Deep learning-based prediction of seismic response histories of base-isolated structures

Abstract

Conventional machine learning approaches in earthquake engineering have primarily focused on predicting peak structural response quantities, often neglecting the time-dependent nature of seismic behavior. This study applies a sequence-learning framework to directly map ground-motion time series to the full dynamic response history of base-isolated structures. A numerical dataset is generated using a MATLAB/Simulink simulation environment, where a multi-degree-of-freedom base-isolated system is subjected to recorded earthquake motions through linear time-history analysis. Both seismic input and structural response are represented as time series, allowing the model to learn temporal relationships without relying on predefined intensity measures or simplified response descriptors. A sequence-based Long Short-Term Memory (LSTM) model is employed to reconstruct response histories, including displacement and acceleration. The predicted responses show strong agreement with numerical simulation results, demonstrating the ability of the model to capture both amplitude variation and temporal evolution of structural behavior. By combining physics-based simulation with deep learning, the proposed methodology provides an efficient alternative for time-dependent seismic response estimation and offers potential for rapid structural assessment under earthquake loading.