Abstract: With the advancement of intelligent civil engineering, physics-informed neural networks (PINNs) have demonstrated the capability to rapidly predict structural dynamic responses under conditions of limited or even absent sensor data. However, constrained by the black-box representational nature of multilayer perceptrons (MLPs), conventional PINNs struggle to accurately characterize structural responses involving high-frequency components and multifrequency coupling, and they generally exhibit limited physical interpretability. Under seismic excitations, structural dynamic responses typically consist of low-frequency dominant modes coupled with high-frequency local vibrations, and this pronounced multiscale spectral characteristic further exacerbates spectral bias, convergence degradation, and interpretability issues in PINNs. To address these challenges, this study proposes a novel physics-informed framework, termed SDR-FPIKAN, which integrates Fourier feature mapping (FF) with the Kolmogorov–Arnold network (KAN) for seismic structural dynamic response analyses. In this framework, Fourier feature mapping is introduced at the network input to project the temporal domain into a multiscale frequency space, thereby explicitly enhancing the model’s ability to resolve high-frequency and multifrequency components. Within the network architecture, a KAN structure is employed to replace conventional fixed activation functions with learnable univariate functions, enabling the structural response to be expressed as explicit analytical forms composed of trigonometric and polynomial functions, while remaining the consistent with the governing dynamic equations. The deep integration of FF and KAN endows the model with both enhanced spectral representation capability and interpretable functional representations, fundamentally alleviating the spectral bias and interpretability limitations inherent in conventional PINNs. Numerical case studies involving a ten-storey shear frame structure and a two-storey steel frame structure under seismic excitations are conducted to systematically compare SDR-FPIKAN with the conventional PINN. The numerical results demonstrate that SDR-FPIKAN outperforms the traditional PINN in terms of prediction accuracy, of convergence efficiency, of physical interpretability, and of numerical stability, thereby validating its effectiveness in addressing complex structural dynamics problems.