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

This study utilizes machine learning (ML) techniques to predict the performance of slag-based cemented tailings backfill (CTB) activated by soda residue (SR) and calcium carbide slag (CS). An experimental database consisting of 240 test results is utilized to thoroughly evaluate the accuracy of seven ML techniques in predicting the properties of filling materials. These techniques include support vector machine (SVM), random forest (RF), backpropagation (BP), genetic algorithm optimization of BP (GABP), radial basis function (RBF) neural network, convolutional neural network (CNN), and long short-term memory (LSTM) network. The findings reveal that the RBF and SVM models demonstrate significant advantages, achieving a coefficient of determination (<i>R</i>²) of approximately 0.99, while the <i>R</i>² for other models ranges from 0.86 to 0.98. Additionally, a dynamic growth model to predict strength is developed using ML techniques. The RBF model accurately predicts the time required for filling materials to reach a specified strength. In contrast, the BP, SVM, and CNN models show delays in predicting this curing age, and the RF, GABP, and LSTM models tend to overestimate the strength of the filling material when it approaches or fails to reach 2 MPa. Finally, the RBF model is employed to perform coupling analysis on filling materials with various mix ratios and curing ages. This analysis effectively predicts the changes in filling strength over different curing ages and raw material contents, offering valuable scientific support for the design of filling materials.