Attention-enhanced neural networks and ensemble learning for accurate prediction of concrete carbonation

Concrete carbonation is a major contributor to reinforcement corrosion and durability loss in reinforced concrete, threatening structural integrity over time. While substantial research has been conducted on predicting carbonation depth using machine learning and deep learning techniques, many existing studies are limited by small datasets (300–600 samples) and a narrow focus on fly ash concrete, which restricts the robustness and generalizability of the models. In contrast, this study presents a novel approach by integrating a comprehensive dataset of 2163 samples from both natural and accelerated carbonation experiments, which is not confined to fly ash concrete. Eight key input variables were selected, and the dataset was randomly divided into training (80 %) and testing (20 %) subsets. Five predictive models were evaluated: Random Forest (RF), two baseline deep learning models (ANN and CNN), and two enhanced models that incorporate feature interaction terms, carbonation-related equations, and attention mechanisms (ATT-ANN and ATT-CNN). The results show that the enhanced models significantly outperform the baseline approaches in both accuracy and generalization. Specifically, ATT-CNN achieved the highest performance, reducing MSE by 24.6 % and 11.4 % relative to RF and CNN, respectively, and improving R² to 0.9142. ATT-ANN also demonstrated notable improvement, achieving an R² of 0.9063. Additionally, the integration of SHAP, Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) methods provided deep insights into the decision-making process of the best-performing models. The findings indicate that carbonation time, temperature and humidity interactions, and the interaction between water-binder ratio and CO₂ concentration are the most influential factors in carbonation prediction. This study offers a novel, interpretable framework that not only enhances predictive accuracy and robustness but also deepens understanding of carbonation mechanisms, contributing to more scientifically rigorous service life assessments of reinforced concrete structures.