Interpretable ensemble machine learning for the prediction of the expansion of cementitious materials under external sulfate attack

External sulfate attack (ESA) is a key degradation mechanisms of cementitious materials. Although the advantages of low-C3A cement and supplementary cementitious materials (SCM) have been confirmed, there remains a need for a better understanding of the phenomenon and guidance on accelerated testing due to the numerous parameters affecting this degradation. This study introduces a machine learning framework for predicting the expansion of cementitious materials incorporating SCM because of ESA. A comprehensive database is constructed, and four optimized machine learning models are compared. Among them, extreme Gradient Boosting (XGBoost) showed the best performance with a R2 accuracy of 0.933 and 0.788 on the training and the test set resp. Additionally, SHapley Additive exPlanations (SHAP) enabled the identification of the most influential inputs and their relative influence. It has been found that clinker composition, mix proportion, sample geometry, and sulfate solution characteristics play an important role, with their relative contribution being 34%, 36%, 3% and 27% resp. Furthermore, a thorough analysis of the model predictions on some expansive and non-expansive mortar and concrete samples demonstrated its reliability. Finally, the model was shown to be able to accurately predict the time required to reach a given expansion.

Automated summary

This study introduces a machine learning framework for predicting the expansion of cementitious materials due to external sulfate attack (ESA). Four optimized machine learning models are compared and extreme Gradient Boosting (XGBoost) showed the best performance. SHapley Additive exPlanations (SHAP) enabled the identification of the most influential inputs and their relative influence. The model has been shown to accurately predict the time required to reach a given expansion.