Inverse design of reconfigurable terahertz metasurface with wide phase modulation range via a high–precision CNN–transformer framework

Le résumé fourni par la source

Abstract Terahertz communication promises Tbit s −1 -level connectivity for 6G but is hindered by severe path loss and atmospheric absorption. While reconfigurable metasurfaces offer a promising solution via dynamic wavefront control, conventional design relying on empirical optimization and constrained geometric templates inherently restricts phase tuning ranges. To address these limitations, we propose an intelligent inverse design framework for liquid crystal metasurface with three synergistic innovations: (1) a non-uniform rational B-splines method is introduced to explore smooth topological geometries of metallic patterns, reducing the solution space from 100 to 8 degrees of freedom and simplifying subsequent optimization; (2) a hybrid convolutional neural network−Transformer model is developed to jointly learn local and global electromagnetic features, achieving high prediction accuracy (phase error: 1.03°, amplitude error: 2.8 × 10 −3 ); (3) the proposed network is integrated with a particle swarm optimization algorithm for inverse design, rapidly identifying the globally optimal design within high-dimensional parameter spaces across 0.25−0.55 THz. Finally, the optimized metasurface achieves 310° phase tuning and the reflection amplitude exceeding 0.9, with the liquid crystal thickness being only 1.4% of the working wavelength, surpassing the performance in prior literature. Based on array-level optimization algorithms, we further demonstrate highly directional beam steering with flexible control over beam quantity, power ratio, and beamwidth. This work presents an innovative design approach for reconfigurable terahertz metasurfaces, paving the way for their practical implementation in next-generation communication systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.