Machine learning-based prediction of carbonation depth in alkali-activated materials: Integrating physics knowledge and data augmentation

Carbonation remains a critical barrier to the widespread structural application of alkali-activated materials, and predictive models capable of capturing the complex, nonlinear relationships between carbonation depth and its governing factors are still limited. To address this gap, we developed five machine learning models and laid emphasis on an artificial neural network (ANN) model embedded with a symbolic physics-informed formula and a dropout layer. The results demonstrated that the physics-informed ANN model outperformed the other four approaches in terms of accuracy and robust generalization. The employed conditional tubular generative adversarial network (CTGAN) proved to be less effective, evident from the compromised prediction accuracy and generalization performance. SHAP analysis indicated that carbonation time and CO₂ concentration were primary contributors to carbonation depth, while MgO content and modulus of the activator solution made minimal contributions. Additionally, CaO content showed a notably positive effect on carbonation resistance. Overall, the authors believed that the proposed ML framework was effective and offered practical potential for field applications.