An efficient on-demand inverse design of terahertz metasurfaces via deep learning

Deep learning-assisted metasurface inverse design has accelerated the development of terahertz (THz) wave technologies for next-generation communications and sensing systems. However, conventional approaches are limited by the task-specific architecture paradigms and redundant dataset regeneration when new needs arise. In this work, we propose an efficient on-demand design method (EDDM) for THz metasurfaces to enhance design flexibility and reduce resource consumption. The method integrates genetic algorithms (GA) with transfer learning-accelerated artificial neural networks (ANN) for rapid target-driven optimization. The physics-partitioned ANN models with shared parameters are employed to reduce data dependency. Concurrently, the GA-driven adaptive fitness functions are integrated to translate electromagnetic (EM) requirements into quantifiable design objectives. EDDM integrates structured dataset arrangement with hyperparameter optimization of ANN to enable full reuse of pre-trained models, significantly reducing computational demands while maintaining design precision. As a demonstration of applications, the spectrum-customized metasurfaces, high-efficiency transmitarray antennas, and polarization converters are validated based on EDDM. The results indicate the design exhibits multi-adaptability across amplitude, phase, and polarization specifications respectively. EDDM establishes a versatile paradigm and offers a practical approach to expedite the applications of metasurfaces in demand scenarios.