Fast Prediction Method For Electromagnetic Response Of Reflectarray Antenna Unit Cell Based On Machine Learning

In recent years,reflectarray antennas have become a technical solution with low cost and high reliability in many applications,including satellite communications,microwave communications,aerospace technology,etc.Compared with the previous reflectarray antennas,the new reflectarray antenna has a flat reflector surface,simple structure,convenient processing,manufacturing and transportation,and no need of complex power feed network.At the same time,the emergence of reconfigurable reflectarray antenna enables the antenna to change the basic working characteristics of the independent radiation unit cell to meet different working requirements.However,since there is no analytical solution,synthesizing such a high-performance reflectarray antenna is a very challenging task.The traditional method is to use the electromagnetic simulation software to calculate the scattering matrix of the reflectarray antenna unit cell.The commonly used software is HFSS and CST,and the results obtained by them are very accurate.However,the time required for simulation grows exponentially with the degree of freedom of the antenna unit cell,which is not conducive to our real-time analysis of antenna performance,in addition to requiring sufficient storage space.Therefore,they may not be used in practical antenna design and optimization.In order to solve the computational time and storage problems associated with traditional full-wave simulation methods,in this paper,based on the now well-established machine learning related theories,machine learning is mainly studied the application of metal reflectarray antenna,graphene reconfigurable patch antenna and graphene reconfigurable reflectarray antenna to speed up the design and optimization of antennas.The main research work is as follows:(1)The problem of predicting the electromagnetic response of a reflectarray antenna unit cell based on Support Vector Regression(SVR)is investigated.Firstly,the calculation of the scattering coefficients of a periodically spaced passive reflectarray antenna unit cell is converted into a regression estimation problem,and then an analytical model is built up by SVR to quickly predict the electromagnetic response of the antenna unit cell.For this purpose,the full-wave simulation software CST is used to obtain a set of random samples of the scattering coefficient matrix of the reflectarray antenna unit cell,which is used for SVR training.Under the same condition,the Radial Basis Function Network(RBFN)is used to predict the electromagnetic response of antenna unit cells.The comparative results show that the SVR algorithm is effective in the electromagnetic response prediction of reflectarray antenna elements.(2)The problem of predicting the electromagnetic response of graphene reconfigurable patch antennas based on SVR is investigated.The machine learning method of SVR is used for fast reconfiguration prediction of graphene patch antenna parameters to address the time-consuming problem of full-wave simulation of the radiation characteristics of conventional graphene reconfigurable antennas.The electromagnetic response of a graphene patch antenna with different parameters(patch size,chemical potential,frequency,etc.)is transformed into a vector regression problem.With the antenna unit cell parameters as input and the corresponding S-parameters as output,a regression model is established and the full-wave simulation software CST is used to build the SVR training dataset and test dataset for fast prediction of the electromagnetic response of graphene reconfigurable antenna unit cells.By predicting the parameters of S₁₁,and comparing with the results of RBFN and CST,the effectiveness of the method is verified.(3)The problem of predicting the electromagnetic response of graphene reconfigurable reflectarray antenna unit cells based on deep learning is investigated.Considering that the SVR algorithm used in the previous two works is not outstanding compared with RBFN in terms of computational efficiency,the Convolutional Neural Network(CNN)in deep learning algorithm is considered to be applied to the study in this section.The method first discretizes the input vectors(patch geometry,chemical potential,frequency,incident angle,etc.)of the graphene reflectarray antenna,then pre-processes the data into a 2D picture suitable for CNN training,and finally computes the electromagnetic response of the graphene reflectarray antenna using a CNN training model instead of extensive full-wave simulation.The training results of the three algorithms,CNN,SVR and RBFN,are compared together.The experimental results show that CNN has a better performance in the prediction of the electromagnetic response of graphene reconfigurable reflectarray antennas,with further improvement in accuracy and a time improvement of nearly 35.6%compared to the SVR algorithm used in the previous two works,verifying the effectiveness of the method.