Prediction of Carbonation Depth using Machine Learning

Carbonation is one of the main durability mechanisms for reinforced concrete as the atmospheric CO2 enters the concrete cover and reacts with the alkaline hydration products, lowers the pH of the pore solution, can initiate corrosion of the reinforcement when the carbonation front reaches the steel surface. Typical carbonation models can be useful but are often based on simplifications of the nonlinear effects of water-to-binder ratio, exposure time, relative humidity, carbon dioxide concentration, curing, supplementary cementitious materials, recycled aggregate content and compressive strength. The paper proposes a machine learning-based approach for the prediction of the carbonation depth of a set of 120 concrete observations obtained from a structured accelerated-carbonation testing plan, consisting of primary experimental data. The set consists of mix-design, mechanical, environmental and exposure variables, with the carbonation depth (mm) being the response variable. Seven regression models were used for testing and compared: multiple linear regression, decision tree, random forest, support vector regression, artificial neural network, gradient boosting and extra trees. Model accuracy was measured in terms of R2, RMSE, MAE, MAPE and five-fold cross validation. The best performing model was Artificial Neural Network, with test R2 = 0.837, RMSE = 2.96 mm, MAE = 2.48 mm, and MAPE = 25.51%. The feature importance analysis revealed that the duration of exposure, the compressive strength, water to binder ratio, CO2 concentration and relative humidity were the most important variables. The results validate the hypothesis that carbonation prone concrete structures can be accurately and interpretable be predicted using nonlinear ensemble models. The proposed framework can aid in the service-life assessment, maintenance planning, and sustainable concrete mix evaluation.