Integrating Bayesian, ensemble, and hybrid machine learning models for predicting overbreak and underbreak in underground stope mining

Overbreak and underbreak in underground stoping operations cause significant challenges, impacting safety and operational efficiency, leading to dilution and ore loss. This study employs machine learning methods to predict overbreak and underbreak at individual drillhole locations, enabling dynamic prediction of stope shapes before blasting in underground mining. Based on a comprehensive database collected from a longhole sublevel stoping mine in Western Australia, different machine learning-based models, including multiple linear regression, random forest regression, gradient boosting machine, and Gaussian process regression, were developed systematically. The best-performing model was the initial Gaussian process regression with the Matérn 3/2 kernel, achieving R² value of 0.820 and a root mean squared error value of 0.452. A series of hybrid models were also developed to improve prediction accuracy by combining Gaussian process regression, gradient boosting machine[A5.1], random forest regression, and artificial neural network with linear meta-learners. The best hybrid configuration achieved with R² value of 0.721 and RMSE value of 0.425, but further complexity led to overfitting with no gains in generalisation. SHapley Additive exPlanations analysis across all models highlighted the Mathews stability number, ratio of burden to spacing, and void ratio as the most influential variables, varying in impact between the random forest regression, gradient boosting machine, and Gaussian process regression models.