A residual multi-layer perceptron framework for transmission prediction and physical interpretation of terahertz responses in multilayer compound metal–dielectric metasurfaces

Abstract

Metamaterials have emerged as promising candidates for terahertz applications. Among different structures, metal–dielectric composite periodic arrays have gained significant interest due to their ability to efficiently manipulate electromagnetic wave propagation and enhance transmission in the terahertz range. In this work, we present a deep-learning-driven modeling framework based on a Residual Multi-Layer Perceptron (ResMLP) to predict the transmittance of multilayer metal–dielectric composite periodic annular-aperture arrays. A large-scale dataset comprising geometric parameters and transmittance spectra obtained from full-wave electromagnetic simulations was used to train the ResMLP. The trained model achieves highly accurate predictions of optical responses, with a validation mean squared error of 0.000 166. To further assess reliability, a “closed-loop verification” was performed, in which the predicted geometric parameters were reintroduced into the forward simulator to regenerate the transmittance spectra, yielding high consistency with the target spectra. Compared to full-wave simulation, our approach offers up to 105-fold improvement in computational efficiency without compromising accuracy. This study demonstrates the potential of deep learning in accelerating the design and optimization of multilayer metal–dielectric metasurfaces, thereby facilitating the development of terahertz photonic devices.