GAT-Net: Inverse design of multifunctional metasurface based on graph attention network

Metasurfaces have revolutionized the miniaturization and multifunctionality of optical devices and systems, offering unprecedented solutions for advanced applications. Yet, the intrinsic optical interactions between meta-atoms are a critical factor that can markedly affect and impose limitations on the overall performance of these metasurfaces. In this study, we present an inverse design methodology that employs a graph attention network (GAT-Net) to comprehensively address the mutual coupling effects in metasurface design. The GAT-Net-based approach stands out from traditional design methods by fully taking into account the complicated coupling interactions between meta-atoms in the design process. We show that this integration could improve the focusing efficiency of the designed metalens, showcasing the superiority of our model in terms of performance optimization. To further validate the effectiveness of design models with the GAT-Net, we successfully engineered an achromatic metalens, a metasurface showing multi-wavelength multiplexing optical routing, and a metasurface for color holographic image projection. The results highlight GAT-Net’s capability of handling multiple wavelengths simultaneously and its potential to fully account for the coupling effect to a full extent. This advancement in design efficiency and multichannel capability benefits the development of highly functional metasurfaces.