Neural Network-Based Inverse Design of Metasurfaces with a Physics-Informed Intelligent Algorithm

This paper proposes a novel physics-informed machine learning approach for metasurface inverse design, which integrates a physics-assisted genetic algorithm with a residual neural network (ResNet) to generate optimal antenna structures based on input performance requirements. To reduce parameter dimensionality and enhance training accuracy, we incorporate coupled mode theory (CMT) into the genetic algorithm for efficient parameter fitting. The proposed inverse design methodology ingeniously combines ResNet with CMT principles, significantly streamlining the traditionally cumbersome metasurface design process. The CMT framework effectively reduces network complexity and scale while elucidating the intricate relationship between geometric configurations and reflection spectra. By simply specifying target performance metrics, our system can produce optimized patch unit structures that surpass existing dataset limitations. As a test case, we applied this inverse design network to a liquid crystal-based metasurface. Experimental results demonstrate the capability to achieve phase differences exceeding 360°, thereby validating both the effectiveness and practical utility of our proposed methodology.