ABSTRACT Accurate prediction of nonlinear structural responses under strong earthquakes is essential for performance‐based assessment and rapid decision‐making for a given structural configuration. Deep learning (DL) offers a promising approach, yet existing models often face challenges in robustness, long‐sequence modeling, and physical interpretability. Physics‐informed neural networks (PINNs) typically impose constraints through the loss function, which can create conflicting objectives and slow or destabilize convergence, especially in complex nonlinear dynamics. To overcome these limitations, we propose a Physics‐Guided Mamba‐Attention Hybrid (PG‐MAH) model for computationally efficient seismic response prediction after offline training. The framework comprises the following three components: (1) a front‐end Temporal Multi‐Head Self‐Attention (TMHA) mechanism applied to the seismic excitation sequence to identify and emphasize critical time steps; (2) a Mamba model for capturing long‐range temporal dependencies from the TMHA‐enhanced sequence; and (3) a physics‐guided residual correction model embedding the Newmark‐ scheme to promote dynamic consistency. Numerical examples indicate that the proposed PG‐MAH can achieve high predictive accuracy and robustness while maintaining favorable computational efficiency. This study considers a fixed structural configuration in each example. The global dynamics operators used by the physics‐guided correction layer are assembled from the corresponding OpenSees model; when the structural configuration changes, these operators must be updated and the surrogate generally requires fine‐tuning using nonlinear time‐history analysis (NLTHA) data for the new configuration.
Zheng et al. (Tue,) studied this question.