In this study, a hybrid physics-based and data-driven model is proposed to reliably forecast the dynamic responses of structures under time-varying excitations. While detailed physics-based models are accurate, they are computationally intensive; conversely, purely data-driven models are efficient but highly dependent on data quality and volume. The proposed hybrid approach bridges this gap by combining a fast, low-fidelity physics-based model with a data-driven corrector to predict the residuals between the low-fidelity approximation and the high-fidelity target. Specifically, the low-fidelity physical model relies on a multi-degree-of-freedom mass-spring-damper system, while the data-driven corrector is based on a selective state-space neural network. The viability of the proposed model is demonstrated using a multi-story structural case study featuring nonlinear behavior. The model achieves high predictive accuracy, with a Mean Absolute Percentage Error of 4.76% across the test set. Furthermore, it significantly reduces computational time compared to purely physics-based models and requires less training data than purely data-driven approaches. By enabling rapid and reliable predictions of structural behavior under extreme dynamic loading, this framework provides an efficient tool for near-real-time structural health monitoring and post-event vulnerability assessment, directly contributing to the safety and resilience of the built environment.
Mai et al. (Mon,) studied this question.