Abstract: To address the challenges of low modeling efficiency and high simulation cost of frequency selective surfaces (FSS) under high-degree of Freedom (High-DOF) parameter spaces and complex topologies, this study proposes an artificial intelligence-based electromagnetic forward modeling approach. A neural network architecture based on deep operator network (DeepONet) is constructed, where the branch network incorporates an improved ResNet-18 to effectively extract multi-scale spatial features from FSS topology images, while the trunk network takes frequency as an explicit input to enhance the modeling capability of frequency responses. By learning the nonlinear mapping between topology and frequency response, the model achieves efficient prediction of the S₂₁ parameter across the 2–20 GHz frequency band. The proposed framework adopts an offline training and online testing strategy, enabling the model to rapidly generate full-band responses from a single structural input, thereby significantly improving inference efficiency. Experimental results show that the model's average relative error on the validation set is 0.0478, the coefficient of determination (R²) is 0.99441, and the average prediction time per sample is 6 ms. Compared with traditional finite element method (FEM) and finite-difference time-domain (FDTD) approaches, the proposed AI-based model eliminates the need for repeated meshing and re-modeling, substantially reducing computational cost. It provides a feasible and efficient path for rapid modeling and intelligent computation of complex electromagnetic structures such as FSS.