Fast Optimization Of Complex Electromagnetic System Based On Simulation And Machine Learning

At present,the traditional design pattern for electromagnetic?EM?structures is mainly about manual optimization or iterative optimization based on intelligent optimization algorithms,which require a lot of simulation calculations,resulting in long design cycles and low design efficiency.Surrogate model technology uses a small amount of sample data generated by EM simulation software to construct an approximate mathematical model to predict the response value of unknown location,and then optimizes the global optimal solution of the specified parameter space,which is a feasible scheme to improve the efficiency of engineering project.However,the commonly used surrogate models have weak generalization ability,and poor fitting effect on high-dimensional and strong nonlinear EM structures optimization problems.Machine learning is introduced to build surrogate model to overcome these problems,improve the reliability and accuracy even using small sample data set.Therefore,it is one of the urgent problems to be solved to improve the design efficiency of EM structures on how to combine computationally accurate full-wave simulation software with machine learning surrogate models,and apply intelligent optimization algorithms with powerful global optimization capabilities to the EM field.In order to solve the above problems,CST simulation and machine learning surrogate models are combined to build a complex EM system for quick optimization design of single objective and multi-objective EM structure in the paper.The mian reseach and results are shown as follows.1)The theory of the complex EM system based on machine learning is analyzed.The optimization problem of EM structure is transformed into the optimization based on surrogate model.The basic principles of three sampling algorithms?MCS/LHS/LHSMDU?,four machine learning surrogate models?polynomial regression,support vector regression,decision tree and random forest?and two genetic algorithms?SEGA and NSGA-II?are described respectively.2)The general framework and design flow of the EM system are introduced in detail.The modules including joint simulation based on Python and CST,data processing,prediction and analysis based on machine learning algorithm,searching for the best design point based on genetic algorithm are all realized.The effectiveness of the design pattern proposed by the system is verified by taking Lévi nonlinear test function as an example.3)Based on the above EM system,the fast optimization design of T-waveguide power divider and TE₄₂ mode input coupler is completed respectively.First,the T-waveguide power divider is divided into two objective designs to realize the low reflection and equal power output:the single target task with the average reflection coefficient of Port 1 less than-25 dB in the 9 GHz?12 GHz working frequency band,and the multi-target task with the reflection coefficient less than-18 dB for each working frequency point.Then,a high-order TE₄₂ mode input coupler is designed,whose parasitic modes are well suppressed from 221 GHz to 230 GHz.Besides,its broadband and low reflection transmission is realized.