Application Of Deep Learning In The Inverse Design Of Metasurface Absorbers

Metasurface absorber（MA）is a kind of functional device with a strong absorption effect on the incident electromagnetic waves.It has important application value in the fields of electromagnetic stealth,infrared detection,color display and microwave communication.With the rapid development of science and technology,people have higher and higher requirements for the performance of the MAs.How to design and optimize the metasurface quickly and efficiently has become one of hot topics in the research field today.Usually,the design of the metasurface depends on the designer’s experience and physical inspiration.It is possible to obtain the desired optical properties by using electromagnetic simulation software to simulate the pre-designed metasurface hundreds of times.This empirical trial and error method has the disadvantages of long design time,low efficiency,high calculation cost,and waste of resources.Deep learning（DL）,a sub-field of artificial intelligence and machine learning,has become a new method of metasurface design.It is a data-driven method,which allows the network model to automatically find useful information from a large amount of data,which is in sharp contrast to the methods based on physics or rules.Based on deep learning,thesis designed three generalized metal-insulator-metal（MIM）type metasurfaces with a single size and different surface morphology.The research work of this paper mainly includes the following two parts:Firstly,the convolutional neural network was used to inverse design MA.In order to accelerates the design and optimization process of metasurface and improve the prediction accuracy of the optical response of metasurface,a deep learning model composed of classification network and prediction network was proposed in this thesis.The model takes the optical response of three kinds of metasurfaces as the inverse design object,and takes the data pair composed of the absorption spectrum and structural parameters（or labels）of the metasurface as the learning object.On basis of this model,we can discover the hidden law between the data,and realize the classification and inverse design of the metasurface.To prove the effectiveness and practicability of this method,we designed six hand-drawn spectra as the prediction object according to human will.The prediction results show that the predicted spectra are in good agreement with the hand-drawn spectra.Secondly,the convolutional autoencoder network was used to inverse design MA.When training the network model,the larger the amount of data,the longer the training time of the network model.To further accelerate the design of metasurface absorber,a depth neural network model composed of classification network,convolutional autoencoder network and prediction network was proposed in this thesis.It not only accelerates the classification and inverses the design process of metasurfaces,but also realizes the reconstruction of data after dimensionality reduction.To prove the effectiveness and practicability of the model,we took three hand-drawn spectrums as the prediction object.The results show that the prediction spectrum is in good agreement with the hand-drawn spectrum.