PSI - Issue 78

Gianluca Quinci et al. / Procedia Structural Integrity 78 (2026) 845–851

848

main beams, either precast or cast in place, supported by elastomeric or metallic bearings. The superstructure also includes reinforced concrete deck slabs, transverse beams, bearing plinths, edge curbs, and other secondary elements. For each bridge, an in-depth geometric survey was conducted to gather the fundamental structural information. The data acquisition campaign included: • total bridge and individual span lengths, • dimensions of piers (height, width, cross-section), • deck width, • quantity and dimensions of main and transverse beams, • geometry of curbs and bearing plinths. This information was organized into a structured digital database that accurately reflects the geometrical attributes of each asset. In parallel, original design documentation was recovered and analyzed to obtain mechanical properties of the materials used during construction. The extracted information included: • concrete compressive strength, elastic modulus, and ultimate strain, • steel yield and tensile strengths, and Young’s modulus, • layout and thickness of the longitudinal and transverse reinforcement. The integration of archival documents with field survey data offered a comprehensive understanding of the as-built condition of the structures, an essential input for subsequent numerical modeling and seismic performance evaluation. To simulate the seismic behavior of these bridges, a procedure based on nonlinear static analysis (pushover) was adopted, following the N2 method guidelines, Fajfar 2000. This approach allows for estimation of structural capacity under seismic loading, taking into account inelastic behavior, which cannot be captured by linear dynamic procedures. To reduce computational complexity, each bridge was idealized as a single-degree-of-freedom (SDOF) system, with an equivalent mass reflecting the deck portion associated with a representative pier. The stiffness was derived from the ratio of the yield shear force to the corresponding yield displacement, obtained through moment – curvature analysis of the piers. These curves were built using the mechanical properties of concrete and steel, enabling identification of key parameters such as yield and ultimate displacements and shear capacities. The pushover procedure was performed by developing the Acceleration-Displacement Spectrum Response (ADSR) and constructing a bilinear capacity curve for each structure. The intersection of the capacity curve with the demand spectrum defined the Performance Point, which corresponds to the expected displacement of the system under the considered seismic input. A total of 3750 nonlinear static analyses were carried out (25 bridges × 150 input spectra), using a suite of 150 accelerograms selected to represent the regional seismic hazard, based on probabilistic hazard curves for Sicily. These analyses were performed using a custom MATLAB code, designed to handle the simulation workflow efficiently by dividing the process into computational blocks to manage large-scale execution. The primary output from each simulation was the maximum displacement reached by the deck under each seismic event. These displacement values served as the Engineering Demand Parameters (EDPs) used to develop fragility curves. The method followed the Cloud Analysis approach, in which each bridge was associated with 150 PGA – EDP pairs, forming a statistical cloud describing structural demand over a broad range of intensities. By applying log-linear regression to these data in the natural logarithmic domain, the median trend of the demand with respect to PGA was derived. The dispersion around this trend was quantified through the logarithmic standard deviation, under the assumption of lognormal behavior, i.e., the natural logarithm of displacement is normally distributed. The probability of exceedance was then calculated for each PGA level using the standard cumulative distribution function. The displacement threshold adopted in this study was the ultimate displacement (d u ) estimated via pushover analysis, representing a collapse-level performance state. Each fragility curve thus expressed the probability of exceeding this critical displacement as a function of increasing ground motion intensity. From each curve, the parameters μ (median demand) and β (dispersion) were extracted and used as output targets for training the ML models. The inputs to the learning algorithm were the geometric and mechanical features gathered earlier for each bridge. As a result, the ML models were trained to predict the seismic fragility characteristics, μ, β, and d u , directly from structural features, enabling efficient large-scale screening of seismic vulnerability across bridge networks.