Unravelling the correlation between natural and accelerated carbonation of low-carbon concrete using machine learning

Understanding correlation between accelerated and natural carbonation is paramount to accurately predicting concrete's long-term carbonation resistance in real-world conditions. However, this relationship is highly dependent on material properties, mix design, and environmental exposure, making the development of a generalized correlation formula unrealistic and nonviable. To address this complexity, this research proposes a machine learning framework to estimate the correlation index for “low-carbon” concrete specific to mix design and regionally relevant climatic exposures. Two probabilistic deep learning models, achieving testing R² of 0.95 in predicting natural and accelerated carbonation depths, were utilized to perform 768 carbonation simulations. The results demonstrate that the developed models provide a unique capability to link the carbonation rates of mixtures under different accelerated testing conditions (e.g., CO₂ concentrations) to the carbonation rates of the same mixtures under region-specific climatic exposure. This framework offers a practical tool for the rapid evaluation of long-term carbonation in low-carbon concrete.