Explainable AI model for predicting equivalent viscous damping in dual frame-wall resilient system

A prominent challenge in applying the direct displacement-based design (DDBD) method to the proposed dual frame-wall lateral force-resisting system lies in determining the equivalent viscous damping ratio (EVDR). However, the strong nonlinearity and complexity behind the equivalent procedure lead to limited choice, mostly trial and error based on experience, to explain and predict the EVDR in the context of traditional research. This study employs the XGBoost method to unravel intricate relationships of EVDR using over 5 million data points from nonlinear time–history (NLTH) analyses, encompassing various parameters including the fundamental period, ductility, subsystem stiffness ratios, post-yielding stiffness ratios of the subsystems and ground motion types. SHapley Additive exPlanations (SHAP) values consistently identify critical features relevant to the equivalent procedure. Comprehensive feature ablation tests further illuminate the robustness and susceptibility of each model. Additionally, the incorporation of Local Interpretable Model-agnostic Explanations (LIME) for local interpretability provides insights into the intricate decision-making mechanisms inherent in each model’s predictions. Both the predicting results from machine learning (ML) and traditional method are also compared. Findings highlight the relative importance of features for EVDR and present a refined prediction model. It underscores the pivotal role of model interpretability in reinforcing confidence in complex models and advocates for leveraging ML techniques to enhance the effectiveness and efficiency of the DDBD method in structural design.