typical network structure,namely recurrent neural network(RNN)model,in sequential electromagnetic problems is studied.To tackle with the misalignment detection of vortex wave carrying orbital angular momentum,a more generalized vortex misalignment theory is firstly proposed.Based on the proposed theory,RNN with gated recurrent units is designed.Through comparisons under different misalignment cases,the efficiency of proposed RNN model is verified.To further compare and analyze the performance of different network architectures in sequential electromagnetic problems,the misalignment detection models based on CNN and ANN are also designed,validating the superiority and reliability of RNN in electromagnetic problems with sequence structures.3.The accuracy problem of artificial neural network in microwave field is investigated.Taking the fast multipole method,a key algorithm for large-scale problems,as a specific example,the problems lying in its translation process are studied.To tackle with the problems,several data-driven neural network models are proposed.Further,the limitations of the single network models in accuracy and storage are pointed out.Based on the physical information lying in the calculation of translation process,a hybrid network architecture is proposed and applied to various examples.With high precision and low storage,the proposed hybrid architecture is promising to relevant problems.4.The performance degradation phenomenon of convolutional neural network method in microwave engineering application is studied.Accordingly,a CNN model based on physical knowledge is constructed and the effects of the amount of physical prior knowledge is then analyzed.Taking the orbital angular momentum modes detection of vortex electromagnetic waves as a specific example,the effectiveness of CNN in the case of sufficient sampling points is first demonstrated by the data-driven method.Then,the dilemma over the high resolution required by CNN and the less sampling points needed in engineering application is discussed.And the performance degradation problem of CNN is pointed out.To tackle with this,based on the characteristics of helical phase front,a convolutional network model with different degrees of physical knowledge is proposed.It is shown that the proposed model presents high reliability and robustness under limited sampling points,suggesting great potential in the practical application of radio vorticity communication.This work builds a systematic modeling method for electromagnetic problems incorporated with deep learning,which solves the problems lying in the traditional electromagnetic methods,as well as promotes the development of the emerging field,i.e.,the electromagnetic modeling based on deep learning.Consequently,this dissertation can provide promising guidance to future relevant research.