Bridging Data Gaps in Structural Fragility Modeling through Transfer Learning: Methodology and Case Studies

A new methodology-centered framework applies transfer learning to structural fragility modeling, addressing domain shift, class imbalance, and scarce target labels. The approach, detailed in a paper posted to arXiv, demonstrates four adaptation strategies across three case studies involving hurricanes and an earthquake [1]. The paper, authored by Narges Saeednejad, presents a framework designed to preserve engineering interpretability while supporting decision-making under uncertainty [1]. It tests four transfer learning strategies: instance-based, parameter-based, hierarchical Bayesian, and multi-source [1]. Each strategy is evaluated through a distinct case study. Instance-based transfer learning via importance weighting is demonstrated on coastal bridge fragility using observations from Hurricane Katrina [1]. Parameter-based transfer learning, combined with hierarchical Bayesian transfer learning, is applied to residential building fragility using data from Hurricane Ian. That combination enables partial pooling across strata and posterior uncertainty quantification [1]. The third case study uses multi-source transfer learning, which fuses multiple analytical fragility models with learned source weights and regularized target-domain adaptation, to examine seismic bridge fragility using observations from the 2001 Nisqually earthquake [1]. A key finding is that directly transferring existing state-of-the-art source models fails under domain shift and severe class imbalance [1]. Targeted adaptation, however, substantially improves failure detection and predictive stability in low-data regimes [1]. The authors argue this underscores the need for systematic guidance on diagnostics, strategy selection, and uncertainty reporting when developing and adapting fragility models [1]. Related work highlights the broader challenge of unifying disparate fragility estimates. A 2026 NIST study by Hariri Ardebili and Sattar introduced a framework for combining multi-source and multi-fidelity fragility functions, noting that diverse numerical models with varying fidelity produce divergent outcomes and that methodologies are needed to synthesize these into a cohesive risk evaluation [4]. That study evaluated approaches including decision tree selection, uniform and non-uniform weighting, and variance-bias decomposition, finding substantial variability in outcomes based on data accessibility and model fidelity [4]. A separate 2026 arXiv paper proposed an online Bayesian learning framework for fragility modeling that continually refines physics-based predictions as new observations become available [3]. That framework uses a two-stage process: a local updating stage that assimilates heterogeneous evidence to correct physics-based fragility estimates, and a global propagation stage employing a structured Gaussian process to generalize corrections across space and archetypes [3]. The result is a spatially coherent fragility field that connects local evidence to regional understanding [3].