SA-PhyGRU: A Self-Attention-Enhanced Physics-Informed GRU for Structural Seismic Response Prediction with Small Datasets

Accurate prediction of structural dynamic responses is critical for seismic analysis and decision-making throughout the structural life cycle. While model-driven and data-driven approaches have advanced practice, reliable prediction under limited data remains challenging due to the high cost of acquisition and simulation. This study proposes a Self-Attention-Enhanced Physics-Informed Gated Recurrent Unit network, SA-PhyGRU, for efficient and accurate seismic response prediction. The proposed network integrates GRU dynamics with a self-attention mechanism to capture long-range temporal dependencies and improve computational efficiency, while embedding physical constraints to enhance fidelity and generalization. Numerical and experimental validations on a three-story frame and a California hotel building show that SA-PhyGRU consistently outperforms conventional baselines in both accuracy and runtime, achieving improvements of up to 11.6% in R2, with pronounced gains in small-sample regimes. These results highlight SA-PhyGRU as an effective and generalizable approach for structural seismic response prediction and performance evaluation.