Efficient inverse design of single-substrate complex-amplitude holographic metasurfaces via Kolmogorov–Arnold networks

Single-substrate transmissive metasurfaces with independent amplitude and phase control enable high-quality holographic imaging. However, conventional design highly depends on computationally intensive and time-consuming simulations. To streamline the design process, the amplitude-phase response is often discretized into multiple steps, which inevitably introduces quantization errors, potentially missing the global optimum. Additionally, the complex residual coupling among parameters always reduces the design accuracy and increases the design complexity of holographic metasurfaces. To address these challenges, this work establishes an accurate mapping between structural parameters and electromagnetic (EM) responses by employing a Kolmogorov–Arnold network (KAN) with learnable activation functions. This network achieves high prediction accuracy and fewer parameters compared to conventional methods. By further integrating KAN with the particle swarm optimization (PSO) algorithm, the inverse design process is performed in a continuous parameter space, enabling rapid and precise retrieval of meta-atom parameters. Based on this hybrid framework, two complex-amplitude holographic metasurfaces are designed to demonstrate high-quality holographic imaging. The proposed intelligent framework is suitable for efficient metasurface design and promising for broad applications in holography, metalenses, and vortex wavefront control.