Method to Obtain Structures in Multi-port Cavity Based on Physics-Informed Deep Learning

In this paper, a method to obtain structures in multi-port cavity is proposed. The time-reversal method is applied to obtain a rough physical topology of the intracavity structure. Further, a U-net is introduced to extract the final device structure. In addition, networks based on adversarial autoencoder with physics constraints have been proposed for multi-port device design with better results. Both methods achieve a better means of initial value construction, which will provide an initial solution for the iterative optimization method of the multi-port devices inverse design, and is even expected to be an alternative method.