Flexible On-Demand Design of Terahertz Multi-Resonant Metasurfaces Through Deep Learning

Multi-resonant metasurfaces, with their capability to achieve precise electromagnetic control across multiple frequency bands, play a crucial role in applications such as optical displays, sensing, absorbers and so on. However, designing multi-resonant structures is challenging due to the nonlinear complexities inherent in multi-frequency responses, which pose significant difficulties for traditional numerical optimization methods. In this work, we introduce an innovative deep learning network for the flexible, on-demand design of multi-resonant terahertz (THz) metasurfaces. This method utilizes masked- attention to enable precise control over the number of resonances and allows for the direct generation of desired metasurface structures by inputting specific resonance frequencies and amplitudes. Our approach significantly enhances the efficiency and flexibility of multi-resonant metasurface design, demonstrating broad potential for application in the design of other electromagnetic devices.