Transfer Learning Framework Bridges Data Gaps in Structural Fragility Models

This paper presents a methodology-centered transfer learning framework for adapting structural fragility models when faced with domain shift, class imbalance, and very few labeled target observations. The authors—Narges Saeednejad and Jamie Ellen Padgett from Rice University—demonstrate four distinct transfer learning strategies: instance-based (via importance weighting), parameter-based, hierarchical Bayesian (partial pooling with posterior uncertainty), and multi-source fusion. Each is validated through three complementary case studies using real disaster data.

In the first case study, instance-based transfer learning (importance weighting) is applied to coastal bridge fragility after Hurricane Katrina. The second uses parameter-based and hierarchical Bayesian transfer learning on residential building fragility from Hurricane Ian, enabling uncertainty quantification across strata. The third fuses multiple analytical models via multi-source transfer learning for seismic bridge fragility using observations from the 2001 Nisqually earthquake. Across all scenarios, directly applying existing state-of-the-art models fails under domain shift and severe class imbalance, while targeted adaptation significantly improves failure detection and predictive stability.

The findings underscore the need for systematic guidance on diagnostics, strategy selection, and uncertainty reporting when developing and adapting fragility models. The framework preserves engineering interpretability and supports decision-making under uncertainty, making it a practical tool for risk assessment in data-scarce environments.