PSI - Issue 84

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

205

Table 1. R 2 values of the best ML model for deriving μ, β, and d u Parameter ML model R 2 β GPR 0.82 d u GPR 0.81 μ GPR 0.79

5. Conclusions This study introduces a Machine Learning (ML)–driven framework to generate seismic fragility curves for bridge structures using only a limited set of easily retrievable structural descriptors. In particular, regression models are employed to infer the core fragility parameters— median capacity (μ), dispersion (β), and ultimate displacement capacity (d_u)—from compact geometric and mechanical inputs such as total bridge mass, pier height, and longitudinal reinforcement characteristics. By doing so, the proposed strategy provides an efficient alternative to conventional fragility derivation workflows (e.g., simulation-based Cloud Analysis), which typically require extensive nonlinear analyses and significant computational effort. Across the investigated algorithms, Gaussian Process Regression (GPR) systematically delivered the most reliable predictions for all target parameters, as confirmed by the adopted validation indicators. Moreover, fragility curves reconstructed from the ML-predicted parameters showed a close agreement with those obtained from detailed nonlinear reference analyses, supporting the robustness of the approach and its capability to preserve accuracy despite a reduced input space and limited data availability. The main strength of the proposed methodology is its scalability: once trained, the ML models can rapidly produce fragility estimates for large bridge inventories, avoiding time-intensive numerical simulations during the initial assessment phase. This makes the framework particularly suitable for first-level seismic risk screening, portfolio ranking, and the definition of prioritization strategies for inspections and retrofitting interventions within infrastructure management programs. Future work will focus on broadening the method’s applicability by considering additional bridge typologies and alternative performance/damage criteria, as well as by explicitly accounting for epistemic uncertainty within the ML stage. These developments are expected to further improve the generality, predictive performance, and reliability of the proposed framework for large-scale vulnerability assessment. Acknowledgements This work was carried out within the Framework Agreement (Convenzione Quadro, ex art. 15 Law 241/1990) signed on 16 March 2021 between ANAS S.p.A. and the FABRE Consortium, supporting research for the implementation of the 2022 Italian Guidelines (Consiglio Superiore dei Lavori Pubblici) on risk classification/management, safety assessment and monitoring of existing bridges. References Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5 – 32. El - Maissi, A. M., Argyroudis, S. A., & Nazri, F. M. (2021). Seismic vulnerability assessment methodologies for roadway assets and networks: A state - of - the - art review. Sustainability, 13, 61. https://doi.org/10.3390/su13010061 Fajfar, P. (2000). A nonlinear analysis method for performance - based seismic design. Earthquake Spectra, 16(3), 573 – 592. https://doi.org/10.1193/1.1586128 Maulud, D., & Abdulazeez, A. M. (2020). A review on linear regression comprehensive in machine learning. Journal of Applied Science and Technology Trends, 1(2), 140 – 147. Quinci, G., Paolacci, F., & Phan, H. N. (2023). Artificial Neural Network Technique for Seismic Fragility Analysis of a Storage Tank Supported by Multi - Storey Frame. ASME J. Pressure Vessel Technol., 145(6), 061901. https://doi.org/10.1115/1.4063242 Quinci, G., Phan, N. H., & Paolacci, F. (2022). On the Use of Artificial Neural Network Technique for Seismic Fragility Analysis of a Three Dimensional Industrial Frame. ASME PVP Conference, Las Vegas, USA. https://doi.org/10.1115/PVP2022 - 83874