PSI - Issue 84

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

394

complex nonlinear relationships between structural parameters, intensity measure (IM), and damage outcomes (Mangalathu et al., 2018; Li et al., 2025). In the context, this paper proposes an ML-based framework for fragility assessment of multi-span simply supported prestressed reinforced concrete (RC) bridges under two different exogenous hazards: traffic loading in combination with material degradation (i.e., corrosion). In detail, the methodology is described in Section 2 and is based on a new taxonomy in which categories and classifications are designed on the bridge attributes collected in the prioritization steps of the IG and involved in the Structural-foundational risk classification. For each taxonomy branch, nonlinear fiber-based finite element (FE) models of the superstructure are analyzed through simplified numerical analysis. Results, referred to the collapse limit state (LS) for flexural and shear mechanism, are used to train XGBoost ML algorithm according to a classification problem: determine whether the LS is reached based on the set of input parameters. The physical description of the ML surrogate model results is performed through an eXplainability approach. The results of the application are presented in Section 3 where the effectiveness of the ML surrogate model for fragility assessment was validated over FE numerical results. 2. Methodology The proposed methodology is illustrated in Fig. 1. As preliminary step of the framework, the identification of a bridge superstructure system. In this case, as shown by the state-of-the-art (D’Angelo and Salvadori, 2024; Natali et al., 2023; Salvatore et al., 2024) simply supported multi-girder reinforced concrete bridges is the most recurrent bridge typology in the Italian transportation network. The proposed taxonomy (Table 1) is based on structural and geometric information available at the end of the risk classification level of the IG: maximum span length ( L_MAX ), and superstructure defectiveness level ( SDL ). These parameters are strictly related to the expected behavior of the structure, as required for a taxonomy, and their influence is demonstrated by previous literature studies (Miluccio et al., 2022; Nettis Al. et al., 2024). In this way, each taxonomy branch is identified using a set of labels representing different combinations of the considered categories L_MAX/SDL . In detail, the maximum span length, L MAX , and the equivalent percentage reduction of the corroded steel area, ξ , are the so-called “associated parameters” to the two categories L_MAX and SDL , respectively. In the case of Structural-foundational risk classification of the IG, both maximum and average span length are taken into account. Still, L MAX is considered in the classification of structural-foundational vulnerability (MIT, 2020) and therefore is selected. ξ is expressed as: where, A s is the initial area of non-corroded steel, and A corr is the steel area after corrosion process. As second step, a plane grillage nonlinear fiber-based FE model of the superstructure is defined for the push-down displacement-control analysis under Traffic Load Model 1 (LM1) reported by the Eurocode 1 – Part 2 (CEN, 2003). The grillage model is composed of longitudinal beams linked by diaphragms in the transverse direction. The slab is modelled as part of the beam section for an effective width (CEN, 2004), while one-dimensional elements independent of the diaphragm (Hambly, 1991) are considered in the transverse direction. Each fiber of beams and diaphragms sections, depending on the assigned material, is characterized in the elastic and post-elastic field through proper stress strain relationship, excluding the ones referring to the slab which is assumed elastic in a global assessment analysis. Modified stress-strain relationships are used to account for corrosion effects. Based on the ξ value, modified elastic modulus, ultimate strength, and ultimate strain of steel are derived according to the model proposed by Yu et al. (2022). Additional information to those provided by the abovementioned parameters (i.e., “associated parameters”) are required to generate FE models for the analysis. For this reason, independent and dependent variables are defined, such as material mechanical properties (e.g., concrete compressive strength, f cm,b ) and geometric properties of the deck (e.g., height of beams, H b ). Every independent parameter is characterized by a proper statistical distribution and the Latin Hypercube Sampling (LHS) by Iman Conover (1982) is used as sampling technique to generate random samples s A A A − corr s  = (1)