Safer Underground Stopes with Quantified Uncertainty: An Interpretable and Data‐Driven Update to the Stability Graph

ABSTRACT

Accurate stope stability assessment is essential for safe underground mining. This study develops an interpretable, data‐driven framework to update the conventional stability graph and reduce its reliance on predefined linear boundaries. Based on 426 cases, a multilayer perceptron optimized by balancing composite motion optimization achieves a test accuracy of 90.70%. Stratified bootstrap resampling confirms stable lower bound performance, with 95% confidence intervals of [0.85, 0.95] for accuracy, [0.83, 0.95] for balanced accuracy, and [0.85, 0.95] for weighted F1 score. Monte Carlo perturbation of rock mass and geometric inputs produces only about a 1% label‐flip rate, indicating robustness to moderate input uncertainty. Sensitivity and interpretability analyses identify the stability number as the dominant factor, while hydraulic radius acts as a secondary variable with strong interaction effects. Residual Kriging is then used to calibrate predicted probabilities in the stability graph, select an optimal decision threshold of 0.3, and define an uncertainty zone for ambiguous conditions. Relative to existing stable–unstable prediction boundaries, the updated graph shows stronger class separation, and external validation on two independent datasets yields conservative accuracies above 90%. The framework is mainly intended for open stoping and open‐span stability assessment; application to mining methods governed by different failure mechanisms requires site‐specific recalibration and validation. Overall, the study provides an interpretable and uncertainty‐aware pathway for quantitative stability graph refinement.