PSI - Issue 78

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

850

Among all models tested, Gaussian Process Regression demonstrated the best predictive accuracy, yielding a coefficient of determination R² = 0.82, indicating a strong correlation between the predicted and actual values. This result underscores the model’s capacity to effectively learn from sparse data while preserving robustness and precision.

Table 1. An exam

Parameter

ML model

R 2

β

GPR GPR GPR

0.82 0.81

d u

μ 0.79 For conciseness, Table 1 reports the best-performing model for each fragility parameter. In all cases, GPR emerged as the most accurate model, outperforming other methods in terms of generalization and reliability. To validate the model’s practical applicability, the fragility curve predicted by GPR for a new bridge configura tion, excluded from the training dataset, was compared against a reference curve obtained via full nonlinear time history analysis. As shown in Figure 2, the two curves are closely aligned, with a mean absolute error (MAE) of less than 2% across the considered PGA range. This high level of agreement confirms that the ML-based estimation approach, although approximate, offers sufficient accuracy for early-stage seismic risk screening. It enables infrastructure managers to efficiently generate fragility curves for a wide inventory of structures, significantly reducing computational burden and supporting large scale prioritization efforts.

Fig. 2. Comparison between the fragility curve obtained with the ML model and the one derived using the traditional method

5. Conclusions This work proposes an innovative strategy for generating seismic fragility curves of bridge structures through the application of Machine Learning (ML) techniques, relying exclusively on a compact set of structural input features. By leveraging regression models to estimate the primary parameters of fragility functions, namely the median capacity