PSI - Issue 78

Andrea Nettis et al. / Procedia Structural Integrity 78 (2026) 1404–1411

1407

2.4. Preliminary results Initial performance evaluations indicate that the proposed model achieves high accuracy, with a correct classification rate of approximately 90%. The use of attention mechanisms enhances the model’s ability to generalize across varying image conditions and corrosion patterns, reducing the rate of misclassification. Fig. 2 provides a representative output, highlighting the model's ability to accurately identify a region affected by high-level corrosion. 2.5. CV-informed mass loss sampling Following the image-based classification, each predicted corrosion level is translated into a corresponding mass loss range ( Q corr ), as defined in the CONTECVET guidelines (Fagerlund, 2001). These ranges provide probabilistic input for structural modeling, allowing corrosion effects to be stochastically incorporated into simulations. For each severity class — Low, Medium, and High — a uniform probability distribution is assumed within the specified mass loss interval, enabling consistent and statistically informed degradation modeling within the simplified structural analysis workflow.

Fig. 2. Representative application of CNN prediction.

3. Structural Modelling and Fragility Analysis 3.1. Material Properties Degradation and Modeling Uncertainties

Mechanical property variability due to limited structural knowledge was modeled via Gaussian sampling, using statistical distributions from the Italian bridge dataset by Zelaschi et al. (2016) for parameters such as concrete compressive strength ( f c ), steel yield strength ( f y ), and elastic modulus ( E s ). The shear modulus of neoprene bearings ( G neo ), commonly used in Italian bridges, was also treated as a random variable based on ranges from Cardone, 2014. Corrosion reduces the effective area and degrades mechanical properties, as described in Di Mucci et al. (2025). Residual steel properties were estimated using models by Du et al. (2005), with ultimate strain values informed by Opabola (2022), to simulate advanced degradation. Corroded transverse reinforcement further compromises confinement, reducing the strength and ductility of confined concrete. This was modeled using a corrosion-adjusted Mander formulation (Mander et al., 1988).

3.2. Numerical Model

The numerical modelling strategy developed in this study focuses on capturing the seismic response of RC bridge piers under variable corrosion conditions (Di Mucci et al., 2025) using OpenSees (McKenna et al., 2000). A fiber based multi-degree-of-freedom (MDoF) model was implemented in OpenSees, where the pier shaft is represented