PSI - Issue 78

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

2

can mask poor performance on the damaged class, as shown for usability-tag prediction after quakes [29, 2]. Parallel e ff orts also address social and economic consequences [11, 9, 35, 27, 12, 14, 5, 36]. Using nearly 300 non-code-conforming reinforced-concrete and masonry buildings assessed under the Italian Seis mic Code (Province of L’Aquila), the authors train and compare linear / logistic regressions with advanced ML algo rithms. The dataset contains about 15 categorical and numerical features obtained from surveys, tests, and numerical analyses. Results reveal that balanced-threshold artificial neural networks attain prediction accuracies above 85%, and SHAP values clarify each variable’s influence. The proposed framework therefore provides decision-makers with a fast, transparent aid for allocating seismic-mitigation resources where they are most needed.

1. Problem formulation

The authors try to predict the seismic vulnerability index of almost 300 Italian buildings. Each index follows the Italian Seismic Code and is described by 15 numerical or categorical features. They built both regression and classifi cation models to see whether these features are enough to reproduce the engineers’ results. The seismic vulnerability index α is the ratio between structural capacity C anddemand D :

C (⃗ x ,⃗ Θ ) D (⃗ x ,⃗ Θ )

(1)

.

α =

Here⃗ x holds deterministic parameters, and⃗ Θ covers uncertain ones. Engineers estimate C and D with a procedure set by the Italian code. Their choices about modelling, material tests and “confidence factors” introduce subjectivity into α . Each survey gathers:

• geometry, location and visible damage; • construction history; • lateral-load resisting elements and their connections; • material condition and mechanical tests; • subsoil and foundation information.

All buildings reach knowledge level LC2, so the confidence factor on material properties is 1.2 [15, 13]. Using commercial finite-element software, the engineers create the elastic model; Then they derive the site-specific design spectrum from the code, run linear dynamic (response-spectrum) analysis and compute capacities of members and the overall building. Finally, they obtain α with Eq. (1). The authors ask whether α (or a class of α ) can be predicted directly from the 15 features. They fit four models:(i) linear regression [24], (ii) decision-tree regression [8], (iii) support-vector regression (SVR) [30], (iv) artificial neural network (ANN) [19]. Each model estimates

α = f (⃗ x ,⃗ Θ ) + ϵ,

(2)

where ϵ is the residual error. Because exact α values are uncertain, the authors also group them into k classes C 1 , . . . , C k . Given features⃗ x , the predicted class is

ˆ C = argmax C i

f (⃗ x ) .

(3)