PSI - Issue 78

Franco Ciminelli et al. / Procedia Structural Integrity 78 (2026) 921–928

922

1. Introduction Italian road infrastructures — particularly bridges and viaducts — were largely built between the 1950s and 1970s, during the country’s post -war economic boom and engineering innovation. These structures were designed under technical and regulatory frameworks that differ significantly from current standards. Although advanced for their time, many of them did not fully consider seismic and hydrogeological risks, whose recognition and mitigation have evolved only in recent decades. Today, many of these bridges are operating beyond their intended service life and are subject to demands exceeding their original design loads, mainly due to increased traffic and the extreme effects of climate change (Nasr et al., 2021). Seismic risk, in particular, remains a critical concern for structures located in high-hazard areas, as shown by several international studies (Padgett & DesRoches, 2007; Pinto & Franchin, 2010). Even moderate earthquakes can compromise the stability and functionality of bridges (Marshall & Coulston, 2013), especially in the presence of pre-existing defects, material degradation, or outdated design standards (Zucca et al., 2020). In recent decades, various approaches have been developed to estimate the seismic fragility of road infrastructure (Muntasir Billah & Shahria Alam, 2014) aiming to assess the probability of structural damage based on ground motion intensity and structural characteristics. These include fragility curve methods (Hwang et al., 2001), probabilistic models (Hawk & Small, 1998; Thompson et al., 1998), numerical analyses, and more recently, data-driven techniques (Alogdianakis et al., 2022; Pallante et al., 2024). Despite extensive research, most of these methods focus on specific risk scenarios and do not fully integrate the combination of factors that contribute to overall seismic risk (Santarsiero et al., 2021). In Italy, the lack of a unified management system — tragically highlighted by the collapse of the Morandi Bridge in Genoa in 2018, which resulted in 43 fatalities — revealed the consequences of neglect and ineffective maintenance strategies (Beltrametti et al., n.d.; Calvi et al., 2019). Similar incidents, such as the collapses of the Albiano Magra bridge and the A6 motorway Madonna del Monte viaduct, have further confirmed the vulnerability of the national infrastructure network (Malerba, 2024). In response, the “Guidelines for the classification and management of risk, safety assessment, and monitoring of existing bridges” (GL) were issued by Italian institutions (ANSFISA, 2022; MIT & CSLLPP, 2022), now serving as the national reference framework. These Guidelines adopt a multi-level and multi-risk approach, introducing the Class of Attention (CoA) as a key indicator to support the prioritization of inspections and interventions. However, the operational application of the GL presents significant challenges, particularly in the transition from initial assessment levels (0 and 1, involving census and visual inspection to estimate the Level of Defectiveness — LoD) to more detailed evaluations. The estimation of LoD itself often proves burdensome both economically and logistically and is strongly influenced by inspector subjectivity and experience (Abdallah et al., 2022). In this context, the present study proposes advanced tools for the seismic risk management of existing bridges by leveraging Artificial Intelligence and, specifically, Automated Machine Learning (AutoML). The proposed framework is fully aligned with the GL methodology and focuses on the initial levels of analysis (0, 1, and 2). The main goal is to accurately predict the seismic Class of Attention (CoA-S) using only information available from census data or technical documentation. A similar approach has been developed for predicting structural and foundational CoA and structural LoD using Artificial Neural Networks (ANN) (Principi et al., 2025). It is important to stress that inspections are not eliminated, as they remain essential for detailed structural assessments. Instead, the proposed tool aims to use the predicted CoA-S to support the prioritization of inspections, enabling a more efficient allocation of resources and a preliminary estimation of the management costs associated with follow-up interventions outlined in higher levels of the GL. Furthermore, the method can be used as a rapid tool for verifying already classified bridges and viaducts. The methodology, detailed in Chapter 2, includes the preparation of synthetic datasets using rule-based algorithms, sampling techniques, and feature selection procedures. Predictive models were trained using AutoGluon-Tabular (Erickson et al., 2020) on synthetic data, and later experimentally validated on a real-world dataset comprising 100 bridges from the ANSFISA database. A learning curve analysis was also performed to identify the optimal dataset size for achieving the best trade-off between accuracy and generalization. Chapter 3 presents the results, focusing on the relative importance of predictive features and the confusion matrices derived from model training. Finally, Chapter 4 provides a concluding discussion on the strengths, limitations, and future development opportunities of the proposed framework, highlighting its potential for extension to other infrastructure risk contexts.