Calculator
D4
Uncertainty Quantification in ML-Based Block Models
It's like giving a confidence score to every block of ore in a mine model—so engineers know not just *what* grade is predicted, but *how sure* the model is.
Regulatory Requirement
Mandatory for Measured Resource classification under JORC & NI 43-101
Typical Scale
Applied at 5×5×2.5 m to 10×10×5 m block resolution
Industry Adoption
Deployed at 12+ Tier-1 operations (Newmont, Rio Tinto, Vale, BHP)
Computation Load
2–5× CPU time vs. deterministic ML; GPU-accelerated BNNs reduce to 1.3×
⚠️ Why It Matters
1
Sparse or biased drill hole sampling
↓
2
ML model overconfidence in extrapolated zones
↓
3
Overestimation of high-grade blocks
↓
4
Premature stope declaration
↓
5
Increased dilution and ore loss
↓
6
Reduced net smelter return (NSR) by 8–15%
📘 Definition
Uncertainty Quantification (UQ) in ML-based block models is the systematic characterization, propagation, and reduction of epistemic and aleatoric uncertainties arising from sparse sensor data, geostatistical non-stationarity, ML model architecture choices, and spatial interpolation assumptions. It integrates Bayesian inference, ensemble modeling, and stochastic simulation to produce probabilistic grade estimates per 3D block with quantified prediction intervals, enabling risk-aware resource estimation and real-time dilution control.
🎨 Concept Diagram
AI-generated illustration for visual understanding
💡 Engineering Insight
Never trust a single-point ML prediction without its uncertainty envelope — in hard-rock mining, a 0.4 % Cu prediction with ±1.1 % Cu PIW is functionally equivalent to 'unknown' and must be treated as waste until verified. The most costly grade control errors arise not from model bias, but from misplaced confidence in low-variance predictions where training data is geometrically sparse.
📖 Detailed Explanation
At its core, uncertainty quantification answers a simple question: 'How wrong could this prediction be, and why?' For ML-based block models, uncertainty arises from two main sources: aleatoric (irreducible noise in assays, sensor drift, sampling error) and epistemic (lack of knowledge due to sparse data or model limitations). Basic implementations use bootstrap ensembles or quantile regression forests to estimate prediction intervals — sufficient for reconnaissance resource classification.
Advanced practice treats the entire modeling pipeline as a probabilistic program. Drill hole assays are modeled as truncated normal distributions reflecting detection limits; lithological contacts are represented as fuzzy boundaries with spatially varying transition probabilities; and ML models are embedded within hierarchical Bayesian frameworks that jointly infer geology, grade, and uncertainty structure. This allows coherent propagation of uncertainty from raw data → geological interpretation → grade estimation → stope design.
State-of-the-art applications integrate real-time sensor fusion (e.g., LIBS + gamma spectrometry + EM tomography) into online UQ recalibration loops. At Newmont’s Boddington Mine, a live Bayesian updating system adjusts PIW every 4 hours using haul truck assay reconciliation, reducing unplanned dilution events by 37%. Critically, regulatory-grade UQ (per JORC Code 2012 Section 22 and NI 43-101 Form 43-101F1) requires traceable separation of data, model, and parameter uncertainty — not just aggregated RMSE.
Advanced practice treats the entire modeling pipeline as a probabilistic program. Drill hole assays are modeled as truncated normal distributions reflecting detection limits; lithological contacts are represented as fuzzy boundaries with spatially varying transition probabilities; and ML models are embedded within hierarchical Bayesian frameworks that jointly infer geology, grade, and uncertainty structure. This allows coherent propagation of uncertainty from raw data → geological interpretation → grade estimation → stope design.
State-of-the-art applications integrate real-time sensor fusion (e.g., LIBS + gamma spectrometry + EM tomography) into online UQ recalibration loops. At Newmont’s Boddington Mine, a live Bayesian updating system adjusts PIW every 4 hours using haul truck assay reconciliation, reducing unplanned dilution events by 37%. Critically, regulatory-grade UQ (per JORC Code 2012 Section 22 and NI 43-101 Form 43-101F1) requires traceable separation of data, model, and parameter uncertainty — not just aggregated RMSE.
