Calculator
D3
Geostatistical Machine Learning for Orebody Modeling
It's like using smart computer models trained on real rock data and geological patterns to draw a more accurate 3D map of where valuable ore is — so miners waste less rock and hit the target grade every time.
Typical Scale
Used in deposits >50 Mt reserves; standard block size: 5–20 m edge length
Industry Adoption
Deployed at Rio Tinto’s Koodaideri, Newmont’s Boddington, and BHP’s Olympic Dam
Regulatory Context
Required for JORC/NI 43-101 compliant resource statements when traditional methods yield >15% grade variance
Computational Demand
128-core CPU + 4×A100 GPUs for 100-realization generation on 10M-block model (~4 hrs)
⚠️ Why It Matters
1
Sparse or biased drill sampling
↓
2
Poor spatial interpolation of grade
↓
3
Over- or under-estimation of ore tonnage
↓
4
Suboptimal stope design and sequencing
↓
5
Increased dilution and loss
↓
6
Reduced net smelter return (NSR) and shortened mine life
📘 Definition
Geostatistical Machine Learning for Orebody Modeling integrates geostatistical frameworks (e.g., kriging, multiple-point statistics) with supervised/unsupervised ML algorithms (e.g., random forests, graph neural networks) to honor spatial continuity, uncertainty quantification, and domain constraints while conditioning on multi-source data (drill assays, geophysics, hyperspectral imaging, borehole logs). It bridges stochastic simulation and data-driven prediction to generate conditional, probabilistic orebody realizations that preserve geological realism and support risk-informed mine planning decisions.
🎨 Concept Diagram
AI-generated illustration for visual understanding
💡 Engineering Insight
Never let ML replace geology — it must *constrain* ML. The most robust implementations embed geological rules as hard constraints (e.g., fault offsets, lithological contact angles) in the loss function or use them to mask invalid predictions post-inference. A model that honors structure but mispredicts grade is recoverable; one that violates geometry is operationally catastrophic.
📖 Detailed Explanation
At its core, geostatistical ML replaces traditional interpolation with learned spatial relationships: instead of assuming stationarity and fitting a single variogram, it learns how grade varies *conditional* on lithology, structure, and geophysics — much like a geologist interprets patterns across datasets. This requires careful feature engineering: drill hole assays are composited to consistent lengths, geophysical responses are co-located and normalized, and structural orientation tensors are encoded as spherical harmonics to preserve rotational invariance.
Advanced implementations go beyond point prediction. Graph Neural Networks (GNNs) treat drill holes as nodes and geological contacts as edges, enabling propagation of grade information across fault-bounded domains. Deep Gaussian Processes allow hierarchical uncertainty propagation — from sensor noise → assay error → spatial correlation → economic cutoff variability — resulting in full posterior predictive distributions per block, not just mean grades.
The frontier lies in real-time assimilation: modern systems ingest face mapping LiDAR, XRF scans, and in-borehole gamma logs directly into the model pipeline. This demands lightweight architectures (e.g., quantized neural nets) deployed on edge hardware, coupled with Bayesian updating schemes that adjust kernel hyperparameters without full retraining — ensuring grade control decisions reflect the latest 24-hour production data, not last-quarter drilling.
Advanced implementations go beyond point prediction. Graph Neural Networks (GNNs) treat drill holes as nodes and geological contacts as edges, enabling propagation of grade information across fault-bounded domains. Deep Gaussian Processes allow hierarchical uncertainty propagation — from sensor noise → assay error → spatial correlation → economic cutoff variability — resulting in full posterior predictive distributions per block, not just mean grades.
The frontier lies in real-time assimilation: modern systems ingest face mapping LiDAR, XRF scans, and in-borehole gamma logs directly into the model pipeline. This demands lightweight architectures (e.g., quantized neural nets) deployed on edge hardware, coupled with Bayesian updating schemes that adjust kernel hyperparameters without full retraining — ensuring grade control decisions reflect the latest 24-hour production data, not last-quarter drilling.
