Predictive Modelling of Alkali-Slag Cemented Tailings Backfill Using a Novel Machine Learning Approach

Area of Science:

• Geotechnical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Slag-based cemented tailings backfill (CTB) is crucial for mine waste management and ground support.
• Optimizing CTB performance requires accurate prediction of its mechanical properties.
• Soda residue (SR) and calcium carbide slag (CS) are potential activators for CTB, but their performance needs robust modeling.

Purpose of the Study:

• To evaluate the efficacy of seven machine learning (ML) techniques in predicting the performance of slag-based CTB.
• To develop a dynamic growth model for predicting CTB strength development.
• To identify the most accurate ML model for CTB performance prediction and material design.

Main Methods:

• An experimental database of 240 CTB test results was compiled.
• Seven ML techniques were employed: SVM, RF, BP, GABP, RBF, CNN, and LSTM.
• Model accuracy was assessed using the coefficient of determination (R²).

Main Results:

• The RBF and SVM models achieved the highest accuracy (R² ≈ 0.99).
• The RBF model accurately predicted the time to reach specific strengths, outperforming BP, SVM, and CNN.
• RF, GABP, and LSTM models showed overestimation tendencies for strength predictions near 2 MPa.

Conclusions:

• ML techniques, particularly RBF and SVM, are highly effective for predicting CTB performance.
• The RBF model provides a reliable tool for dynamic strength prediction and optimizing CTB mix designs.
• This study offers valuable data-driven insights for the sustainable design of mine backfill materials.