PSI - Issue 78

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

847

approaches integrated with cloud analysis, these curves are commonly modeled using a lognormal cumulative distribution, Shinozuka et al. 2000. This probabilistic model is generally characterized by two parameters: • μ, representing the median structural capacity, i.e., the intensity level (e.g., PGA) at which there is a 50% chance of reaching the chosen performance limit; • β, the logarithmic standard deviation, which captures variability stemming from both seismic input and model-based uncertainties. • For bridge structures with relatively simple configurations, it is customary to relate the median capacity μ to the seismic intensity using a power-law relationship, such as: = ∙ where a and b are empirically calibrated coefficients, and μ may represent a critical response parameter like the maximum pier displacement. To fully characterize the fragility function, the ultimate displacement capacity of the pier (denoted as d u ) must also be known. Once these quantities are defined, the probability of exceedance at each intensity level is calculated through the standard lognormal formulation: ( > | ) = 1 − ( ( )) where Φ is the standard normal cumulative distribution function. The strategy adopted in this study proposes the use of supervised Machine Learning (ML) to estimate the fragility parameters (μ, β, and d u directly from key geometric features of the bridge. The core concept is to develop regression models that learn to predict these parameters based on a subset of easily accessible structural descriptors, such as: • total structural mass, • pier height, • reinforcement areas (top and bottom) in the longitudinal direction. To this end, a reference database of bridge configurations is created by performing nonlinear time-history simulations on a diverse set of representative bridge archetypes. The corresponding fragility parameters are then extracted using cloud analysis, yielding a labeled dataset suitable for ML training. Several ML regression techniques are explored to model the mapping between input features and fragility parameters. Once trained, these models can be deployed to rapidly estimate fragility curves for new bridges, eliminating the need for time-intensive simulations in the early stages of seismic risk screening. While the resulting estimates are approximate by nature, the methodology offers a practical solution for scalable vulnerability assessments over large infrastructure portfolios. In particular, it enables the prioritization of assets based on predicted seismic fragility, allowing decision-makers to better allocate limited resources and schedule interventions accordingly. In the following sections, this approach is tested on a real-world case study in Sicily, an Italian region with both significant seismic exposure and a dense road network. The case study is used to assess the predictive capabilities of the trained ML models and fine-tune their application for broader regional evaluations. 3. Case Study: Seismic Fragility of Simply Supported Girder Bridges along the A19 Motorway . The proposed methodology was applied to a real-world scenario involving 25 bridges located along the A19 Catania – Palermo highway, a primary infrastructure corridor in Sicily that traverses regions of significant seismic hazard. All structures under consideration belong to the category of simply supported girder bridges, a typology widely employed in Italy for viaducts constructed between the 1950s and 1980s. These bridges are composed of longitudinal