Inverse Design of Frequency Selective Surface Using Physics-Informed Neural Networks
This paper uses Physics-Informed Neural Network (PINN) to design Frequency Selective Surface (FSS). PINN integrates physical information into the loss function, so training PINN does not require a dataset, which will be faster than traditional neural networks for inverse design. The specific implementation process of this paper is to construct a PINN using field solutions of mode matching method, and given the design goal, the PINN can train the shape of the diaphragms. The single frequency FSS that meets the design goal was designed using the inverse design method proposed in this paper without a dataset, verifying the rationality of using PINN to design metasurface. Using PINN for inverse design is not limited to single frequency FSS, but can also be used for more complex metasurface.