Training and validation of a machine learning model for seismic vulnerability of masonry buildings in aggregate and selection of key parameters

The seismic vulnerability assessment of masonry buildings in aggregate, typical of historical centers, remains a critical and widely debated topic. The complexity of this issue, driven by numerous parameters and uncertainties, has left many challenges unresolved by traditional analysis methods. This study addresses the problem through an innovative approach, leveraging state-of-the-art Machine Learning (ML) techniques to analyze the small historical center of Casentino (AQ), significantly affected by the 2009 L'Aquila earthquake. The availability of highly detailed data on structural characteristics and damage enables the identification of correlations between multiple vulnerability parameters and in plane and out-of-plane damage mechanisms. From these correlations, a subset of key factors was identified as most relevant for interpreting seismic vulnerability while being easily obtainable through in-situ surveys, offering an efficient basis for large-scale risk assessments. The robustness of this subset was validated on a second case study, Visso (MC), by comparing ML-predicted damage to observed damage following the 2016–2017 Central Italy earthquake. The results highlight the potential of ML techniques to enhance seismic risk evaluation and streamline data collection processes for broader applications.