PSI - Issue 62

Fabio Parisi et al. / Procedia Structural Integrity 62 (2024) 701–709 F. Parisi et al. / Structural Integrity Procedia -- (2024) _ – _ 3 RTs in the prediction task (Breiman, 2001) by building different regression trees and averaging the results. 2.1. Methodology We first defined the seismic demand dataset by performing a consistent number of NLTHA carried out on a population of numerical models of the investigated bridge subjected to a preliminarily selected ground-motion suite. Then, a representative portion of the dataset is selected, and the informative content of the features is investigated to identify feature importance; a candidate group of features is identified and selected to train the RF regressor. Note that represent the dataset used for the training: it is quite smaller than , representing the potential of the investigation. An RF is trained on and used generate data configuring a new ′ dataset, equal to D in terms of instances numerosity. To conclude, both and ′ are then used to fit a power-law relationships consistently with the procedure outlined in (Nettis et al., 2021a). The potential of the methodology is then tested by comparing the power-law relationship obtained with the two different approaches. 3. Case-study bridge In this section, the case study bridge is described, together with the knowledge-based uncertainties and the ground motion suite. The modelling and analysis assumptions used to generate the seismic demand database to train the adopted ML algorithms are reported. 3.1. Description of the case-study bridge, knowledge-based uncertainty, ground motion selection and intensity measure parameters The presented methodology is tested for a multi-span RC bridge located in Southern Italy built between 1980-1984 and designed according to old Italian reference codes and shown in Figure 1. The case-study bridge exhibits an isostatic structural scheme with five 30 m-long spans. The superstructure is composed of three prestressed girders simply supported on single-column piers and abutments. The piers comprise a RC pier cap and a column having a circular cross-section measuring a diameter of 2.60 m. The height of the piers, measured from the ground level to the top of the pier cap, ranges between 8.60 and 15.60 m. The girder-pier connections are composed of fixed bearings and sliders which are fixed in the transverse direction and, therefore, transfer seismic inertia forces between the superstructure and the substructure components. Further information on bridge geometry is reported in (Nettis et al., 2020). Since blueprints or design documents are unavailable, reinforcement details and mechanical parameters are modelled as knowledge-based uncertainties through the statistical distributions reported in Table 1. The adopted statistical distributions are derived from literature studies (Celik and Ellingwood, 2010). A uniform discrete distribution is used to model the variability in the design strength of concrete and steel , subsequently used to perform a simulated design considering the old italian design reference codes and to derive the longitudinal reinforcement in the RC piers. The strength values for concrete and steel ( and ) are calculated depending on the corresponding design characteristic values assuming a normal distribution in which the 5 th fractile is equal to and and the coefficient of variation is equal to 0.09 and 0.18 for steel and concrete, respectively (Monteiro, 2016). The volumetric ratio of the transverse reinforcement ( ) is modelled through a discrete uniform distribution. 703

Figure 1. Geometry of the case-study bridge.

Table 1. Statistical distribution representing knowledge-based uncertainty.