Life-cycle performance prediction and interpretation for coastal and marine RC structures: An ensemble learning framework

Abstract

Long-term exposure to coastal and marine environments accelerates the aging of reinforced concrete (RC) structures, impacting their structural safety and society impact. Traditional assessments of long-term performance deterioration in RC structures involve complex, nonlinear, and time-intensive studies of physical mechanisms. While existing machine learning (ML) methods can assess the lifetime of these structures, they often prioritize data regression over mechanistic interpretation. To enhance the efficiency and interpretability of predicting the life-cycle performance of RC structures, this study introduces a generic framework based on interpretable ensemble learning (EL) methods. The framework predicts life-cycle performance efficiently and accurately, with optimal hyperparameters automatically tuned through Bayesian optimization. Interpretability algorithms clarify the influence of environmental, durability, and mechanical parameters on structural durability and mechanical predictions. Validation employs real-world cases of RC hollow beams in the coastal area of the Guangdong-Hong Kong-Macao Greater Bay Area (GBA). The comprehensive model for RC structures integrates actual data on temperature, humidity, and surface chloride content in the GBA, considering diffusion, convection, and binding effects of chloride ions, corrosion non-uniformity, and crack impact on durability estimation. Comparative analysis with existing ML methods underscores the effectiveness of the framework. The findings highlight the dynamic evolution of feature importance rankings throughout the service life, shedding light on the continuous changes in the significance of different factors when predicting mechanical resistance.