PSI - Issue 78

Marco Martino Rosso et al. / Procedia Structural Integrity 78 (2026) 1382–1388

1383

1. Introduction The evaluation of building damage and usability under earthquake conditions represents a manifold activity of critical importance both for risk management, earthquake engineering, and structural engineering purposes. At least two different strategies can be identified considering different time scales. Indeed, in a preventive perspective, territorial-scale vulnerability assessment enables the development of a knowledge model about the built environment. This could potentially provide its predisposition to high seismic risk levels, thus ensuring the possible application of mid- and long-term seismic risk mitigation strategies on a wide regional scale. Indeed, this may allow conducting cost-benefit analyses to finance retrofitting programs for vulnerable structures, whilst optimizing and prioritizing the most important or most vulnerable ones. From a completely mirrored perspective, there is post disaster building damage and usability evaluation, which instead require rapid evaluation for prioritizing and timely activating first aid and coordination of emergency services on a regional scale on one side, rapidly taking immediate life-safety decisions regarding building occupancy and/or coordination of emergency response and temporary housing solutions, and also for defining the disaster volume, especially to quantify the reconstruction costs on the entire regional scale for the subsequent recovery planning. Various methodologies have been employed for large-scale damage and usability assessment, each with distinct advantages and limitations. Mechanics-based approaches rely on structural modeling of the building behavior, but this requires a high knowledge level of the built environment, which is prohibitive for large-scale evaluation. On the other hand, data-driven strategies make use of historical statistics of earthquake occurrences in the area under investigation and pattern recognition to extrapolate information valid for the territorial scale. In recent years, data driven approaches have gained popularity since their promising and attractive solutions, entirely driven by extracted information hidden in meaningful data. For instance, Zucconi et al. (2017) developed a linear regression model for masonry buildings after the L'Aquila earthquake, Bertelli et al. (2018) introduced nonlinear regression to derive usability-based fragility curves, and Tocchi et al. (2023) applied ML to assess building usability post-L'Aquila earthquake. This study mainly focused on this second perspective of rapid post-disaster building usability assessment, illustrating an Artificial Intelligence (AI) supported strategy for building usability classification in several municipalities of the Abruzzi Region (Italy) after the 2016 – 2017 Central Italy earthquakes. Besides the disaster that occurred, for scientific studies, studying these kinds of events represents an important opportunity to learn from real-world data how the built environment reacts to certain input earthquake events. In particular, the 2016-2017 seismic sequence provides an exceptional testbed for damage assessment methodologies since multiple high-magnitude sequences of events occurred over a period of more or less 18 months, activating complex fault mechanisms and aftershock interactions. Moreover, in the struck Central Italy area, the building stock presents a mix of historic masonry and modern construction, with a relatively dispersed situation in terms of code compliance, construction quality, and seismic vulnerability. All of these aspects underline the importance of continuing to study this kind of real-world data to further inspect possible strategies for rapid and reliable building usability assessment, and even to evaluate the possibility of transferring the acquired knowledge for promoting effective and sustainable predictive maintenance programs, attempting to avoid, or at least contain, future disasters. The following Section 2 illustrates the current available dataset and the Machine Learning (ML) classification strategy herein adopted. In Section 3, the results of the optimized ML classifier have been reported, and the adopted interpretability tools are discussed. 2. Machine Learning Classifiers for Building Usability Class The currently available dataset comprises 12,662 buildings in the Abruzzi region (Italy) struck by the 2016-2017 Central Italy earthquake event, surveyed using the Rapid Post-Earthquake Damage Evaluation (AeDES) forms, i.e., Italy’s standard post -earthquake assessment tool organized into nine sections across four pages. These forms document usability ratings and damage severity, aiding reimbursement and recovery planning. Specifically, the first three sections of AeDES forms refer to the building description data, i.e. containing information about geographic location and positioning, if isolated, edge, or being part of a building aggregate, structural characteristics, such as number of floors above the ground, inter-story height, floor area, construction era,