PSI - Issue 84

Gianluca Quinci et al. / Procedia Structural Integrity 84 (2026) 199–206

201

based frameworks coupled with cloud analysis—they are typically represented through a lognormal cumulative distribution, Shinozuka et al. 2000. Within this formulation, the fragility model is commonly defined by two key parameters: • μ, the median capacity, i.e., the intensity measure value (e.g., PGA) associated with a 50% probability of reaching the selected limit state; • β, the logarithmic standard deviation, accounting for dispersion due to record-to-record variability and modeling-related uncertainties. For bridges with relatively straightforward structural layouts, the median response μ is often expressed as a power law function of seismic intensity, for instance: = ∙ where a and b are calibration coefficients obtained empirically, and μ may correspond to a structural response quantity such as the maximum pier displacement. In order to define the fragility function entirely, the ultimate displacement capacity of the pier (d_u) must a lso be specified. Once μ, β, and d_u are available, exceedance probabilities at each intensity level can be computed through the standard lognormal expression: ( > | ) = 1 − ( ( )) where Φ denotes the standard normal cumulative distribution function. Building on this framework, the present study introduces a supervised Machine Learning (ML) strategy aimed at predicting the fragility-related quantities μ, β, and d u directly from a limited set of bridge geometric descriptors. The underlying idea is to train regression models capable of inferring these parameters from readily retrievable characteristics, including: • total structural mass, • pier height, • reinforcement areas (top and bottom) in the longitudinal direction. To generate training data, a benchmark database of bridge configurations is assembled by running nonlinear time history simulations over a broad set of representative bridge archetypes. Fragility parameters are then estimated via cloud analysis, producing a labeled dataset for ML model development. Different regression-based ML techniques are investigated to approximate the functional relationship between the selected input features and the fragility parameters. Once calibrated, the trained models allow rapid fragility-curve estimation for previously unseen bridges, avoiding computationally expensive simulations during the initial screening phase. Although the predicted fragility functions are intrinsically approximate, the approach supports scalable vulnerability evaluation across large bridge portfolios. In particular, it enables asset ranking based on estimated seismic fragility, thus helping decision-makers allocate resources more effectively and plan detailed assessments or retrofitting actions where they are most needed. In the next sections, the method is applied to a real-world case study in Sicily, an Italian region characterized by significant seismic hazard and a dense transportation network. The case study is used both to evaluate predictive performance and to refine the procedure for future large-scale regional applications. 3. Case Study: Seismic Fragility of Simply Supported Girder Bridges along the A19 Motorway The proposed framework was validated on a real infrastructure portfolio consisting of 25 bridges along the A19 Catania–Palermo motorway, one of the main transportation corridors in Sicily, crossing areas characterized by considerable seismic hazard. The examined assets are all simply supported girder bridges, a configuration extensively adopted in Italy for viaducts built between the 1950s and the 1980s. From a structural standpoint, these bridges feature