Inverse Design Method Of Electromagnetic Metamaterials Based On Machine Learning

In the past two decades,the theories,design and engineering applications of metamaterials have received great attention from academia and industry and achieved considerable results.At present,mature and effective metamaterial design methods mainly rely on the rapid development of computational electromagnetics and HPC(High Performance Computing),with the help of electromagnetic simulation software such as CST,HFSS,COMSOL,etc.to directly numerically solve Maxwell's equations.With the continuous improvement of the complexity of the structure,process and function of metamaterials,in order to achieve the performance indicators in terms of amplitude and bandwidth,researchers must do a large number of electromagnetic simulations repeatedly adjusting and modifying models and materials parameters,or rely on optimization algorithms such as PSO to optimize the parameters of a single model.Such a design process is cumbersome,and it places certain requirements on the professional skills of researchers.How to automatically excavate the physical connotation behind the electromagnetic response of metamaterials and achieve on-demand design for specific design goals is still one of the major challenges in the future.In recent years,AI(Artificial Intelligence)technology represented by ML(Machine Learning)has achieved considerable results in the fields of image segmentation and recognition,image super-resolution reconstruction,machine translation,human-computer interaction,and data mining.At present,ML has been gradually introduced and served in metamaterial design.In the simulation design stage,if these simulation results can be collected and processed,it will greatly reduce the participation of designers,design costs,and significantly improve design efficiency,even be excavated to find suitable potential physical mechanisms.Based on machine learning theory,this paper studies the inverse design of absorber metamaterials and gratings.The DNN(Deep Neural Network)is used to predict the absorption spectrum of the absorber.The GAN(Generative Adversarial Network)is leveraged to realize the on-demand inverse design of the absorber.The interactive learning method of reinforcement learning is utilized to accomplish the structure inverse design of the plane grating.The first chapter introduces the research background and the current research results of domestic and foreign scholars.The second chapter introduces the basic electromagnetic properties of metamaterials and related theories of machine learning,then discusses the feasibility and advantages of artificial metamaterial design based on ML.Chapter 3 introduces a DNN is constructed to predict the electromagnetic properties of the absorber,and discusses a data preprocessing method to make the model training more stable.From the experimental results,the trained model has high efficiency and accuracy.Chapter 4 shows the on-demand inverse design of the absorber structure based on the GAN,introduces the concept of fuzzy design,and proposes a feature compression method that enables the model to achieve "one-to-many" mapping.Finally obtains a broadband design scheme in the range of 8?20GHz.The fifth chapter introduces the interactive learning method to inverse design the grating unit.The model does not need to obtain any data set in advance,which means the model can "learn while simulating".It is quickly and accurately to obtain the transmission efficiency at a specific working wavelength and deflection angle.Throughout the experiments in this article,humans do not have much design intervention.As a typical data-driven design method,once the mathematical model is constructed,it will be able to quickly produce a design scheme that meets the requirements.This paper compares the design results based on machine learning with electromagnetic simulation,proves that these methods effectively meets the design criteria and provides new possibilities for the reverse design of metamaterials.