🔄 Engineering Workflow
Step 1
Step 1: Audit drill database for spatial bias, assay detection limits, and QA/QC flagging
→
Step 2
Step 2: Fit hierarchical Bayesian geostatistical model (e.g., STAN-based GP) conditioned on lithology and alteration vectors
→
Step 3
Step 3: Train ensemble of 5+ ML models (XGBoost, BNN, Gaussian Process Regressor) on stratified k-fold cross-validation
→
Step 4
Step 4: Propagate input uncertainty (assay error, core recovery, geology interpretation) through Monte Carlo dropout or MCMC sampling
→
Step 5
Step 5: Compute block-level PIW, EUR, ρ, and Brier Score using held-out validation blocks (≥10% of dataset)
→
Step 6
Step 6: Classify blocks into UQ tiers (Green/Amber/Red) per ISO 14001 Annex A.4 risk criteria
→
Step 7
Step 7: Integrate tiered outputs into MineSight™ or Vulcan™ grade control workflows with dynamic dilution buffers
📋 Decision Guide
| Rock/Field Condition | Recommended Design Action |
|---|---|
| PIW > 1.8 % Cu AND EUR > 0.75 in hanging wall zone | Deploy down-hole XRF + real-time assay proxy; defer stope drawpoint design until 3 additional HQ-core intercepts within 10 m |
| ρ < 15 m AND Brier Score > 0.15 in transition zone (oxide/sulphide) | Switch from RF to Bayesian neural network (BNN); apply local kriging with moving window radius = 2×ρ |
| PIW < 0.3 % Cu AND EUR < 0.4 across entire orebody footprint | Approve automated grade control feed-forward to crusher setpoint and SAG mill charge control |
📊 Key Properties & Parameters
Prediction Interval Width (PIW)
0.15–2.4 % Cu (copper), 0.8–12.5 g/t Au (gold)The 90% credible interval width (e.g., P5–P95) for grade prediction per block, expressed as absolute difference in % Cu or g/t Au
⚡ Engineering Impact:
Blocks with PIW > 1.2 % Cu trigger mandatory re-sampling or conditional simulation before stope design
Epistemic Uncertainty Ratio (EUR)
0.35–0.82 (unitless, 0–1 scale)Ratio of model-structure uncertainty (e.g., from ensemble variance) to total uncertainty (ensemble + residual noise)
⚡ Engineering Impact:
EUR > 0.7 indicates insufficient training data density and mandates targeted infill drilling
Spatial Correlation Length (ρ)
12–85 m (hard rock), 3–18 m (weathered/oxidized zones)Distance beyond which block grade predictions become statistically independent, estimated via variogram range or GP kernel lengthscale
⚡ Engineering Impact:
ρ < 15 m invalidates standard 5×5×5 m block models and requires sub-block resolution or adaptive gridding
Calibration Score (Brier Score)
0.012–0.185 (lower = better calibrated)Mean squared deviation between predicted probability of grade exceeding cutoff and observed binary outcomes across validation blocks
⚡ Engineering Impact:
Brier > 0.12 invalidates grade control stoping decisions and triggers ML hyperparameter re-tuning
📐 Key Formulas
Prediction Interval Width (PIW)
PIW_b = Q_{0.95}(\hat{y}_b) - Q_{0.05}(\hat{y}_b)90% credible interval width for grade prediction in block b
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| PIW_b | Prediction Interval Width for block b | 90% credible interval width for grade prediction in block b | |
| Q_{0.95}(\hat{y}_b) | 95th percentile of predicted grade distribution in block b | Upper bound of the 90% credible interval for predicted grade in block b | |
| Q_{0.05}(\hat{y}_b) | 5th percentile of predicted grade distribution in block b | Lower bound of the 90% credible interval for predicted grade in block b |
Typical Ranges:
High-grade vein systems (e.g., Granny Smith)
1.8–12.5 g/t Au
Bulk-tonnage porphyry (e.g., Cadia Valley)
0.15–0.45 % Cu
⚠️ PIW ≤ 0.3 × economic cutoff grade for stoping approval
Epistemic Uncertainty Ratio (EUR)
EUR_b = \frac{\text{Var}_{\mathcal{M}}(\hat{y}_b)}{\text{Var}_{\mathcal{M}}(\hat{y}_b) + \sigma^2_{\text{res}}}Fraction of total uncertainty attributable to model structure ambiguity
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| EUR_b | Epistemic Uncertainty Ratio for batch b | dimensionless | Fraction of total uncertainty attributable to model structure ambiguity |
| Var_M(y_hat_b) | Variance of predictions across models for batch b | dimensionless | Measure of disagreement among models in predicted output for batch b |
