Interpretable damage state-dependent fragility models for multi-hazard interactions based on Bayesian modeling and causal inference

• Bayesian fragility models for sequential earthquake–flood loadings. • Damage state–dependent models improve predictive power and reliability. • Period elongation is a key predictor of structural degradation. • Causal DAGs enable explainable and transparent multi-hazard modeling. • Supports scenario analysis and digital twin applications for resilience. Considering the increasing exposure of cities to multiple hazards, disaster risk management requires trustworthy fragility models that account for cumulative damage at the building level. This study develops damage state- and hazard-dependent fragility models for a reinforced concrete building subjected to sequential earthquake and flood events. Using Bayesian inference with Hamiltonian Monte Carlo, we construct and compare statistical models based on predictive accuracy, guided by information criteria, a proposed causal directed acyclic graph, and domain knowledge. Results highlight the importance of the existing damage state in improving predictive power. Incorporating structural damage proxies, hysteretic energy, and especially period elongation, further enhances model accuracy. Notably, once period elongation is accounted for, the type of hazard causing the prior damage has minimal additional effect, consistent with the causal assumptions. These findings support applications in life-cycle assessment and decision support systems, while the developed models enable “what if” scenario analysis involving the existing damage and other predictors related to structural performance. The physically interpretable and causally structured models align with the requirements for safety–critical and digital twin applications where transparency and reliability are essential.