Inverse design of metamaterial structure based on AdaBelief residual neural network

In recent years, the design of metamaterial devices based on deep learning has made unprecedented development. However, for the reverse design of two-dimensional materials, it is difficult to solve the problems of falling into the local optimal value in a small sample space by using the traditional artificial neural network. Meanwhile, a lot of computational cost will be needed with the increase of the complexity of the structure. To solve these defects, a residual neural network based on AdaBelief optimization algorithm was proposed. The validity of the network was verified by choosing the design of multilayer alternating thin film structure based on graphene. The structure parameters of the multi-resonant perfect absorption spectra of samples were constructed by using the characteristic matrix method. The results show that the network model reaches 97% of prediction accuracy within a shorter training time. Compared with the prediction results of other neural networks, this network shows the advantages of high prediction accuracy and fast convergence rate, and achieves the design goal of perfect absorption metamaterial structure based on graphene.