Inverse Design Of Metamaterials Via Deep Learning For Electromagnetically Induced Transparency

Electromagnetically induced transparency（EIT）,a quantum interference effect that eliminates light absorption in opaque media,has associated non-optical properties that allow light to pass through the medium in a narrow spectral range.Metamaterials,an artificial electromagnetic medium with a negative refractive index on the subwavelength scale,have become a paradigm for electromagnetic space engineering and controlled wave propagation.Since experimental conditions of extremely low temperature and high laser intensity are usually required to realize the EIT effect in atomic systems,however,metamaterials can generate EIT phenomena at room temperature,which not only avoids the strict experimental conditions,but also metamaterials can cause more pronounced EIT phenomena,making it necessary to study EIT metamaterials.The study of EIT metamaterials usually requires the calculation of the electromagnetic response of all structural parameters by using numerical simulations,and the design and optimization of metamaterials is often a time-consuming and effortful task due to the lack of empirical relationships between structural parameters and the corresponding electromagnetic response.Traditional optimization algorithms such as genetic algorithms and particle swarm algorithms require several trials and errors to find the best design solution,and these trials and errors are easily affected by the local extremes,which leads to the failure to obtain the global optimal solution.Further,the physical limitations of metamaterial structures could not be solved in the past under the conditions of high time cost and complex algorithms in inverse engineering.This thesis proposes limiting the value range of metamaterial structural parameters through a single structural parameter acquisition method（SSPAM）for the first time,which will meet the expected values of our predictions and obtain high-quality and effective data in a relatively short time.The first attempt to use this method to effectively solve the physical limitation problem in the inverse design of metamaterials,which is a further improvement of the inverse design,once again enhances the credibility of the inverse design of metamaterials and realizes the idea of on-demand design.The mean squared error（MSE）of our best deep learning model is 0.00075 and 0.00026 in the training set and validation set,respectively,and 3.0×10^-5 in the test set.We input three specific points of the EIT spectrum into our optimal model to predict the corresponding EIT structural parameters inversely,verified by numerical simulation calculation,and obtained satisfactory results.The main work of this thesis is as follows:1.The basic theory of EIT metamaterials and the working principle of the neural network are introduced and the advantages of using deep learning for the inverse design of EIT metamaterials are discussed.Explains the physical limitations in the inverse design of metamaterials.An introduction to Py Torch,the deep learning framework used in this thesis.2.The SSPAM method is applied to limit the numerical range of the three variable structural parameters of the selected EIT structure,and its transmission spectrum is used as the optical response.Subsequently,the collection of the dataset is performed with the joint simulation of Python and CST software,and pre-processing was performed for the primary data,mainly including conversion of data values,and removal of invalid data values.A total of 15,000 sets of valid data were collected,with 5,000 sets of data collected for each of the 3 structural parameters.3.Build deep learning models,which are Least Squares Generative Adversarial Networks（LSGAN）,one-dimensional residual networks（1D-ResNet）,and deep neural networks（DNN）,respectively.The activation function and the optimizer are then chosen.By analyzing the MSE of the training sets and test sets,as well as the smoothness of the model training,the most suitable activation function for this work is Rectified Linear Unit and the optimizer is Adam.4.By building a comparative model of LSGAN,1D-ResNet,DNN is selected as the final research model of the work in this thesis.After that,the number of neural network layers and the number of samples in the dataset are analyzed for the DNN model.The MSE on the test set is used as a metric to evaluate the performance of the model after the optimal model is obtained by training on the training and validation sets.The optimal model of this thesis is obtained:the number of neural network layers is 5 and the number of samples is 5000.The DNN model-based inverse design of EIT metamaterials finally designed in this thesis is simple and has very high prediction accuracy,and this work can provide new ideas and methods for the inverse design of metamaterials in other models.