Machine learning model for predicting post-earthquake reconstruction grants of masonry buildings from seismic vulnerability features

In the aftermath of the 2009 L’Aquila earthquake in Italy, public authorities established a procedure to determine reconstruction grants for buildings on the basis of evidence from visual inspections, as documented in a dedicated form (i.e., the AeDES form). The calculation is driven by two main aspects, namely the observed damage and the assessed vulnerability level. For masonry buildings, which account for more than 90% of structures in small urban Italian centres, a set of nine features was selected to represent the key factors governing the overall vulnerability assessment. The base value established through this procedure is next subjected to a series of adjustments that may increase the initial estimate to nearly twice its original amount. Overall, this procedure imposes significant uncertainties and difficulties on territorial authorities in both planning and allocating the required financial resources. This study demonstrates that machine learning algorithms can accurately predict the final reconstruction grant category for masonry buildings deemed unusable after an earthquake (per the AeDES form), relying only on their vulnerability characteristics and a measure of the seismic ground motion intensity. Using real data collected after the 2009 L’Aquila earthquake, the predictive performance of various machine learning algorithms is evaluated through two alternative classification approaches (i.e., multiclass and cascade classification approach). The results demonstrate that a gradient boosting model can provide a satisfactory approximation of the reconstruction grant, and a feature importance analysis is also performed to identify the inputs that most strongly influence the model’s predictions. • ML for post-earthquake reconstruction grant for masonry buildings at regional scale. • Inputs limited to vulnerability indicators and scalar ground-motion intensity measure. • Hyperparameter-tuned classifiers benchmarked across multiple learning algorithms. • Model-agnostic uncertainty and misclassification patterns quantified by grant class. • Decision-support workflow for rapid, consistent public-aid screening and risk management.