Explainable machine learning for predicting compressive strength of alkali activated material: a comparative modelling and interpretability study

To promote the design of low-carbon alkali-activated materials, precise yet interpretable compressive strength prediction based on mix design variables is essential to minimize trial-and-error experimentation. This study employs an interpretable machine learning approach to predict compressive strength based on important mix design variables, including fly ash content, ground granulated blast furnace slag content, Na 2 O content, SiO 2 /Na 2 O ratio and water to binder ratio. Various ensemble machine learning models, such as Random Forest, Gradient Boosting Decision Tree, XGBoost, and LightGBM were developed and compared for model performance evaluation using R 2 , RMSE and MAE. Among the models, the Random Forest model had the best generalization performance with a test coefficient R² of 0.762, along with RMSE of 9.23 MPa and MAE of 6.38 MPa. Explainable artificial intelligence techniques were then used to interpret the machine learning model predictions and determine the variable contributions. The results consistently showed that the water to binder ratio is the most dominant factor influencing the compressive strength, followed by Na 2 O content, fly ash content and SiO 2 /Na 2 O ratio. A clear nonlinear threshold was found at a water to binder ratio of approximately 0.5, beyond which the predicted compressive strength dropped rapidly. Therefore, the interpretable machine learning framework presented in this study minimizes trial-and-error experimentation by providing accurate compressive strength prediction and clear, mix-design-related explanations for the development of low-carbon alkali-activated materials.