Large Scale Mining

Large Scale Mining Optimization for Gold Extraction

Region

Multi-Year Data

Historical Operations

Timeline

Productivity Up

With Model Guidance

Annual Savings

Downtime Down

Reduced Shutdowns

Variance Reduction

Very Low Error

Recommendation Accuracy

Large Scale Mining Optimization for Gold Extraction

Executive Summary

A temporal deep learning model was trained on years of plant data to recommend optimal sequences for water injection rate, oxygen flow, temperature, and hold time. The system serves as operator decision support with actionable set points, improving productivity and reducing downtime while maintaining very low recommendation error.

Global producer of gold and gold-copper concentrates needed to optimize complex autoclave operations across multiple sites with varying operator expertise.

Business Challenge

Complex Multivariable Control

High operator-to-operator variability in managing interrelated process parameters affecting gold extraction efficiency.

Asset Protection vs. Productivity

The need to maximize extraction while protecting asset health and stability in high-pressure, high-temperature operations.

Operational Risk Management

Risk of reduced recovery and unplanned downtime when operating conditions are not optimal, leading to significant financial losses.

Industry Context

In pressure oxidation, crushed ore is mixed with water to form a slurry that is processed in an autoclave at high temperature and pressure with oxygen
Oxidation of sulfide minerals releases the gold and removes non-gold materials from the slurry
The process is sensitive to control choices, and suboptimal operation can cut recovery or force a shutdown
Experienced operators mitigate these risks, but variability remains significant

What We Built

Data and Signals

Historical Production Data

• Multiple years of autoclave operation data
• Time-series measurements at high frequency
• Process upset and recovery events
• Maintenance and shutdown records

Key Process Variables

• Water injection rate profiles
• Oxygen flow rate measurements
• Temperature trajectories
• Hold time durations

Outcome Metrics

• Gold recovery rates
• Throughput measurements
• Energy consumption
• Equipment health indicators

Modeling Approach

Temporal Deep Learning Architecture

Production-grade deep learning model with temporal architecture that learned the relationship between control sequences and downstream outcomes. Sequential modeling of industrial time series enables the system to respect plant dynamics.

Decision Support Framework

Model recommends operator actions as clear set points for the four control levers, packaged as decision support rather than automatic control. Preserves operator oversight while reducing cognitive load.

Stability-Aware Recommendations

System proposes stable, feasible adjustments that respect equipment constraints and process dynamics, avoiding recommendations that could trigger upsets or shutdowns.

Planning and Simulation Tool

The solution provides real-time recommendations through an operator interface that displays suggested set points alongside current operating context, historical performance, and confidence indicators.

Operator Interface

Intuitive display of recommended set points with clear visualization of current vs. suggested states

Context Integration

Recommendations presented alongside relevant process context and historical trends

Shift Handover Support

System captures operational knowledge for smooth transitions between shifts

Change Management

Training on years of labeled operating history to capture a wide range of conditions

Validation against held-out periods to verify low average recommendation error

Deployment as an operator aid that surfaces recommended set points alongside current context

Gradual adoption starting with advisory mode before full implementation

Results and Impact

Productivity ↑

Throughput Improvement

After adopting model-guided set points

Downtime ↓