PSI - Issue 78

Nicola Di Battista et al. / Procedia Structural Integrity 78 (2026) 412–417

413

In a probabilistic framework, SRA combines three fundamental constituents— hazard , vulnerability , and expo sure —through a convolution integral that yields annualised loss metrics. Probabilistic seismic-hazard analysis (PSHA) remains the prevailing tool for quantifying the first term, although deterministic scenario approaches are increasingly employed for region-specific studies [6, 24]. Vulnerability encapsulates the propensity of the built environment to sustain damage under ground–motion demand [24]. It is typically expressed through fragility functions that link an intensity measure to the probability of exceed ing discrete damage states. These functions may be derived (i) analytically, by subjecting numerical archetypes to nonlinear time–history analysis, or (ii) empirically, by interrogating post-event damage databases. Exposure, the third ingredient, monetises the built inventory through socio-economic and demographic data, enabling the conversion of physical damage into direct and indirect losses [24]. Beyond safeguarding life safety, contemporary SRA must provide credible estimates of economic conse quences—repair costs, downtime, and reconstruction outlays—that can disrupt public finances and regional socio economic equilibrium [19]. Accurate loss forecasts are invaluable to governments (resource allocation), insurers (pre mium calibration), and planners (risk-mitigation prioritisation). Post-earthquake observations have traditionally been mined to refine structural fragilities; comparatively few in vestigations have focused on models that predict reconstruction costs [18, 11, 13]. Studies exploiting data from the 2009 L’Aquila (Italy) event revealed limitations when PHSA-based cost models were applied to diverse construction typologies [10, 12]. FEMA P-58, while comprehensive, was conceived for U.S. construction practice; Italian applica tions therefore substituted non-structural fragilities and consequence functions with U.S. proxies [1, 22]. Subsequent e ff orts have begun tailoring these functions to Mediterranean building stocks [8, 23]. Parallel to these refinements, data-driven paradigms—especially machine-learning (ML) algorithms—have gained traction for predicting damage metrics, usability classifications, and other consequence measures [21, 4, 5, 3]. ML’s capacity to capture complex, nonlinear relationships suggests strong potential for cost-prediction tasks, yet, to date, no study has harnessed the L’Aquila post-event dataset to develop ML models for masonry-building reconstruction costs. This work fills that gap. Leveraging records from the 2009 L’Aquila reconstruction-grant programme, the authors benchmark several ML techniques for predicting total reconstruction costs of masonry buildings. The adopted feature set comprises vulnerability-related indices consistent with the grant-allocation procedure. Following Fung et al. [16], the dependent variable encompasses structural repair, ancillary direct costs (e.g. permits), and indirect expenditures such as temporary relocation. The working dataset contains 2 230 masonry buildings located in eleven municipalities that fall inside the so-called Municipalities of the Crater (MIC), the zone accorded highest priority for public reconstruction funding after the 6 April 2009 event. Each record stores: • the nine vulnerability attributes used by the Italian post-event procedure to compute a global vulnerability index (0–36); • the building’s usability rating derived from the AeDES rapid-assessment form (classes A, B, C, E); • the peak ground acceleration (PGA) expected at the municipality centroid, estimated with regional attenuation laws [20]; • the plan area [m 2 ]; • the base reconstruction grant authorised for that building. Figure 1 provides a visual synopsis: (a) spatial distribution of PGA; (b) breakdown of AeDES usability classes; (c) histogram of the global vulnerability index; (d) frequency of authorised grant amounts ( € / m 2 ). The Mw 6.3 main shock struck the Abruzzo region at 03:32 AM local time, claiming roughly 300 lives, injuring more than 1 500 people, and displacing about 70 000 residents [2, 14]. Early macro-economic appraisals placed direct losses near euro 10 billion [14], underscoring the fiscal magnitude of the reconstruction e ff ort that followed. Immediately after the earthquake, trained engineers performed “comb” street-level inspections and filled out the nine section AeDES form. Besides assigning a usability class, inspectors evaluated five EMS-98 damage grades (D 1 –D 5 ). Concurrently, they scored nine structural attributes to obtain a categorical vulnerability level: V 1 (low), V 2 (medium) 2. Dataset derived from the 2009 L’Aquila earthquake