A Sustainable and Practical Machine Learning Approach Using Scikit-Learn for Predicting Stope Instability: Identification of Critical Geotechnical Factors

Abstract

Stope instability remains a persistent and hazardous challenge in underground mining, impacting safety, efficiency, and sustainability. Traditional stability assessment methods, while valuable, are often limited by site-specific calibration, simplifications, and adaptability issues in dynamic underground conditions. While machine learning shows potential for improved accuracy, a critical gap persists in understanding how geotechnical factors interact in practice. This study introduces a novel, practical machine learning framework (Scikit-Learn) to predict stope instability, and crucially, to quantify the nuanced, non-linear influence and interaction of critical geotechnical factors in a shallow gold mine. Comprehensive geotechnical investigation (observations, lab tests, rock mass classifications, blast damage assessments) and advanced data analysis (Random Forest feature importance, RFE, decision boundary analysis) identified water ingress, blast-induced damage, and rock mass quality (RMR) as the most significant instability factors. Water ingress profoundly impacted stability, with moderate blast damage exacerbating instability under high water ingress. Rock strength showed comparatively lower significance. The developed model achieved robust predictive performance (accuracy: 0.83, precision: 0.88, recall: 0.83, F1-score: 0.83). Based on these insights, tailored support patterns (e.g. 22mm/16mm cone bolts, timber props) are proposed to mitigate specific risks. This research significantly advances targeted rock mechanics solutions by providing a deeper, quantifiable understanding of complex instability mechanisms, enhancing mine safety and operational efficiency in shallow gold mining.