Metasurfaces offer an unprecedented avenue to facilitate light-matter interactions. However, traditional design methodologies rely on computationally intensive trial-and-error processes. Moreover, existing deep learning (DL) schemes are predominantly hindered by their massive data requirements and limited exploration of freeform design spaces. To overcome these challenges, a multi-model-driven generative-evolutionary strategy (GES) is proposed, for the on-demand inverse design of bespoke Terahertz (THz) metasurface sensors. By leveraging a Conditional Diffusion Generator (CDG) and an Attention-Enhanced Residual Network (ARN), this framework enables the exploration of an expansive design space encompassing 2100 possible configurations. The GES effectively overcomes the data bottleneck by selectively generating high-potential data in stages. Full-wave simulations confirm that the inversely designed metasurfaces exhibit high-contrast resonance peaks and exceptional sensitivity across low, mid, and high THz bands. This work provides a versatile paradigm for the efficient design of high-performance functional metamaterials, significantly accelerating the advancement of application-specific THz sensing.
Zhang et al. (Sat,) studied this question.