Early-age compressive strength prediction of cemented phosphogypsum backfill using lab experiments and ensemble learning models

Abstract

The unconfined compressive strength (UCS) of cemented phosphogypsum (PG) backfill is an important mechanical index that determines the stope safety. However, the traditional mechanical test is costly and time-consuming. In this study, UCS test were conducted to establish the dataset for model construction. The dataset contained 81 UCS results with different combinations of influencing variables, including the pH of PG, the electrical conductivity (EC) of PG, the pH of backfill slurry, the binder/PG ratio, the solid concentration, and the curing age. Six ensemble learning models, including AdaBoost, gradient boosted decision tree (GBDT), extreme gradient boosting (XGBoost), the light gradient boosting model (LGBM), random forest (RF), and extremely randomized trees (ExtraTrees), were constructed to predict the UCS of cemented PG backfill. The results show that the six ensemble learning models obtained values of root means square error (RMSE), mean absolute error (MAE) and coefficient of determination (R2) in the test set ranging from 0.050 MPa to 0.113 MPa, 0.034–0.154 MPa, and 0.933–0.990, respectively, and these exhibited smaller errors than standalone machine learning models. Moreover, the XGBoost model performed the best among the six ensemble learning methods, with RMSE= 0.050 MPa, MAE= 0.034 MPa, and R2 = 0.990 for the test set. The study revealed the satisfactory prediction performance of ensemble learning models for the UCS of cemented PG backfill. Among six models, this study recommends XGBoost model as a reliable tool for quick and accurate prediction of the UCS of cemented PG backfill.

Keywords

• Machine learning
• Ensemble learning
• Cemented PG backfill
• Unconfined compressive strength
• Prediction
• Modeling