Accurate prediction of seismic demand in nuclear power plant reactor containment buildings is essential for reliability assessment and safety management. While nonlinear time-history analysis (NTHA) provides realistic responses, its high computational cost limits its applicability in large-scale probabilistic studies that account for material and structural uncertainties. To address this challenge, this study proposes an explainable multi-target machine learning framework based on Extreme Gradient Boosting to simultaneously predict two key engineering demand parameters): peak floor acceleration (PFA) and peak floor displacement (PFD). The model is trained on a dataset of 2000 NTHA simulations incorporating uncertainties in damping ratio, concrete compressive and tensile strengths, steel yield strength, and structural mass. Shapley Additive Explanations (SHAP) are employed to quantify the relative importance of input variables and enhance interpretability. Results show that the proposed approach achieves high predictive accuracy with significantly lower computational requirements compared to conventional NTHA. Hyperparameter sensitivity analysis identifies regularization strength, minimum child weight, and learning rate as the most influential parameters for model performance. SHAP analysis further reveals that damping ratio predominantly governs PFA, while damping ratio, concrete strength, and mass strongly influence PFD. These findings demonstrate the effectiveness of explainable multi-target learning for efficient, accurate, and transparent seismic demand prediction of critical nuclear containment structures.
Azad et al. (Tue,) studied this question.