The terahertz (THz) frequency range has emerged as a promising spectral window for broad applications, including next-generation wireless communication, high-resolution imaging, and ultrafast spectroscopy. Among the essential components in these systems, amplitude modulators with high quality (Q) factors can provide sharp, selective frequency responses, which are key requirements for scalable and high-performance THz systems. However, designing high-Q THz modulators remains challenging, as conventional full-wave simulations are time-consuming and inefficient. In this study, we propose a deep learning-based inverse design framework tailored for THz metasurfaces composed of split-ring resonators (SRRs). The framework is built on a tandem neural network architecture that couples a forward model with an inverse network to retrieve structural parameters from desired spectral responses. To enhance physical feasibility and predictive stability, we introduce an autoencoder-based spectral projection method. Our model accurately reconstructs SRR geometries across a wide range of spectral targets by learning the underlying physical relationships. Notably, we demonstrate the inverse design of Fano resonant geometries characterized by high-Q factors and sharp asymmetric resonances, which are essential features for achieving deep modulation. By extending the tandem deep learning approach to the THz domain and incorporating an autoencoder-based spectral projection, our framework provides a scalable and efficient pathway for the rapid prototyping of tunable, high-Q THz devices and lays the foundation for artificial intelligence-driven design of advanced THz photonic components.
Jeong et al. (Mon,) studied this question.