Abstract
Metasurface is a revolutionary platform to achieve desired properties by artificially engineering meta-atom's arrangements. However, the explosively expanding design space of advanced metasurfaces with multiple degrees of freedom (MDOF) has made the traditional human-guided design methods increasingly ineffective, limiting the development of the metasurfaces. Intelligent design methods have been presented to tackle these challenges by introducing innovative computational models, but they are predominantly data-driven and faced the issues of data scarcity, poor physical interpretability, and weak generalization capability. Here, a physics-driven intelligent design (PDID) paradigm is proposed and demonstrates its application to design MDOF multiplexed metasurfaces. The PDID method integrates the physical prior knowledge into a deep neural network, thereby enhancing its physical interpretability and reducing its reliance on extensive databases. Compared to the traditional intelligent designs, this can reduce both design time and database size by two orders of magnitude. Through experimental validation of MDOF multiplexed metasurfaces, the versatility and computational efficiency of PDID are showed. This method not only presents a novel intelligent design tool but also exemplifies the integration of physical knowledge with machine learning to address the challenges. Its interdisciplinary insights offer significant potentials for innovative applications across the materials science, computational science, and information technology.