Context
Exploration drilling is one of the largest costs in mining. Traditional grid-based sampling ignores what’s already been learned — placing holes at uniform intervals whether the geology is well-understood or completely unknown.
Problem
- Redundant samples: Budget spent where uncertainty is already low
- Coverage gaps: Faults, intrusions, and complex geology under-sampled
- No feedback loop: Strategy doesn’t adapt to collected data
- Model uncertainty ignored: All samples treated as equally valuable
What We Built
Multi-Model ML Pipeline
Implemented four approaches for ore grade prediction, each with uncertainty quantification:
- Gaussian Processes: GPFlow with Matern52/RBF kernels — best on smooth geological variations
- Bayesian Neural Networks: TensorFlow Probability, 3×128 layers with MC sampling — handles complex non-linear patterns
- Random Forest Ensembles: 100-500 trees with inter-tree variance — fast, robust to outliers
- Deep Ensembles: Multiple BNNs with aleatoric/epistemic uncertainty decomposition — most reliable confidence estimates
Synthetic Data Generation
Built a configurable geology simulator for algorithm validation:
- 14 geological patterns: Layered deposits, fault displacements, ore body intrusions, vein systems, fractal variations
- 8 noise models: Measurement error, outliers, missing data, spatial correlation
- Parametric control: Tune complexity to match real deposit characteristics
Adaptive Sampling Engine
Six sampling strategies with head-to-head comparison:
- Uniform grid (baseline)
- Random / stratified random
- Space-filling curves (spiral, Hilbert)
- Uncertainty-driven adaptive: Next sample placed where model is least confident
- Along-feature: Follows predicted geological boundaries
3D Drill Data Integration
- Spatial coordinates with depth profiles
- Iron content assays (Fe_PPM)
- Surface elevation interpolation
- Handles missing/erroneous measurements
Uncertainty Quantification
- Confidence intervals with calibration validation
- Aleatoric (inherent noise) vs epistemic (reducible with more data) decomposition
- Real-time uncertainty maps showing where to drill next
Results
- ~40% cost reduction for equivalent geological model confidence
- Adaptive strategy outperforms grid across all sample densities
- Largest gains in complex geology (faults, intrusions)
- Well-calibrated uncertainty estimates across all methods
Technical Stack
GPFlow, TensorFlow Probability, scikit-learn, NumPy/SciPy, GeoPandas, Matplotlib
Work conducted with a top-3 global mining company; specific data anonymized.