🎓 Lesson 1
D1
Why AI Changes the Game in Orebody Delineation
AI helps mining engineers draw more accurate, real-time boundaries around ore deposits by learning from vast amounts of geological and geophysical data.
🎯 Learning Objectives
- ✓ Explain how supervised ML models (e.g., random forests) outperform inverse distance weighting in grade interpolation for complex mineralization
- ✓ Analyze the impact of input data resolution and uncertainty propagation on AI-predicted ore-waste boundaries
- ✓ Apply conditional simulation with Gaussian process regression to quantify spatial grade uncertainty at a 5-m block scale
- ✓ Design an AI-ready data pipeline for integrating core logging, assay composites, and downhole geophysics into a unified 3D training dataset
📖 Why This Matters
In traditional mining, orebody boundaries are often drawn manually or with static geostatistics—leading to 15–25% average grade misclassification and costly dilution or ore loss. At Rio Tinto’s Koodaideri mine, AI-driven delineation reduced waste mining by 12% in the first year by updating ore contacts using real-time blasthole assay feedback. This lesson shows how AI transforms orebody modeling from a one-time interpretation into a living, self-correcting system—fundamental to modern grade control and automated haulage.
📘 Core Principles
AI-powered delineation rests on three pillars: (1) Data fusion—integrating heterogeneous, multi-resolution inputs (e.g., gamma-ray logs, magnetic susceptibility, XRF scans) into a common voxel space; (2) Probabilistic learning—using ensemble or Bayesian models to output not just a 'best estimate' grade, but a full posterior distribution per block; and (3) Active learning—where the AI identifies high-uncertainty zones and recommends optimal next drill locations. Critically, unlike deterministic interpolators, AI models learn geological continuity rules implicitly—e.g., recognizing shear-hosted gold veining patterns without explicit structural modeling—and generalize across domains when trained on transferable features like spectral absorption ratios or texture gradients.
📐 Uncertainty-Aware Grade Prediction Confidence Interval
This formula computes the 90% prediction interval width for a block-grade estimate using Gaussian Process Regression (GPR), a widely adopted AI method in grade modeling. Narrower intervals indicate higher confidence in boundary placement—critical for cut-off grade decisions.
GPR Prediction Interval Width (90%)
PI_width = 2 × z_{0.95} × σ_predWidth of the 90% prediction interval for a block-grade estimate from Gaussian Process Regression, used to assess boundary confidence.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| PI_width | Prediction Interval Width | g/t | Total span (lower to upper bound) of the 90% grade prediction interval |
| z_{0.95} | Standard Normal Quantile | dimensionless | z-score corresponding to 95th percentile (1.645 for 90% two-sided interval) |
| σ_pred | Predictive Standard Deviation | g/t | Model-estimated standard deviation of the grade prediction at that location/block |
Typical Ranges:
High-confidence ore zone (well-drilled): 0.1 – 0.4 g/t
Marginal ore/waste contact zone: 0.6 – 1.8 g/t
Undrilled or geophysically ambiguous zone: 2.0 – 5.0+ g/t
💡 Worked Example
Problem: A GPR model predicts a 5×5×5 m block grade of 1.82 g/t Au with a standard deviation σ = 0.47 g/t. Compute the 90% prediction interval width.
1.
Step 1: Identify z-score for 90% two-sided confidence: z = 1.645
2.
Step 2: Apply formula: PI_width = 2 × z × σ = 2 × 1.645 × 0.47
3.
Step 3: Calculate: 2 × 1.645 × 0.47 = 1.5463 ≈ 1.55 g/t
4.
Step 4: Interpret: A width < 0.8 g/t indicates high-confidence classification at 1.0 g/t cut-off; this block (width = 1.55 g/t) requires additional sampling before stope design.
Answer:
The 90% prediction interval width is 1.55 g/t, exceeding the recommended threshold of 0.8 g/t for confident ore/waste assignment at a 1.0 g/t cut-off.
🏗️ Real-World Application
At Newmont’s Boddington Mine (Western Australia), a convolutional neural network (CNN) was trained on 22,000 m of core scan images (RGB + SWIR), 3,800 drill hole assays, and airborne magnetic data to delineate porphyry copper-gold zones. The AI model updated ore contacts every 72 hours as new blasthole assays arrived—reducing average grade variance in production blocks by 34% compared to ordinary kriging. Crucially, it flagged a previously undetected low-grade halo (0.25–0.45 g/t Au) surrounding the main ore zone, enabling selective mining and boosting recoverable ounces by 6.2% annually.
🔧 Interactive Calculator
🔧 Open AI-Powered Orebody Delineation & Grade Control Calculator📋 Case Connection
📋 Gold Mine Real-Time Grade Control at Development Drift Face
Manual chip sampling caused 24–48 hr delay in stope boundary decisions, leading to 14% dilution
📋 Iron Ore Mine Sensor Fusion for Banded Iron Formation (BIF) Delineation
Conventional geophysics failed to resolve thin hematite bands (<2m) within jaspilite, causing 22% grade variance in ROM...
📋 Limestone Mine Digital Twin for Karst-Related Grade Uncertainty
Solution cavities caused unpredictable grade drops (CaCO₃ purity <85%) in otherwise uniform deposits, resulting in 11% p...