Advancing seismic risk assessment: leveraging pre-trained neural networks for seismic fragility models

Fragility curves are indispensable for quantifying the seismic risk of highway bridges, offering critical insights into the resilience and safety of transportation networks in earthquake-prone regions.Existing methodologies for developing fragility curves rely heavily on extensive post-event reconnaissance data or computationally intensive numerical simulation, posing significant challenges due to high resource demands and time constraints.While recent advancements in machine learning-based fragility models have introduced alternative approaches that reduce computational overhead, these methods typically require large training datasets, which are often inaccessible for older bridge structures or regions with limited seismic data.To address these challenges, this study introduces a novel transfer-learning-based fragility modeling framework designed to minimize data requirements without compromising accuracy or generalizability.The proposed approach leverages a sequential neural network pre-trained on a large fragility analysis dataset and fine-tunes it using a smaller, region-specific dataset of earlier bridge designs.By freezing all but the output layer and bottom-most hidden layers during fine-tuning, the transfer learning process effectively tailors the model to predict damage probabilities for underrepresented bridges while preserving the learned representations from the larger dataset.The model's performance was rigorously validated against traditional analytical fragility curves using R-squared values, demonstrating substantial improvements in predictive accuracy compared to models trained solely on smaller datasets.Ten-fold cross-validation confirmed the framework's robustness, ensuring generalizability without overfitting, even with limited training data.This innovative transfer learning approach offers a scalable, effective, and data-efficient solution for seismic fragility analysis, addressing critical data limitations while enhancing the accuracy of risk assessments.By enabling informed decision-making, this methodology contributes to the resilience of transportation infrastructure, supporting the design, retrofitting, and maintenance of highway bridges in seismic regions and promoting sustainable, disaster-resilient communities.