Prediction and interpretation of concrete corrosion induced by carbon dioxide using machine learning

Abstract

The CO2-induced corrosion of concrete containing solid waste materials (slag and fly ash) rich in Ca, Si, Al, Fe, and Mg oxides is investigated. In this work, more than 1700 samples are extracted, each accompanied by 29 feature data points. Subsequently, 8 key supervised machine learning (ML) algorithms were explored to predict the concrete corrosion induced by CO2, including Decision Tree (DT), Random Forest (RF), AdaBoost, CatBoost, and XGBoost, as well as K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Multilayer Perceptron (MLP). Filling the gap in the research of the ML model in predicting the carbonation depth of concrete incorporating solid waste materials. After optimizing the hyperparameters, the XGBoost model exhibits the best prediction accuracy and generalization ability, with MSE, RMSE, MAE, and R2 values of 7.94, 2.82, 1.83, and 0.92, respectively. Based on the optimally trained XGBoost model, Shapley Additive exPlans (SHAP) and Partial dependency plots (PDP) were used to evaluate the independent impacts of ‘carbonation time’, ‘water’, ‘aggregate’, and each raw material, as well as the interactive effects between features and oxide components.