Stacking-ensemble ANN model to predict the carbonation depth of concrete structures exposed to natural degradation environments

Abstract

This study proposes an approach to enhance the prediction of concrete carbonation depth by combining Artificial Neural Networks (ANNs) with Ensemble learning techniques, with a specific focus on the Stacking fusion strategy. The proposed model was trained exclusively on carbonation data from concrete elements exposed to atmospheric CO₂ under natural degradation conditions, over short and long periods, and across diverse climate zones in thirteen countries. A total of 1031 data were compiled from the literature, including concrete elements both with and without exposure to rain. The model incorporates ten input variables, including concrete mixed proportions, environmental conditions, and exposure time. Results demonstrate that the Stacking-Ensemble approach improves generalization and accuracy performances by 9 % compared to base ANN models. Additionally, the proposed model outperforms existing mathematical formulations, particularly for long-term concrete structures, achieving a coefficient of determination (R²) of 0.91 and a root mean squared error (RMSE) of 1.43 mm. The final model, integrated into an accessible software tool, offers a practical and data-driven solution to support durability assessment and sustainable concrete design.