AI-based system for planning and optimizing mining operations
A mining company spends $800,000 per month on blasting and is experiencing rock fragmentation that is 23% worse than predicted—drilling rigs deviate from planned trajectories, and the geological model doesn't take into account current data from MWD sensors. This is a common situation: a gap between the static plan and the dynamic reality of the quarry.
Where the mining plan breaks down
Geological model vs. reality
The block geological model is constructed using exploration borehole data on a 25x25 m grid. Interpolation between boreholes is performed using Kriging and Sequential Gaussian Simulation. As the mining front moves, it exposes rock previously unseen by the model. Deviations in metal content from the predicted value are 12–18% for copper deposits and up to 30% for complex polymetallic deposits.
ML solution: integration of MWD (Measurement While Drilling) data in real time. The drilling rig already knows the specific drilling energy (SE, kJ/m³), rate of penetration (ROP), and vibration. These signals correlate with rock hardness and update the block model on the fly.
Using a copper open pit mine as a case study, updating the model using MWD data reduced the predicted Cu grade deviation from 16% to 6%, resulting in $2.1M/year of additional recovery when converted to mill feed. Stack: XGBoost for assay prediction based on MWD features, Sequential Gaussian Simulation (SGS) for block recalculation, and Apache Kafka for streaming from drill rigs.
Optimization of blasting operations
The drilling pattern (burden, spacing, stemming length, delay timing) determines the particle size distribution (PSD) of the blasted rock mass. PSD directly impacts crusher performance and mill specific energy consumption. The traditional Kuz-Ram model provides an accuracy of ±30% based on P80.
Machine learning approach: regression on a historical dataset of post-blast laser scanning blasts. Input variables: geomechanical parameters (UCS, RQD, joint spacing), drilling pattern, explosive type, specific explosive consumption. Output: predicted P80. Pattern optimization: LightGBM + Optuna for Bayesian optimization for the target P80. A measurable result: a 7–11% reduction in mill specific energy consumption.
Production planning: short-term and strategic
Short-term mine planning (1–7 days)
Task: to create a work order for a fleet of equipment, taking into account the requirements of the processing plant for average grade (blending), geomechanical constraints (slope angles, rock pressure zones), and equipment availability.
Solution: MILP problem with ML prediction of excavator productivity depending on rock type. Solver: Gurobi or OR-Tools. 24-hour horizon, recalculated every 2 hours as the situation changes.
Long-term mine planning
Ultimate Pit Limit (UPL) + pushback sequence: the classic Lerchs-Grossmann problem, extended stochastic optimization to account for metal price and geological uncertainty. E-UPL optimizes for NPV using P10/P50/P90 scenarios, rather than a single baseline price forecast.
Predictive maintenance of the mining fleet
A Caterpillar 7495 excavator costs $30 million. Unscheduled shutdown: $180,000/hour of lost production.
The LSTM Autoencoder is trained on normal operating conditions (6 months of data with 1-minute intervals): transmission vibration, gearbox temperature, hydraulic system pressure, and Wear Particle Analysis. Reconstruction error > threshold = anomaly. On a fleet of 12 excavators: 34% reduction in unplanned stops over 18 months. Time series storage: TimescaleDB. Experiment tracking: MLflow.
Dump truck dispatching
Dispatch optimization: distributing dump trucks between excavators and unloading points. With 50+ trucks, it's an NP-hard problem. Reinforcement Learning (PPO) is trained on a quarry simulator. A fleet of 80 trucks: 8–12% productivity increase vs. baseline heuristics. DISPATCH (Modular Mining) and Wenco are industrial systems with APIs for ML integration.
Development stages
- Data audit: geological database, MWD historian, SCADA, dispatch system 2. ML-updating of the geological model by MWD 3. Short-term planning module (MILP + ML constraints) 4. PdM for priority equipment 5. Fleet dispatch optimization 6. Integration with ERP/dispatch system
Development period: 8–16 months depending on the scale of the field and the maturity of the IT infrastructure.