Regional estimation of seismic response sequences for non-instrumented buildings using deep learning and indicator buildings

Regional seismic risk assessment is hindered by the limited availability of instrumented buildings, as most structures lack monitoring systems to record seismic responses. Existing estimation methods face a trade-off between capturing detailed structural dynamic behaviors and computational efficiency, typically providing only peak responses or requiring computationally intensive numerical simulation for full sequence prediction. This study proposes a novel deep learning-based Dynamic-Static Seismic Network (DSS-Net) that enables real-time estimation of full seismic response sequences for non-instrumented buildings. By fusing ground motions, building attributes, and the dynamic responses of a limited set of instrumented indicator buildings, DSS-Net generalizes to new non-instrumented buildings using only their attributes- reducing the need for numerical models or retraining. The model was evaluated in a simulated high-seismic region using synthetic data generated from nonlinear time history analyses (NLTHA) of 123 buildings subjected to 90 ground motions. DSS-Net achieved high accuracy in estimating roof acceleration sequences at real-time speed. An ablation study underscored the dominant contribution of building attributes to model performance. Comprehensive analyses identified the optimal indicator buildings, confirmed the model’s robustness against noise and generalization to out-of- distribution structures, highlighting its scalability and efficiency for regional seismic assessment.