PSI - Issue 78

Angelo Aloisio et al. / Procedia Structural Integrity 78 (2026) 1–8

6

+0.92

Concrete strength [ MPa ]

Volume [ m 3 ]

+0.76

+0.65

Construction year

+0.62

Height [ m ]

Surface [ m 2 ]

+0.59

+0.54

Peak Ground Acceleration [ g ]

+0.43

Number of elevated floors

+0.31

Structural typology

+0.28

Pilotis floor

+0.25

Structural configuration

+0.21

Soil type

+0.19

Number of residential units

+0.08

Basement floors

+0.08

Damage

+0

Structural interventions

0 . 0

0 . 2

0 . 4

0 . 6

0 . 8

1 . 0

mean(—SHAP value—)

Fig. 3. SHAP interpretation of the ANN. Global feature importance.

4. Conclusions

This work explored whether data–driven techniques can replace the standard, code-based procedure for estimating seismic vulnerability. A catalogue of roughly 300 reinforced-concrete and masonry buildings supplied by ATER (Province of L’Aquila, Italy) served as the test bed. Each entry already possessed a vulnerability index, defined as the capacity-to-demand ratio, obtained through the full knowledge-acquisition and structural-analysis route prescribed by the Italian Seismic Code. Neither regression nor multi-class classifiers reproduced those indices with useful accuracy. By contrast, framing the task as a binary decision, Severe versus Moderate, split at a chosen threshold, proved fruitful. Among the four algorithms examined, an artificial neural network (ANN) outperformed logistic regression, a decision tree and XG Boost, especially when the classes were imbalanced. With the threshold set to α t = 0 . 20 (which yields near-balanced classes) the ANN reached about 85 % overall accuracy and kept recall above 60 % for the minority class; the other models often let minority-class recall fall below 20 %. SHAP analysis showed that concrete strength, global geometry (volume, height, floor area) and construction year carry the greatest weight. Material quality thus matters more than the broad structural category (RC or masonry). Site parameters such as soil class and peak ground acceleration entered with lower importance, probably because their range within the sample is narrow. A reliable binary classifier lets authorities flag the most critical buildings quickly, redirecting resources from costly full assessments to actual retrofitting. Because the target index is code-based, the output is directly tied to regulatory requirements, an advantage in countries where legal liabilities follow the code.