Target-driven deep learning for optimization design of electromagnetically induced transparency metasurfaces based on lithium tantalate

We proposed a target-driven deep learning approach for the optimization design of metasurface. Take electromagnetically induced transparency(EIT)Lithium Tantalate metasurface as example, we showed how the method works. In order to accomplish the optimization design of the metasurface both a forward spectral prediction network (FPN) and an inverse design network (IDN) were designed. The FPN serves as a simulator, it enables rapid optical simulation and design; while the IDN is a hybrid neural network composed of Long Short-Term Memory (LSTM) layer and One-Dimensional Convolutional Layer (Conv1D), which enables rapid and accurate inverse design of metasurfaces. To validate the superior performance, comparative experiments were conducted with fully connected neural networks and One-Dimensional Convolutional Neural Networks (1D-CNN), which demonstrates that our network is both efficient and accurate. By comparing the errors of six structural parameters on the test set, we further analyzed the performance of the inverse design network. Finally, we performed in-depth simulations to investigate the impacts of various structural errors on EIT. Our proposed framework is an efficient strategy for metasurface inverse design, providing a scientific guide for the design of complex, multifunctional metasurfaces.