PSI - Issue 78

Livio Pedone et al. / Procedia Structural Integrity 78 (2026) 1609–1616

1615

work, only the IS-V index is considered. Capacity / demand ratios are evaluated by applying the Capacity Spectrum Method (CSM; ATC (1996)). For the post-earthquake scenario, the analysis is repeated for each considered realization (Fig. 6a). Results in terms of safety assessment are summarized in Fig. 6b.

(a)

(b)

Fig. 6: (a) Seismic response analysis through the capacity spectrum method for each realization of the structure in its damaged configuration; (b) results in terms of “safety index” for the intact and damaged structures, considering both analytical and numerical simulations.

As expected, earthquake-related damage to the structural components leads to a reduction of the safety index for both analytical and numerical simulations. Focusing on the numerical results, the IS-V index decreases from 0.61 (intact) to 0.52 (damaged). Overall, a slight IS-V reduction is observed, consistent with the slight-to-moderate damage sustained by the structure after the damaging earthquake scenario. A good agreement is observed between numerical analysis and the SLaMA method for the intact configuration (10% relative error), while a higher error (almost equal to 15%) is obtained for the damaged configuration, suggesting the need for further refinement of the SLaMA-Damaged approach. Finally, Fig. 6b also shows the dispersion ( ± 3 standard deviations) due to uncertainties in the CNN-damage classification. It is interesting to note that these uncertainties do not significantly a ff ect seismic safety evaluation. This result suggests that the proposed procedure may be suitable for a preliminary estimation of the safety level of earthquake-damaged buildings during early emergency phases. This paper has presented a machine learning (ML)-based framework for post-earthquake seismic residual capacity assessment of damaged buildings. In line with state-of-the-art procedure in the literature, the proposed methodology relies on nonlinear static analyses and capacity reduction factors for damaged structural components. A Convolutional Neural Networks (CNN) model, based on a VGG16 architecture, is used to automatically evaluate the damage severity from images. This ML model has been trained on 5,000 RGB images sourced from existing databases and re-labelled by experts. This information is then used to evaluate suitable reduction factors for damaged components in terms of sti ff ness, strength, and ductility. A simplified analytical / mechanical procedure is employed to evaluate the seismic capacity of the structure in its pre- (i.e., intact) and post- (i.e., damaged) earthquake configuration. The proposed framework has been implemented for a case-study reinforced concrete structure. The preliminary results confirmed the feasibility of the proposed approach. The proposed methodology may support the development of a rapid assessment tool for the emergency phase, allowing for a visual-based mechanically-informed safety evaluation of buildings in post-earthquake scenarios from the earliest emergency phases. 5. Conclusions

Acknowledgements

This research is supported by the PNRR - National Research Centre CN1 on “High-Performance Computing, Big Data and Quantum Computing”, Spoke 5 “Environment and Natural Disaster”. The authors acknowledge the support of CN1-Spoke 5 for funding Simone Saquella and Michele Matteoni (PhD). Livio Pedone received fund ing (RTD-A) from the project “Centro di ricerca per l’innovazione sull’economia circolare e sulla salute” (CUP: