Abstract

The development of metamaterials, constituted by periodic unit cells, has enabled exotic wave phenomena, including negative refraction and total reflection, that are not observed in natural materials. While metamaterials show promise in creating such phenomena, the metamaterials’ bulky wavelength-comparable size poses challenges in practical applications. In contrast, metasurfaces have subwavelength-comparable size, resulting in thin structures while retaining capability to produce extraordinary phenomena, including anomalous refraction, by phase control of transmitted waves using generalized Snell’s law. The inverse design of metasurfaces for anomalous refraction requires maximizing transmittance while ensuring phase shifts of the transmitted waves match with the target phase shifts. Conventional inverse design methods have employed analytical and numerical approaches [1]. While analytical methods offer computational efficiency, the simplified modeling assumptions limit performance and applicability. Numerical methods provide superior performance for complex structures and versatility across different problems; but the numerical methods are computationally intensive. Recent deep learning-based inverse design techniques have been shown to offer high performance with rapid inference time. However, most studies have not addressed elastic metasurfaces, which exhibit more complicated characteristics due to mode coupling. The few studies on data-driven inverse design of elastic metasurfaces use deep learning-based techniques only for forward model, relying on genetic algorithms for inverse design, thus failing to overcome the computational limitations of conventional methods [2]. To develop an on-demand inverse design framework for elastic metasurfaces with anomalous refraction, this study proposes an attention-based hierarchical conditional GAN model. The proposed method incorporates three key components: (1) hierarchical conditions in the GAN discriminator to include target responses and whether the designs belong to the bandgap; (2) attention mechanisms to capture complex relationships between design variables and target responses; and (3) a deep learning-based surrogate forward model to determine the optimal generative model. Experimental results present that the generated designs achieve target responses with rapid generation. The study also analyzes attention scores to reveal the relationships between design variables and target responses.