Predicting the initial cracking time of beam bridges in subtropical coastal area using field measured data

Accurate prediction of the initial cracking time (ICT) of reinforced concrete (RC) bridges is critical for durability assessment and maintenance planning. Conventional ICT models, predominantly calibrated on laboratory data, often fail to capture the complex degradation of in-service bridges in subtropical coastal environments. This study develops a dual-track prediction framework integrating a field-calibrated numerical model with eight machine learning (ML) algorithms trained on 756 field records. Within this framework, the modified numerical model enables rapid prediction and serves as an independent plausibility check for data-driven forecasts, while the ML algorithms capture complex field nonlinearities and quantify predictive uncertainty, thereby compensating for the simplified constitutive assumptions of the numerical formulation. Both tracks are rigorously evaluated on an independent validation set of 1102 records across three coastal provinces. The modified numerical model, incorporating field-measured cover thickness, carbonation depth, concrete strength, and a semi-carbonated zone safety margin, reduces the MAE from 10.57 to 3.04 years, substantially mitigating systematic overestimation. For the ML algorithms, the support vector machine achieves the highest cross-regional accuracy with an MAE of 1.96 years. Cross provincial validation reveals a systematic performance gradient that delineates the framework’s generalization boundary. This integrated approach provides a robust decision-support tool for the lifecycle management of RC infrastructure in subtropical coastal areas.