Wind-resistant shape design exploration of civil engineering structures fundamentally relies on the accurate interpolation and extrapolation of bluff-body aerodynamic features. Current methodologies typically emulate global, integrated coefficients, such as flutter derivatives and admittance functions, in the frequency domain, adopting black-box models. However, they often fail to capture sudden aerodynamic transitions due to data sparsity and lack the phenomenological information inherent to the fluid–structure interaction (FSI). This study proposes a novel deep learning strategy combining: (1) a dual-band flow superposition approach separating the prediction of flow fields for different frequency bands, (2) the BiFIS sampling strategy for shape-agnostic and accurate prediction of flow fields, and (3) a novel physics-informed FSE-U-Net architecture that conditions flow reconstruction on global physical parameters through feature-wise linear modulation (FiLM), ensuring physical consistency across variable flow regimes.