This study aims to develop a physics-informed machine learning framework to improve carbonation depth prediction in reinforced concrete by integrating mechanistic knowledge with data-driven learning.
A CatBoost model incorporating ten physics-guided features was trained on a data set expanded to twice the original training set via physics-constrained augmentation and validated using multiple train-test splits.
The proposed model achieved R2 = 0.996 and RMSE = 3.59 mm. The physics-based carbonation coefficient kestimate was the dominant predictor, contributing 55.3% of the normalized SHapley Additive exPlanations importance, while engineered physics features collectively accounted for 77% of the predictive contribution.
The model requires further validation against independent natural-exposure data sets to ensure long-term reliability beyond accelerated testing conditions.
This framework shows potential for supporting carbonation-depth assessment within the experimental domain represented by the available data. Further validation using independent natural-exposure data sets is required before its application to long-term service-life assessment.
Improved durability predictions contribute to public safety and environmental sustainability by extending the lifespan of infrastructure and reducing the carbon footprint associated with premature reconstruction.
The proposed framework bridges deterministic carbonation theory with explainable machine learning, providing a physically interpretable approach for carbonation-depth assessment within the investigated experimental domain.
