Deep Neural Network Training Integrating the Equivalent Circuit Model for Metasurface Design

This article presents a physics-informed deep learning framework using deep neural networks (DNNs) for metasurface (MS) design. By integrating equivalent circuit parameters as physically interpretable features, this framework effectively embeds the electromagnetic characteristics of MSs into the network architecture and enhances the DNN’s prediction accuracy and efficiency. Numerical experiments demonstrate that compared with conventional DNN models, the proposed framework achieves a 25% improvement in prediction accuracy and a 33% enhancement in design efficiency for MS optimization. Providing complementary physical insights, the physics-informed framework will facilitate machine learning applications in electromagnetic systems’ design and accelerate the development of MSs.