Wet muck spill hazards involve a sudden inflow of wet and fine-grained material from drawpoints in underground mines, posing risks to safety and infrastructure. Despite decades of operational experience, gaps exist in understanding the mechanisms governing wet muck spill susceptibility and severity. This research addresses these gaps through a data-driven approach integrating statistical analyses and machine learning techniques to develop improved spill hazard assessment frameworks. A key focus of this work was understanding the impact of draw strategy on spill occurrence. A conceptual model was developed from which two draw-related variables, Draw Rate (ḋ) and Differential Draw Index (DI), were derived. DI, introduced as a new parameter in this work to represent non-uniform draw, proved particularly significant. Statistical analysis of historical data from the Deep Ore Zone (DOZ) mine validated the conceptual model and showed that spill probability could increase by up to eight-fold when both ḋ and DI were in their upper ranges (i.e., heavy, isolated draw). Building on this work and incorporating the proposed draw-related variables, a random forest (RF) machine learning (ML) model was subsequently developed using historical data from the DOZ mine to provide a comprehensive spill susceptibility assessment and constrain the scenarios leading to spill events. The RF model demonstrated promising performance with an accuracy of 85%, with feature importances showing that previous spill history, fragment size, ḋ, DI, and spill history at neighbouring drawpoints had the highest impact on spill probability. SHAP results supported the interpretability of key contributing factors. Spill history at the newer Grasberg Block Cave (GBC) mine was then examined to identify wet muck spill severity trends and to develop an ML-based operational forecasting tool. The analysis showed that the most severe events occurred at drawpoints with higher Heights of Draw located below the pit bottom, where water and pit wall failure material preferentially accumulate. Additional contributing factors included previous spill severity, high ḋ, and high DI. The GBC model achieved an accuracy of 77%, with partial dependence results supporting the interpretability of key drivers. The developed ML model improved understanding of spill severity while complementing wet muck risk management practices.
Sahar Ghadirianniari (Thu,) studied this question.