Prediction of Backfill Slurry Shrinkage Rate and Optimization of Roof-Contact Backfilling Based on an Optimized Neural Network Approach

Abstract

In backfill mining of metal mines, the sedimentation and shrinkage of backfill slurry critically affect roof-contact quality and stope stability. This study investigates these characteristics for unsorted tailings backfill slurry from the Daye Iron Mine through laboratory experiments and develops a prediction model for the shrinkage rate. Tests examined the effects of cement-to-tailings ratio (1:4 to 1:20) and slurry concentration (65% to 71%) on shrinkage rate over curing time. A crested porcupine optimizer-backpropagation (CPO-BP) neural network model was developed to predict the shrinkage rate and compared with the traditional BP model. The robustness and generalization capability of the model were further verified by repeated five-fold cross-validation and by comparison with the GA-BP, PSO-BP, random forest, support vector regression, and XGBoost models. Results show that shrinkage predominantly occurs within the first 8 h and stabilizes after 24 h; high-concentration slurries stabilize earlier. At a fixed concentration, the shrinkage rate increases with decreasing cement-to-tailings ratio; at a fixed ratio, it decreases with increasing concentration. The CPO-BP model significantly outperforms the BP model, achieving a mean squared error (MSE) of 3.0767 × 10−6, a mean absolute error (MAE) of 1.4528 × 10−3, and a coefficient of determination (R2) of 0.94946 on the test set. For the Daye Iron Mine, a slurry concentration of 69–71% reduces shrinkage height and improves roof-contact rate. This study reveals the evolution of shrinkage rate with curing time, cement-to-tailings ratio, and concentration, demonstrates the effectiveness of the CPO-BP model, and optimizes the backfilling scheme, providing a theoretical basis for roof-contact quality control in practice.