🔄 Engineering Workflow
Step 1
Step 1: Geological Conceptual Model Integration — harmonize structural domains, alteration zonation, and lithological contacts into 3D framework
→
Step 2
Step 2: Multi-Source Data Harmonization — align assay composites, geophysical grids, hyperspectral pixel clusters, and downhole sensor logs to common coordinate system and support volume
→
Step 3
Step 3: Geostatistical Preconditioning — compute directional variograms, fit anisotropic covariance models, and derive spatial priors for ML regularization
→
Step 4
Step 4: ML Model Training & Ensemble Generation — train hybrid models (e.g., RF + kriging residuals, GNN + MPS) with cross-validation and uncertainty-aware loss functions
→
Step 5
Step 5: Conditional Realization Sampling — generate ≥100 equiprobable orebody realizations honoring hard data, spatial constraints, and geological rules
→
Step 6
Step 6: Economic Filtering & Reserve Classification — apply cutoff grade, mining cost surfaces, and dilution envelopes to each realization for probabilistic resource estimation
→
Step 7
Step 7: Operational Feedback Loop — integrate short-interval assay data from production faces to update model parameters weekly via online learning
📋 Decision Guide
| Rock/Field Condition | Recommended Design Action |
|---|---|
| High-grade vein system with erratic continuity (range <15 m vertical, entropy >1.6 bits) | Switch from stationary kriging to non-stationary GPR + MPS conditioning; deploy targeted infill drilling at 5–10 m spacing along vein trace |
| Massive sulfide deposit with strong lithological control and dense assay grid (>45 samples/hm³) | Use gradient-boosted tree ensemble (XGBoost) conditioned on lithology, EM response, and density log; suppress variogram modeling in favor of feature importance weighting |
| Low-grade porphyry with gradual transitions and sparse deep drilling (<12 samples/hm³ below 500 m) | Fuse airborne magnetic and IP data via convolutional autoencoder; apply deep Gaussian process regression with depth-varying length scales |
📊 Key Properties & Parameters
Grade Variogram Range
10–250 m (horizontal), 2–40 m (vertical) depending on deposit styleThe distance beyond which assay grade values become spatially uncorrelated; quantifies grade continuity in a specific direction.
⚡ Engineering Impact:
Directly controls block model support size and minimum mining width — undersized range causes over-smoothing and grade blurring.
Training Data Density
8–60 samples/hm³ (open pit), 20–120 samples/hm³ (underground)Number of assay composites per cubic hectometer (hm³) used to train ML models.
⚡ Engineering Impact:
Below 12 samples/hm³, ML models fail to learn structural trends and default to geostatistical priors; above 80, diminishing returns and overfitting risk increase.
Uncertainty Entropy (Shannon)
0.3–2.1 bits per block (lower = higher confidence)Information-theoretic measure of grade prediction uncertainty across ensemble realizations, expressed in bits.
⚡ Engineering Impact:
Entropy >1.4 bits/block triggers mandatory re-sampling or model recalibration before reserve classification.
SVM Kernel Bandwidth (γ)
0.005–0.08 (unitless, scaled to normalized feature space)Parameter controlling locality sensitivity of Support Vector Regression models trained on geochemical vectors.
⚡ Engineering Impact:
γ <0.01 suppresses local grade anomalies (e.g., high-grade veins); γ >0.06 amplifies noise and violates geological plausibility.
