Machine learning surrogates for predicting and interpreting flood fragility parameters of masonry walls

Abstract Flood-induced out-of-plane collapse of masonry walls is a recurrent cause of severe building damage and loss during flood events, particularly in low-rise residential construction. Assessing this failure mode at scale remains challenging: simplified fragility models are fast but often lack physical transparency, while mechanics-based analyses are interpretable but computationally expensive. This study develops a mechanics-informed surrogate modelling framework that bridges this gap. Machine-learning surrogates, including linear sparse regressions, kernel-based methods, and ensemble models, are trained on a constrained synthetic dataset generated from a mechanics-based wall collapse formulation. This enables fragility parameters to be predicted with accuracy comparable to the reference simulator while retaining a direct link to the governing structural mechanisms within the adopted modelling domain. A central contribution of the work is the explicit integration of explainable artificial intelligence, allowing the role of geometry, openings, vertical connectivity, and boundary restraint to be quantified and interpreted in physically meaningful terms. The results reveal stable and intuitive collapse drivers across boundary configurations, with the upper-floor contribution, thickness-related descriptors, and boundary restraint forming the dominant sensitivity group, while masonry density and opening-related variables have secondary or model-dependent influence. The consistency of these trends across linear and nonlinear surrogate models demonstrates that the framework captures not only fragility estimates but also mechanically consistent sensitivity patterns of wall overturning. By combining mechanics-based simulations with surrogate modelling, the proposed approach enables efficient and physically interpretable assessment of masonry wall fragility.