Generative networks-driven inverse design of VO₂ metasurfaces for dual-function terahertz spectral manipulation

Vanadium dioxide-based (VO₂) reconfigurable metasurfaces (RM) rely on the precise engineering of meta-atoms to achieve dynamic manipulation of electromagnetic (EM) waves. However, designing these structures inversely is physically challenging due to the complex mapping required to simultaneously coordinate spectral responses across different phase-transition states. To address the scarcity of high-performance designs in dual-state scenarios, we present a Jones matrix theory-guided spectral design framework tailored for VO₂ integrated meta-atoms. By synergizing a conditional variational autoencoder with a generative adversarial network (CVAE-GAN), our approach efficiently utilizes the vast design space to locate structures that satisfy rigorous spectral requirements in both insulating and metallic states. Furthermore, by integrating a novel prescreening mechanism, our method evaluates and optimizes candidate solutions, reducing target deviation by 6.25% and maximizing the fidelity of the generated structures. We numerically validate three representative THz functionalities-switchable polarization conversion, switchable absorbers, and dual-function switchable RMs integrating both-confirming the capability of the method to tailor multi-state optical responses within a unified design process. The proposed framework offers a practical route for the inverse design of RM, particularly in scenarios where multi-state or multi-function spectral targets challenge conventional optimization approaches.