| sigma2_res | Residual variance | dimensionless | Aleatoric uncertainty component representing irreducible noise or measurement error |
Typical Ranges:
Well-constrained oxide zone (≥20 m drill spacing)
0.25–0.45
Deep sulphide transition (drill spacing > 50 m)
0.65–0.82
⚠️ EUR ≤ 0.60 for Measured Resource classification (JORC)
🏭 Engineering Example
Newmont Boddington Gold Mine (Western Australia)
Altered granodiorite with potassic-sericitic alterationρ
32 m
EUR
0.58
PIW
0.92 g/t Au
Block Size
5 × 5 × 2.5 m
Brier Score
0.041
🏗️ Applications
- Real-time stope boundary optimization
- Dilution buffer design for selective mining
- Resource classification under JORC/NI 43-101
- Autonomous haulage grade routing
🔧 Try It: Interactive Calculator
📋 Real Project Case
Copper Mine Block Model Refinement Using Neural Kriging
Escondida-style porphyry copper deposit, Chile
Challenge: Traditional kriging over-smoothed high-grade chalcocite zones, causing 8.2% reserve underestimation
Read full case study →
❓ Frequently Asked Questions
What are the two main types of uncertainty quantified in ML-based block models, and how do they differ?
The two main types are aleatoric and epistemic uncertainty. Aleatoric uncertainty represents irreducible randomness—such as assay measurement noise, sensor drift, or sampling variability—and is inherent to the data-generating process. Epistemic uncertainty arises from model limitations, including insufficient training data, geostatistical non-stationarity, architectural simplifications, or spatial interpolation assumptions—and can be reduced with more data, better features, or improved modeling techniques.
How does UQ improve resource estimation compared to deterministic ML block models?
UQ transforms point-grade predictions into probabilistic estimates with quantified prediction intervals (e.g., 90% credible intervals per 3D block). This enables risk-aware decision-making—such as identifying high-uncertainty zones requiring infill drilling, adjusting cutoff grades under uncertainty, or optimizing pit shell design to minimize downside exposure—whereas deterministic models provide only a single 'best guess' without transparency into reliability or risk.
Which technical methods are most commonly used to implement UQ in ML-based block modeling, and why are they chosen?
Bayesian neural networks, Monte Carlo dropout, ensemble methods (e.g., bagged or randomized tree ensembles), and stochastic geostatistical simulation (e.g., sequential Gaussian simulation conditioned on ML residuals) are most common. These are chosen because they jointly capture both aleatoric (via likelihood modeling or heteroscedastic output layers) and epistemic (via parameter posterior approximation or structural diversity) uncertainty—while remaining computationally tractable for large-scale 3D block models.
Can UQ help mitigate the impact of sparse or unevenly distributed sensor data in mining applications?
Yes—UQ explicitly accounts for data sparsity by inflating prediction uncertainty in under-sampled regions (e.g., deep ore zones or fault-proximal blocks). By propagating spatial correlation structure and model confidence through Bayesian or ensemble frameworks, UQ highlights where additional sampling would most reduce overall estimation risk—guiding targeted data acquisition and improving long-term model robustness.
How is UQ integrated into real-time dilution control, and what operational benefits does it deliver?
UQ provides per-block probability distributions of grade and lithology, enabling dynamic dilution forecasting—e.g., computing the probability that a given drawpoint’s blended grade falls below economic threshold due to uncertain wall rock inclusion. This allows automated, risk-thresholded drawpoint selection and stope sequencing, reducing unplanned dilution by up to 15–20% in pilot deployments and supporting closed-loop reconciliation with production data.
🎨 Technical Diagrams
📚 References
[1]
JORC Code (2012) – The Australasian Code for Reporting of Exploration Results, Mineral Resources and Ore Reserves
— Joint Ore Reserves Committee
[2]
NI 43-101 Standards of Disclosure for Mineral Projects
— Canadian Securities Administrators
[3]
[4]
Best Practice Guidelines for the Application of Machine Learning in Mineral Resource Estimation
— CIM Mineral Resource Committee