Metasurface Design Based On Neural Network

Metasurface is an important part of new nanophotonic devices.These materials have achieved perfect absorption,super-resolution imaging,nonlinear optics and many other applications.In general,people often use advanced iterative calculation methods,combined with finite element modeling(FEM)or finite difference time domain(FDTD)to realize the prediction of optical properties of metasurface materials and the design of metasurface structures,but this traditional design process will be affected by the inherent human induced error and the process is very time-consuming.In this thesis,we train a deep learning network model to realize the design of metasurface.When a set of desired spectra is input,the trained generation network will output a metasurface pattern with input optical characteristics.Firstly,the paper introduces the theoretical knowledge of metamaterial and neural network,and then designs the neural network architecture according to the research needs.The network is mainly composed of simulator,discriminator and generator,in which the simulator is convolutional neural network,and the generator and discriminator constitute a generation countermeasure network.Finally,a fitting metasurface generator is trained through continuous optimization and adjustment.Then the test and result analysis of the generated network are carried out.Finally,the specific application example of the network is shown.In this thesis,a metasurface generation network is finally implemented.Compared with the traditional design method,it greatly shortens the design time and reduces the professional requirements of the physical field required by the metasurface design,which can serve the users who lack relevant knowledge.On the other hand,because it is generated by neural network and unsupervised learning method,it does not rely on human experience to design,which can provide reference for researchers to design metasurface structures with certain optical properties.