Abstract

Terahertz metasurface holography requires unit cells with prescribed complex transmission responses and sufficient compatibility with practical fabrication processes. For elementary low-bit phase quantization at 0.1 THz, conventional parameter sweeping of simple geometries remains an effective and reliable approach. However, when the design objective is extended from a single phase state to simultaneous consideration of complex transmission matching, transmission amplitude, structural connectivity, and fabrication-related geometric constraints, the design problem becomes a constrained multi-objective inverse-design task.

Here, we propose a generative-model-assisted inverse-design framework for 3D-printable terahertz metasurface unit cells. The framework combines a VGG-based surrogate model with a modified WGAN-GP generator. The surrogate model maps binary unit-cell patterns to the real and imaginary parts of the transmission coefficient, while the generator produces candidate structures conditioned on target complex transmission responses. A surrogate-assisted physical-response loss and a content regularization loss are introduced to guide the generated patterns toward both the desired transmission response and the constrained structural distribution. The generated unit cells are then selected according to target phase states and assembled into 1-bit, 2-bit, and 3-bit holographic metasurface arrays using phase distributions obtained by the Gerchberg–Saxton algorithm.

Model evaluation shows improved generation stability compared with tested baseline models, with a mean absolute error of approximately 0.06, an average accuracy above 0.95, and an average coefficient of determination of approximately 0.96 after convergence. Ceramic 3D printing and 0.1 THz holographic experiments further demonstrate that the generated structures can support terahertz holographic imaging, especially for 1-bit and 2-bit designs. The 3-bit experiment shows that higher-bit phase quantization is more sensitive to accumulated design, fabrication, array-level, and experimental uncertainties, highlighting the need for improved design-to-experiment error control in future high-bit designs.

This work provides a fabrication-aware generative inverse-design workflow for terahertz metasurface holography and clarifies both its potential and current fabrication-related limitations.