A dual-dynamic deep learning model with graph Kolmogorov–Arnold networks and Bi-GRU for seismic power spectral prediction

With the accumulation of earthquake records and advances in structural earthquake engineering, accurately predicting earthquake power spectra has become increasingly important for assessing structural reliability. However, conventional power spectral models struggle to characterize spectral modulation under varying magnitude–distance scenarios. As a result, they often yield overly smooth predictions that fail to reproduce spectral fluctuations and multi-peak structures, thereby compromising the accuracy of safety evaluations. To address this issue, this study proposes a dual-dynamic deep learning (DL) model that integrates dynamic Graph Kolmogorov - Arnold Networks (GKAN) and Bidirectional Gated Recurrent Units (Bi-GRU). The proposed model jointly captures similarity relationships among earthquake records and contextual dependencies along the frequency axis. Specifically, it adaptively constructs a dynamic K-Nearest-Neighbor (KNN) graph based on feature similarity to represent non-Euclidean associations among records under varying magnitudes and distances. In addition, leveraging dynamic graph learning and a dynamic fusion mechanism, the model performs bidirectional sequence modeling and adaptively reweights information contributions, enabling more effective capture of spectral fluctuations and multi-peak characteristics. Experimental results across multiple earthquake datasets demonstrate that the proposed model not only accounts for magnitude and distance effects but also accurately captures complex spectral variations, thereby improving predictive accuracy and generalization performance.