Bayesian updating of seismic fragility curves of the Irpinia 1980 building stock with machine-learning-supported data analysis

Abstract Reliable seismic fragility assessments are essential for risk application, however, within empirical framework it is hard challenging due to the reliability of exposure information and to the need to account explicitly for uncertainty in ground motion intensity. This study develops an integrated method combining machine-learning supported data completion with Bayesian probabilistic methods to derive fragility functions from the database collected after the 1980 Irpinia earthquake and available through Da.D.O. platform. The first methodological innovation concerns the reconstruction of missing construction age information. A supervised machine-learning classifier is trained using the available structural, geometrical, spatial, and damage-related building attributes. The interpretation of the results shows that the reveals that the most influential predictors are governed primarily by physically meaningful variables. Fragility curves are subsequently derived within a Bayesian framework using Markov Chain Monte Carlo (MCMC) methods, allowing for the explicit treatment of uncertainty in both model parameters and seismic demand. A key aspect of the proposed approach is the incorporation of uncertainty in Shakemap-derived intensity measures, modelled as a spatially correlated latent field. Comparisons with models neglecting intensity-measure uncertainty show that its omission may bias the estimated fragility parameters, whereas its inclusion provides more stable and physically consistent estimates. A further contribution is the comparison of alternative intensity measures, namely PGA, PGV, and Sa(0.3s). Their efficiency is evaluated through both Bayesian likelihood-based comparisons and a residual-based approach, showing that PGV generally provides the strongest support for the observed damage explanation, especially for moderate-to-high damage states. The convergence of the two methods therefore strengthens the robustness of the conclusions and provides independent evidence that PGV generally offers the most effective representation of the observed damage patterns within the Irpinia dataset. Unlike approaches based exclusively on broad vulnerability classes, the proposed framework explicitly considers building-specific attributes reported in the inspection forms, including structural typology, horizontal system, and construction period. The resulting fragility relationships also provide a consistent vulnerability ranking among rubble-stone, tuff, brick masonry, and reinforced-concrete buildings. Finally, comparison with fragility models derived from the 2009 L’Aquila earthquake highlights the possible influence of regional construction practices and the evolution of seismic classification on empirical fragility assessment.