On Demand Design Of Metamaterials Based On Deep Learning

Metamaterials have artificially manufactured periodic sub-wavelength structures,which can be artificially designed to meet various required optical responses,so metamaterials have huge development potential.With the rapid advance of the terahertz field in recent decades and the unique advantages of terahertz waves,researchers have begun to pay attention to the research of metamaterials in the terahertz band.Terahertz metamaterials,as an important and new type of components,have a wide range of applications in the fields of terahertz imaging,detection and communication.Therefore,it is particularly important to design an expected terahertz metamaterial structure.The commonly used process of designing metamaterials involves solving Maxwell's equations,and this process has highly nonlinear and complex boundary conditions.For complex structures,this process is almost impossible to achieve.The purpose of this article is used by deep learning methods to solve the above problem.This paper designs a deep learning model.By letting the neural network learn a data set containing the structure of the metamaterial and its corresponding optical response,the network can obtain the ability to predict the structure of the metamaterial through the optical response.And,this network can also accurately predict its optical response based on the structure of the metamaterial.Once the model is trained,it can be transplanted to portable devices and made into instruments for rapidly designing metamaterials.The main contents of this article are as follows:1.This paper studies the basic theories and typical scientific research results of terahertz and metamaterials to understand their main advantages and practical applications.First,this paper conducts theoretical research on deep learning,clarifies the classification and working principles of neural networks,and focuses on the research of BP neural networks.After that,this paper compares the commonly used deep learning frameworks and explores their advantages and disadvantages.Research shows: Pytorch takes into account multiple aspects such as speed,flexibility and simplicity,and is the most suitable deep learning framework for this article.2.This paper explores the basic theory and development process of electromagnetic induction transparency(EIT)phenomenon in detail.Due to the uniqueness and practicality of EIT phenomenon,in this paper,EIT device is taken as an example to present the working principle and performance of the entire model.Define the five variable structural parameters of the selected EIT device,and use its transmission spectrum as the optical response.After that,CST was used to collect data sets,and a total of 15,000 sets of data were collected.The preprocessing of the original data mainly includes the processing of missing data,the processing of data outliers and the normalization of data.3.Construction of deep learning neural network.First,this paper analyzes the four commonly used activation functions in detail,mainly focusing on the function image,formula,and advantages and disadvantages.The results show that the learnable parameter‘a' in PReLU solves the problem that certain neurons will not be activated and can improve optimization performance,so the most suitable activation function for this model is PReLU.Then there is the selection of the optimizer.The three typical optimizers are compared in detail,and the optimization principle of each optimizer is analyzed.Using the MSE of the training set and the test set as an indicator,the performance of the three optimizers is visually displayed.Research shows that: Compared with the other three optimizers,Adam can optimize to a lower MSE and is more efficient,so Adam is used as the network optimizer.4.The optimization of the number of neural network layers and the number of samples in the data set is also an important work.In terms of the number of neural network layers,it mainly optimizes the number of hidden layers containing 800 neurons in the inverse network.The MSE of the training set and the test set is used as an indicator to evaluate the performance of the model,and a comparison is made in the presence of 1 to 7 layers.Research shows that: 5 layers are the optimal situation.In terms of the number of samples in the data set,five different situations are selected for analysis.Using the MSE of the training set and the test set as an indicator,research shows that the optimal number of samples is 10,000,which balances time cost and forecast accuracy.The metamaterial on-demand design model based on deep learning designed in this paper is simple in structure and has extremely high prediction accuracy.It provides a convenient method for the field of metamaterial design,which is of great significance.