Abstract: Metasurface has demonstrated significant value across various fields due to its exceptional ability to manipulate electromagnetic waves. However, the highly complex nonlinear mapping relationship between their geometric configurations and electromagnetic responses has led traditional metasurface design methods to rely on massive electromagnetic simulations, severely constraining design efficiency. To achieve a breakthrough in this bottleneck, this study generated a uniform dataset of 20 000 sets of geometric data-reflection phase pairs across an ultra-wideband frequency range of 2–18 GHz. A high-performance forward-phase prediction neural network constructed using the Kolmogorov-Arnold convolutional neural network (KAN-CNN) module, attention mechanisms, and residual connections, combined with a simulated annealing algorithm to rapidly generate metasurface structural parameters from a target phase. Experimental results show that this metasurface inverse design system achieves a high-accuracy broadband reflection phase prediction accuracy of 92.7%, with an overall model R² of up to 0.8937. Compared with traditional full-wave simulation iterative optimization, this system significantly improves design efficiency and enables the rapid generation of high-performance metasurfaces. Using this system, a focusing metasurface array operating at 8 GHz with a focal length of 100 mm was successfully designed and fabricated. The measured results closely match the design targets, validating the feasibility and reliability of the entire process from design to fabrication.