📐 Key Formulas
Shannon Entropy (Per Block)
H = -∑ₚ pᵢ log₂(pᵢ)Measures uncertainty in grade probability distribution across ensemble realizations
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| p_i | Probability of grade realization i | dimensionless | Probability of observing the i-th grade value in the ensemble |
| H | Shannon Entropy | bits | Measure of uncertainty in the grade probability distribution across ensemble realizations |
Typical Ranges:
High-confidence open-pit blocks
0.3 – 0.7 bits
Deep, poorly drilled underground stopes
1.2 – 2.1 bits
⚠️ H ≤ 1.0 bits required for Measured Resource classification (JORC Code)
Effective Sample Density
ρ_eff = N / (V × f_corr)Adjusted training sample density accounting for spatial autocorrelation (f_corr = 1 − C(h)/C(0))
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| ρ_eff | Effective Sample Density | samples/m³ | Adjusted training sample density accounting for spatial autocorrelation |
| N | Number of Samples | dimensionless | Total count of training samples |
| V | Volume | m³ | Spatial volume over which samples are distributed |
| f_corr | Correlation Correction Factor | dimensionless | Factor adjusting for spatial autocorrelation, defined as 1 − C(h)/C(0), where C(h) is the covariance at lag h and C(0) is the variance |
Typical Ranges:
Well-structured massive sulfides
40 – 110 samples/hm³
Disseminated porphyry with long-range continuity
12 – 38 samples/hm³
⚠️ ρ_eff < 8 samples/hm³ invalidates ML-only approaches; geostatistics must dominate
🏭 Engineering Example
Cadia East (New South Wales, Australia)
Porphyritic monzonite / quartz monzoniteUncertainty Entropy
0.92 bits/block
Training Data Density
32 samples/hm³
Block Model Resolution
10 × 10 × 5 m
SVM Kernel Bandwidth (γ)
0.024
Grade Variogram Range (vertical)
18 m
Conditional Realizations Generated
120
🏗️ Applications
- Probabilistic reserve estimation
- Short-term grade control optimization
- Dilution forecasting in drawpoint management
- Geological risk mapping for stope design
🔧 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
How does Geostatistical Machine Learning differ from traditional kriging or deterministic ML models for orebody modeling?
Unlike traditional kriging—which relies solely on spatial correlation and Gaussian assumptions—Geostatistical ML embeds geological constraints (e.g., facies continuity, structural trends) into ML architectures (e.g., graph neural networks conditioned on drill-hole topologies) while preserving uncertainty quantification. Unlike black-box ML models (e.g., standalone random forests), it enforces spatial consistency, honors hard data (e.g., assay values), and generates *ensembles* of geologically plausible realizations—not just point predictions.
What types of input data can be integrated into a Geostatistical ML orebody model?
The framework natively supports heterogeneous, multi-scale data: high-resolution drill assay composites, 3D geophysical inversions (e.g., resistivity, magnetic susceptibility), hyperspectral mineralogical maps (e.g., ASTER, HyMap), borehole geotechnical and lithological logs, structural interpretations (faults, folds), and even historical production data—all aligned in a common spatial reference frame and encoded as features or graph edges.
Can Geostatistical ML quantify uncertainty—and how is that useful for mine planning?
Yes. By generating multiple conditional realizations (not just a single 'best guess'), it delivers spatially explicit uncertainty metrics—such as grade variance, probability-of-exceedance maps, or connectivity risk scores. This enables risk-informed decisions: optimizing pit shell design under grade uncertainty, scheduling cut-off grades with confidence intervals, or identifying high-risk dilution zones before excavation.
Is domain expertise still required when using Geostatistical ML—or is it fully automated?
Domain expertise remains essential. Geostatistical ML augments—not replaces—geological judgment: experts define conditioning rules (e.g., stratigraphic layering constraints), validate realization plausibility, select appropriate spatial kernels or graph neighborhoods, and interpret ML-derived feature importance in geological terms. The model learns *from* expert-labeled patterns but cannot infer tectonic history or deposit genesis without guidance.
What computational or data prerequisites are needed to implement Geostatistical ML in practice?
Key prerequisites include: (1) spatially registered, quality-controlled drill and auxiliary datasets; (2) a robust 3D geological framework (e.g., implicit modeling of domains/faults); (3) GPU-accelerated infrastructure for training GNNs or deep generative models; and (4) software interoperability (e.g., Python-based ML stacks interfacing with geostatistical libraries like GSTools or commercial platforms like Leapfrog Geo or Vulcan). Pilot implementations often start with a single domain and expand iteratively.
🎨 Technical Diagrams
📚 References
[1]
[2]
Guidelines for the Application of Machine Learning in Mineral Resource Estimation
— CIM Mineral Resource Committee
[3]
JORC Code (2012 Edition)
— Joint Ore Reserves Committee