PSI - Issue 84

Mirko Calò et al. / Procedia Structural Integrity 84 (2026) 392–400

398

XGBoost algorithm predictions for the derivation of fragility curves. As for dispersion of fragility curves, β , the ones derived through the application of XGBoost algorithm systematically recorded lower values for both failure mechanisms, with larger differences for shear failure one. Consequently, the probability of failure is overestimated at low IM levels and underestimated at high IM levels. These differences in β values can be attributed to the limited variability of the FEM results at the lowest and highest IM values, which biased XGBoost predictions in these IM ranges. Despite these differences, the fragility curves derived from the XGBoost model are overall consistent with those obtained in previous studies (Nettis Al. et al., 2024). As expected, corrosion increases the probability of failure for both collapse mechanisms and it affects more the flexural mechanisms than the shear one as shown in Fig 3c and Fig. 3d. Nevertheless, shear governs the failure mechanism across all L_MAX categories, while flexure mechanisms are less critical because of the favorable plastic redistribution. The dominant failure mode, especially for shorter spans, is shifted from shear to flexure as corrosion increases.

Fig. 3. Comparison of fragility curves derived by XGBoost predictions and by numerical analysis (a, b); Fragility curves derived from the application of the XGBoost algorithm (c, d). 4. Conclusions This paper proposed an ML-based framework for fragility assessment of multi-span simply supported prestressed reinforced concrete bridges under two different exogenous hazards: traffic loading in combination with material degradation (i.e., corrosion). The methodology is based on a taxonomy designed to account for structural-foundational parameters involved in the risk classification step of the Italian Guidelines. Simplified finite element (FE) models, accounting for two failure modes (i.e., flexural and shear), were proposed and numerical analysis were used to generate training/test database for eXtreme Gradient Boosting (XGBoost) Machine Learning (ML) algorithm. XGBoost was trained according to a supervised classification problem: determine whether the collapse limit state is reached based on the set of input parameters. Fragility curves derived for the two failure modes were validated through the FE model numerical analysis showing promising results for the application to other superstructure bridge typologies. It should be emphasized that the proposed ML surrogate model is not a surrogate for numerical analysis, i.e., the single output