Interpretable machine learning model for autogenous shrinkage prediction of low-carbon cementitious materials

Abstract

Supplementary cementitious materials (SCM) are key components of low-carbon cementitious materials. However, their effects, especially for the emerging high cement replacement ratio formulations, are sometimes very challenging to predict, leading to restrictions in most standards. Through advanced Machine Learning techniques, this study provides novel insights into the autogenous shrinkage properties of concrete containing various kinds of supplementary cementitious materials such as slag, fly ash, silica fume, calcined clay, and filler. Four machine learning models are optimized and their predictions are compared on a dedicated database, including ternary and quaternary cement blends. The Extreme Gradient Boosting (XGBoost) model exhibited the highest accuracy among these models. Interpretative tools such as SHAP analysis are then used to obtain novel information about the relative influence of the SCM on the autogenous shrinkage depending on their dosage. The model is finally compared to the recently modified versions of the B4 and the CEB analytical models to predict the autogenous shrinkage of unknown concrete formulations. The model showed significantly better predictions than the analytical models, and a web application is proposed, paving the way for precise and interpretable autogenous shrinkage predictions of cementitious materials incorporating SCM.