Inverse Design Method for Electromagnetic Periodic Structures Based on Physics-Informed Neural Network With Embedded Analytical Models

To achieve an efficient inverse design method for electromagnetic devices, this article introduces the physics-informed neural network with embedded analytical models (EAM-PINN). This approach combines embedded physical knowledge and external physical constraints and is applied to the inverse design of electromagnetic periodic structures. In EAM-PINN, we embed the physical knowledge of periodic structures into neural networks, specifically by replacing ordinary neurons with periodic neurons containing Floquet mode solutions to form neural networks and output electromagnetic fields. Then, we use the mode matching method to link the electromagnetic field with the structures, integrating them into the loss function as external physical constraints. Through EAM-PINN, we successfully perform inverse design of artificial magnetic conductors (AMCs) and frequency-selective surfaces (FSSs), demonstrating its effectiveness in designing electromagnetic periodic structures. Compared with traditional neural networks, EAM-PINN inherits the benefits of traditional PINN, requiring fewer training data or even no data at all, and achieves faster inverse design. Moreover, EAM-PINN exhibits stronger learning capabilities and easier convergence compared with the traditional PINN.