Prediction of seismic capacity limit states of corroded bridge piers considering climate change: A multi-task learning approach

Seismic capacity limit states (LSs) of reinforced concrete (RC) bridges deteriorate over time due to corrosion, a process further exacerbated by climate change. This compounding effect necessitates a time-dependent, multi-hazard capacity model. However, the inherent complexity and uncertainty associated with compounding effects of corrosion and climate change, make conventional methods computationally intensive, particularly in probabilistic seismic risk assessments involving large bridge inventories. This study proposes a novel Multi-Task Learning (MTL)-based machine learning framework that enables different tasks to share information for improved prediction of capacity LSs and failure modes of corroded bridge piers. A comprehensive synthetic dataset representing typical highway bridges under different environmental scenarios is generated to train the proposed MTL model and several conventional single-task learning (STL) models. All models are optimized by hyperparameter tuning, and their results were subsequently interpreted to assess the contribution of key input variables. The results show that the MTL framework outperforms the STL models in terms of accuracy, generalization, and computational efficiency. The proposed model is then integrated into a fragility framework to predict LSs. Based on these predictions, the fragility curves are generated and compared with those obtained from finite element analysis for the case studies, showing close agreement and demonstrating the reliability of the proposed method. The framework provides a rapid assessment tool for evaluating the seismic capacity of aging, corroded bridges under climate change conditions, offering utility in both design and maintenance applications. • A time-dependent multi-hazard ML-based capacity model for bridge piers is developed. • Corrosion and climate change effects on bridge pier seismic behavior are considered. • A novel multi-task learning model to enhance accuracy and reduce costs is proposed. • The model predicts section-based and component-level limit states with failure mode. • Multi-task learning model outperforms single-task models in accuracy and efficiency.