Sensitivity analysis of different machine learning models in the seismic response assessment of bridges

Reliable prediction of the seismic fragility of bridges is critical for structural safety and the prioritization of risk mitigation actions. This work explores the application of Machine Learning (ML) techniques to estimate key fragility parameters, median demand (μ), dispersion (β), and ultimate displacement (d u ), from a reduced set of structural descriptors. A database of 25 simply supported bridges located along the A19 Catania–Palermo motorway in Sicily is developed using nonlinear static analyses, and fragility curves are derived via Cloud Analysis. Several supervised regression models, including Gaussian Process Regression (GPR), Random Forest, SVR, and others, are trained to map structural features to fragility parameters. Among them, GPR consistently shows the best performance. The resulting ML-based fragility curves exhibit excellent agreement with those obtained through conventional methods, confirming the validity of the approach. The study demonstrates the potential of ML to support rapid seismic screening of bridge inventories, significantly reducing computational costs while maintaining reliability. Future developments will aim to expand the methodology to different structural typologies and incorporate epistemic uncertainties in the learning process.