PSI - Issue 78

Available online at www.sciencedirect.com

ScienceDirect

Procedia Structural Integrity 78 (2026) 1–8

XX ANIDIS Conference Fast-Track Code-Based Seismic Vulnerability Screening with Machine Learning: Evidence from 300 Italian Buildings Angelo Aloisio a, ∗ , Francesco Irti a , Marco Martino Rosso b , Giuseppe Quaranta c , Cristoforo Demartino d , Massimo Fragiacomo a a Department of Civil, Construction-Architectural and Environmental Engineering, University of L’Aquila, L’Aquila, Italy

b Department of Structural, Geotechnical and Building Engineering, Politecnico di Torino, Turin, Italy c Department of Structural and Geotechnical Engineering, Sapienza University of Rome, Rome, Italy d Department of Architecture, Roma Tre University, Rome, Italy.

© 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of XX ANIDIS Conference organizers Keywords: Seismic vulnerability index; Data-driven model; Italian Seismic Code; Artificial Neural Networks; Binary classification In seismic engineering, the vulnerability index condenses the likelihood of a structure sustaining damage during an earthquake, guiding retrofit priorities. Large-scale assessments usually rely on three families of methods: (i) damage probability matrices, (ii) vulnerability indices, and (iii) capacity-curve procedures, that can be converted into fragility curves [4, 16, 6]. They may link either seismic-hazard intensity measures or physical building attributes (height, age, materials, structural type, workmanship) to expected loss [22, 23]. When a mechanics-based route is chosen, engineers progress from knowledge acquisition through simplified or de tailed structural modelling to an analysis whose output is the ratio of capacity to demand. Macro-models for masonry, for instance, keep computations manageable for portfolio studies [7, 17, 31, 21, 26]. Yet every such index remains only semi-physical, because it inherits modelling assumptions and uncertainties [34]. Machine-learning (ML) o ff ers a shortcut: once a trustworthy training set exists, a classifier or regressor can predict vulnerability directly from readily available descriptors, bypassing time-consuming analyses [32, 28, 33]. Recent work has trained models on post-event damage surveys or numerically generated data [3, 1]. Nevertheless, class imbalance Abstract The study introduces a data-driven method for estimating a code-based seismic vulnerability index for roughly 300 Italian buildings. Detailed surveys, experimental tests, and numerical analyses produced about 15 mixed-type predictors. These inputs fed several predictive algorithms, in particular logistic regression and an artificial neural network (ANN). After rebalancing the classes by adjusting the vulnerability cut-o ff , the ANN classifies buildings into two risk groups with better than 85% accuracy. SHAP analysis reveals how much each feature influences the prediction, providing a transparent tool to help authorities prioritise seismic-risk mitigation.

∗ Corresponding author: angelo.aloisio1@univaq.it

2452-3216 © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of XX ANIDIS Conference organizers 10.1016/j.prostr.2025.12.001