An efficient deep-learning-based framework for optimization-based inverse design of electromagnetic metasurface design is proposed in this paper. A novel unit-cell parameterization strategy generates 16-element structures via symmetry operations governed by ten geometric parameters, overcoming the inefficiencies of pixel-based representations. A dataset of 16,000 parameter–reflection phase pairs is constructed, and a hybrid model combining Convolutional Neural Network (CNN), attention mechanisms, and the Kolmogorov–Arnold Network (KAN) is developed for broadband response prediction. The coefficient of determination (R²) of the proposed model is 0.8837 in the 2–18 GHz band, which is 44.87% higher than the R² without KAN. The proposed chessboard metasurface achieves a 10 dB monostatic radar cross-section (RCS) reduction under normal incidence over a wide frequency band from 7.4 to 15.2 GHz, corresponding to a relative bandwidth of 69%. This approach provides a generalizable, data-efficient solution for intelligent metasurface design.