Bridging data gaps in fragility modeling for coastal infrastructure resilience

Coastal infrastructure is highly exposed to hurricanes and severe storms, underscoring the need for reliable fragility models that predict structural performance under extreme hazards.Over the past fifty years, methodologies for generating and validating building fragility curves have advanced considerably, but calibration and validation continue to lag due to scarce damage records and domain-specific assumptions.While numerical and simulation-based frameworks exist for singleand multi-hazard scenarios, they may lack transferability and robust empirical verification.Machine learning techniques present opportunities to overcome these limitations by enabling adaptive, continually evolving fragility functions.We propose a methodology integrating conventional transfer learning and hierarchical Bayesian transfer learning to enhance model flexibility and precision, especially in data-scarce coastal regions.This approach refines existing simulation-or data-rich region-derived models with limited labeled damage data from new target areas, improving predictive accuracy and closing the gap between theoretical simulations and observed outcomes.We demonstrate its feasibility through residential building case studies impacted by Hurricane Ian.By leveraging both physical simulations and sparse field observations, the framework supports ongoing model learning and adaptation.By delivering more accurate, region-specific fragility assessments, this framework advances vulnerability analysis and supports resilient coastal infrastructure planning.