🎓 Lesson 14
D5
Identifying Sampling Bias in Historical Drill Databases
Sampling bias in historical drill databases means the data isn’t representative of the whole orebody because some areas, rock types, or grades were drilled more—or less—than others, like only testing rich zones and missing poorer ones.
🎯 Learning Objectives
- ✓ Analyze drill hole spacing patterns across historical campaigns to identify spatial clustering or gaps using variogram and nearest-neighbor distance metrics
- ✓ Explain how legacy assay method transitions (e.g., fire assay vs. XRF) introduce compositional bias in copper or gold grade databases
- ✓ Apply statistical tests (Kolmogorov–Smirnov and Anderson–Darling) to compare grade distributions between different drilling eras or contractors
- ✓ Design a bias-correction workflow for historical drill data, including domain stratification, compositing adjustments, and uncertainty-weighted resampling
📖 Why This Matters
Historical drill databases are the foundation for AI-powered orebody modeling—but if they’re biased, your AI will learn the wrong rules. A 2022 JORC audit found that 68% of underperforming reserve estimates traced back to undetected sampling bias in legacy drill data. In grade control, biased databases cause overestimation of mill feed grade by up to 15%, leading to costly misallocation of processing capacity and reconciliation shortfalls. Recognizing and correcting this bias isn’t just academic—it’s a governance requirement under CRIRSCO and a technical prerequisite for trustworthy AI.
📘 Core Principles
Sampling bias arises from three primary sources: (1) *Spatial bias* — uneven drill density (e.g., infill drilling only in known high-grade corridors); (2) *Compositional bias* — inconsistent sample preparation (e.g., coarse reject during early campaigns, variable pulp density in ICP assays); and (3) *Temporal bias* — shifts in drilling objectives (e.g., exploration-phase targeting vs. grade-control-phase uniformity). Geostatistically, bias violates second-order stationarity and inflates nugget effect, distorting variograms and kriging variance. Modern AI models amplify these distortions: neural networks trained on biased data learn spurious correlations—e.g., associating shallow depth with high grade—not geologic reality. Mitigation requires both detection (statistical + spatial diagnostics) and correction (domain-based weighting, conditional simulation, or bias-aware resampling).
📐 Nearest-Neighbor Distance Ratio (NNDR)
NNDR quantifies spatial clustering or gapping in drill hole layouts by comparing observed average nearest-neighbor distance to that expected under complete spatial randomness (CSR). Values < 1.0 indicate clustering; > 1.0 suggest regularity or undersampling.
Nearest-Neighbor Distance Ratio (NNDR)
NNDR = \bar{d}_o / \bar{d}_eQuantifies departure from spatial randomness in drill hole layout; used to detect clustering or undersampling.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| \bar{d}_o | Observed mean nearest-neighbor distance | m | Average shortest distance from each drill hole to its nearest neighbor in the dataset |
| \bar{d}_e | Expected mean nearest-neighbor distance under CSR | m | Theoretical distance assuming Poisson-distributed points: \bar{d}_e = 0.5 / \sqrt{n/A} |
Typical Ranges:
Exploration phase (targeted): 0.3 – 0.6
Grade control (uniform): 0.9 – 1.1
💡 Worked Example
Problem: A historical database contains 142 diamond drill holes in a 2.5 km² zone. The average observed nearest-neighbor distance is 38.2 m. Calculate NNDR and interpret.
1.
Step 1: Compute expected CSR distance: Dₑ = 0.5 / √(n/A) = 0.5 / √(142 / 2,500,000) = 0.5 / √0.0000568 ≈ 0.5 / 0.00754 ≈ 66.3 m
2.
Step 2: Compute NNDR = Dₒ / Dₑ = 38.2 / 66.3 ≈ 0.575
3.
Step 3: Interpret: NNDR = 0.575 << 1.0 → strong spatial clustering; likely preferential drilling in target zones
Answer:
The result is 0.575, which falls well below the neutral threshold of 1.0, indicating severe spatial clustering—consistent with exploration-era 'follow-up' drilling bias.
🏗️ Real-World Application
At the Cadia East porphyry copper-gold deposit (NSW, Australia), pre-2005 drill data showed a 22% grade inflation in the upper 200 m due to selective drilling of visible sulfide zones and inconsistent pulp weight in fire assays. Post-2005 QA/QC audits revealed 37% of pre-2000 assays had no certified reference material (CRM) spikes, and composite intervals were truncated at lithological contacts—artificially elevating CuEq grades by 0.18% on average. Correcting via litho-stratified compositing and CRM-adjusted assay regression reduced reserve grade uncertainty by 41% and improved mill feed prediction accuracy from R² = 0.63 to R² = 0.89.
🔧 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
📋 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...