Research On Application Of Neural Network In Engineering Electromagnetics

In recent years,the development of artificial intelligence based on neural networks,such as Alpha Fold and Chat GPT,has brought new insights to scientific research.If neural networks can be applied to the field of engineering electromagnetics,it will have great potential and a wide range of application prospects.On the one hand,many electromagnetic problems require a large amount of time to be solved using traditional methods and are difficult to solve for some complex problems.On the other hand,neural networks have significant advantages in handling complex problems and improving efficiency,and are expected to improve current solutions.Therefore,this thesis takes the microwave circuits and antennas in the field of engineering electromagnetics as the background,and explores how to apply neural network technology to solve related problems,in order to achieve performance improvement,efficiency optimization,and innovation breakthroughs.The specific research content includes the following aspects:1.A deep investigation into the application of data-driven residual neural networks and convolutional neural networks in the field of microwave circuits has been conducted,attempting to solve the problem of coupling matrix extraction for microwave filters based on this foundation.A one-dimensional deep residual convolutional neural network has been constructed for this problem.In order to solve the problem of long dataset construction time,a fast dataset construction method based on coupling matrix theory has been proposed.Based on this neural network model and dataset construction method,the coupling matrix extraction of fifth-order and eighth-order cross-coupled filters,as well as fifth-order dielectric waveguide filters,has been realized,validating the effectiveness of this method in coupling matrix extraction.2.In another type of coupling matrix problem,namely coupling matrix synthesis,the Off-policy deep reinforcement learning algorithm,D3 QN,has been applied.Unique reward and action functions have been designed for this algorithm,allowing it to smoothly solve combinatorial optimization problems.Finally,the coupling matrices satisfying the target have been successfully synthesized for a six-order cross-coupled filter with transmission zeros and a trisection structure filter,indicating the feasibility and versatility of this method.3.In another major area of engineering electromagnetics,antenna field,specifically the directional pattern optimization and beamforming problem of array antennas,the Onpolicy deep reinforcement learning algorithm,PPO,has been applied to this problem based on the previously proposed reward and action functions.Through Python-HFSS co-simulation,the accuracy and optimization speed of the results have been improved.In the end,directional patterns that meet expectations have been successfully optimized in four examples of low sidelobe directional patterns and flat-top beamforming,demonstrating the feasibility of this method.