PSI - Issue 78

Lorenzo Principi et al. / Procedia Structural Integrity 78 (2026) 1681–1688

1682

1. Introduction Bridges are key components of transportation infrastructure, directly influencing public safety, economic development, and territorial connectivity (Huang et al., 2013). Recent bridges failures (Wardhana et al., 2003, and Zhang et al., 2022) have highlighted their vulnerability to both anthropic and natural hazards and the necessity of optimized maintenance and risk mitigating actions. Several studies have introduced methodologies and guidelines addressing various hazard types, including environmental degradation and traffic loads (Nettis et al., 2024), seismic actions (Tubaldi et al., 2021, Minnucci et al., 2022, Ruggieri et al., 2024, Scozzese and Minnucci, 2024), floodings (Durand et al., 2018, Cerema, 2019, Scozzese et al., 2023), and landslides (Bobrowsky and Highland, 2013, AGS, 2007, Fell et al., 2008). Nevertheless, most studies remain focused on single - hazard scenarios. To fill this gap, the Italian Ministry of Transportation (MIT) developed the national Italian Guidelines (hereafter referred to as IG), for the risk classification and management of existing bridges (MIT, 2020, ANSFISA, 2022, Natali et al., 2023), introducing a multi - hazard methodology that addresses structural, seismic, flood, and landslide risks. A critical point in the approach is the transition from the initial level, usually consisting in collecting census data (denoted as Level 0 in MIT, 2020), to more advanced evaluation stages (Levels 2 to 5, MIT, 2020). This transition consists of time consuming inspections, including detailed surveys of structural components and surrounding areas (Level 1 MIT, 2020). These inspections remain a critical step, as they can represent a bottleneck for large bridge stocks (Yaw Adu Gyamfi et al., 2016). In the IG procedure, inspections are necessary to compute the seismic risk, as it represents a key vulnerability parameter. Recent studies have explored machine learning (ML) techniques for rapid bridge assessments (Yaw Adu - Gyamfi et al., 2016, Assaad and El - Adaway, 2020, Xia et al., 2021, Alogdianakis et al., 2022, Ruggieri et al., 2023). Principi et al. (2025) proposed a framework based on Artificial Neural Networks (ANNs) to estimate both the Class of Damage (CD), which reflects the bridge’s degradation level according to MIT (2020), and the Class of Structural Risk (CSR), associated with traffic loads. The approach was successfully applied to the Italian highway network using data from a pilot IG - based assessments (Natali et al., 2023). Building on this framework, this study extends the use of ANNs to predict the Class of Earthquake Risk (CER) (MIT, 2020). The proposed model introduces seismic hazard data and bridge vulnerability and exposure information, enabling regional - level assessments. An optimized ANN represents the framework output. This ANN is used to a case study of 95 bridges in the Marche Region to predict the CER. Results are also presented through a GIS - based map, providing an efficient tool for seismic risk analysis at territorial scale. 2. Framework and methodology The present study builds on the novel framework proposed by Principi et al. (2025) (Figure 1) for predicting the CER. The framework is organized into three main phases: I) Data Selection and Data Preprocessing, II) Data Processing, and III) Optimal Model and Performance Analysis, each comprising multiple steps (Figure 1).

Figure 1. General framework.