<p>While the seismic response of isolated structures has been extensively studied, the response of nonstructural components within such structures has received significantly less attention, despite their potential to cause economic losses and human casualties. The floor response spectra are the standardized solution for analyzing seismic demands of nonstructural components. However, their calculation faces computational bottlenecks: inherent local nonlinearities increase computational expense, while repeated time-history analyses for different structures constrain efficiency. Therefore, this paper proposed a deep learning-based approach for the accurate and efficient prediction of floor response spectra of isolated structures. First, based on relevant standards and literature, ground motions and structural parameters were comprehensively selected, thus constructing a high-quality and large-scale dataset. Subsequently, the Top Floor Response Spectrum Network (TFRSNet) and the Other Floor Response Spectrum Network (OFRSNet) were proposed to predict the top floor and other floor response spectra, respectively. Then, an “overall-individual-pointwise” multi-scale accuracy verification approach was proposed and employed to verify the accuracy of the TFRSNet and OFRSNet comprehensively. Also, the computation time of the proposed prediction approach was compared with that of the conventional approach. Moreover, a graphical user interface was developed, realizing a fully visualized and interactive workflow. Finally, the proposed method was applied to the Yunnan Provincial Museum, and seismic analysis and optimal design of cultural relics fixed with supports were conducted. Results indicate the high accuracy and substantial efficiency gains of the prediction models, demonstrating their potential and convenience for practical engineering applications.</p>