Interpretable ensemble machine-learning models for strength activity index prediction of iron ore tailings

Abstract

When iron ore tailings (IOTs) are used as supplementary cementitious materials, their strength activity index is an important factor affecting the mechanical properties of IOTs-cement composites. In this study, four ensemble machine learning (ML) models, including Gradient Boosting Regression Tree (GBRT), EXtreme Gradient Boosting (XGBoost), Adaptive Boosting (Adaboost), and Categorical Boosting (Catboost), were used to predict the strength activity index of IOTs. Shapley additive explanations (SHAP) were also utilized to measure the impact of influencing elements in the strength activity index of IOTs. Results showed that the R2 values of all four models were above 0.83 on the testing set. Dosage, water-solid ratio, and D0.5 particle size were found to be the most important factors affecting the strength activity index of IOTs. For high-silica IOTs, the strength activity index increased with the increase of dosage when the dosage was less than 10%, while when the dosage was higher than 10%, the opposite situation occurred. However, for low-silica IOTs, the strength activity index decreased continuously with the increase of dosage. This research developed a new approach for predicting the strength activity index of IOTs and provided some suggestions for the application of IOTs in cement-based materials.