Interpretable natural gradient boosting model for probabilistic response evaluation of RC columns under blast loads

Reinforced concrete (RC) structures are susceptible to partial or complete progressive collapse initiated by column failures under blast loads. Understanding and predicting the RC column’s responses subjected to blast loads is crucial for implementing proactive solutions to protect life and mitigate economic loss. This research aims to develop a probabilistic displacement prediction model for RC columns under blast loads by adopting a new machine learning algorithm, Natural Gradient Boosting (NGBoost), and to understand the influence of design parameters on the peak and the corresponding discreteness of the RC column’s displacement responses by interpreting the developed NGBoost model. The research outcomes demonstrate that the developed NGBoost model achieves superior accuracy compared to the single-degree-of-freedom (SDOF) method, concurrently offering robust estimates of prediction uncertainties. The feature of the blast loads, including the reflected impulse and reflected pressure, shows a much higher influence than the design parameters on the displacement responses of the RC columns under blast loads. Design parameters such as longitudinal steel reinforcement ratio and concrete compressive strength exhibit noteworthy influence on peak displacement, while longitudinal steel yield strength, transverse steel yield strength, and concrete compressive strength significantly impact the discreteness of the RC column's displacement responses under blast loads. These insightful findings can serve as practical references for enhancing the design of RC columns against blast loads.