Data-driven seismic fragility and risk evaluation of existing reinforced-concrete bridge portfolios under limited data

Abstract

Preliminary screening of existing transport infrastructure networks is essential to support timely maintenance and retrofit decisions, improve structural resilience, and reduce the economic and environmental impacts of severe damage, failure, and inefficient resource allocation. This study investigates a machine-learning-based pathway for quantitative, simplified seismic risk screening of typologically homogeneous reinforced-concrete girder-bridge portfolios under limited data, in accordance with the Italian Guidelines for existing bridges. Fragility labels are derived through a streamlined cloud-analysis benchmark for 75 simply supported reinforced-concrete lattice-girder bridges, yielding the lognormal fragility parameters: median intensity measure μ , logarithmic dispersion β , and ultimate displacement capacity d u . Supervised regression models are trained to map routinely available inventory descriptors, including geometric and mechanical proxies, to ( μ , β , d u ) . Among the tested models, Gaussian Process Regression provides the best balance between accuracy and stability in the small-sample regime, with cross-validated R 2 values close to 0.9 for all targets. The selected surrogate is applied to 10 additional bridge configurations to assess predictive capability on new cases. Finally, inferred fragility is coupled with site-specific seismic hazard to compute the mean annual frequency of exceedance λ , used to rank 25 bridges of the A19 motorway network in Sicily. Results show prioritization closely consistent with the cloud-analysis benchmark, while requiring substantially lower computational effort. Compared with qualitative guideline-based classes of attention, λ offers useful intra-class granularity for transparent, scalable prioritization. Limitations include reduced accuracy outside the training manifold, typology-specific validation, and lack of explicit condition and defectiveness features.

The authors' abstract, as published at the source. Engineering Applications of Artificial Intelligence, 2026 · DOI ↗