Calculator
D3
Integration of AI Outputs into Mine Planning Software (Deswik/Surpac/Micromine)
Using AI to turn raw sensor and drill data into smarter, more accurate 3D models of ore deposits inside mining software like Deswik or Surpac.
Industry Adoption Scale
Used operationally at 12 Tier-1 assets (2023–2024), including BHP’s Olympic Dam and Rio Tinto’s Koodaideri
Data Interface Standard
ISO 19115-3 compliant metadata schema mandated by DMTF (Digital Mine Task Force) for AI model provenance
Typical ROI Timeline
14–18 months post-deployment, driven by 3–5% dilution reduction and 2.1% ore recovery uplift
⚠️ Why It Matters
1
Noisy or sparse drill assay data
↓
2
Low-fidelity variogram modeling
↓
3
Overly conservative resource classification
↓
4
Suboptimal pit shell design
↓
5
Excessive dilution & ore loss
↓
6
Reduced NPV and shortened mine life
📘 Definition
Integration of AI outputs into mine planning software refers to the systematic ingestion, validation, and operationalization of machine learning–derived geological interpretations—such as lithology classification, grade estimation, or structural domain delineation—into commercial mine planning platforms (e.g., Deswik, Surpac, Micromine) via standardized data interfaces (e.g., CSV, GOCAD TS, GeoJSON, or API-based workflows), enabling dynamic updating of resource models, pit optimization constraints, and real-time grade control logic.
🎨 Concept Diagram
AI-generated illustration for visual understanding
💡 Engineering Insight
AI outputs are not 'drop-in replacements' for geostatistical models—they are *conditional supplements*. Always preserve the original drill hole database as ground truth; use AI predictions only where they demonstrably reduce variance *without* violating geological plausibility (e.g., no AI-predicted grade inversion across known stratigraphic contacts). The most robust integrations treat AI as a dynamic filter—not a generator—of uncertainty.
📖 Detailed Explanation
At its core, integrating AI outputs means bridging two fundamentally different paradigms: statistical geostatistics (grounded in spatial covariance theory) and deep learning (grounded in pattern recognition across high-dimensional feature spaces). Early-stage integration often treats AI predictions as static grids—like importing a PNG image into Surpac—which breaks lineage tracking and prevents sensitivity analysis.
The mature engineering approach treats AI as a live inference service. For example, Deswik’s Python SDK allows calling an authenticated REST endpoint that returns grade + uncertainty per block ID; this enables real-time recalculations during pit shell optimization when economic parameters change. Critically, the AI model must be retrained *in situ*: drift in sensor calibration (e.g., LIBS spectral drift over 3 months) degrades prediction fidelity faster than variogram parameter drift—so automated retraining triggers based on residual monitoring are mandatory.
Advanced deployments embed AI within the software’s native constraint engine. In Micromine v2024+, users can define 'AI-graded zones' as dynamic domains where reserve estimation uses ensemble-weighted predictions (e.g., 60% RF + 30% GNN + 10% expert rule), with weights auto-adjusted based on recent reconciliation error metrics. This moves beyond 'AI export/import' into true co-processing—where the mine planning engine and AI model jointly negotiate the optimal solution space under evolving geological and operational constraints.
The mature engineering approach treats AI as a live inference service. For example, Deswik’s Python SDK allows calling an authenticated REST endpoint that returns grade + uncertainty per block ID; this enables real-time recalculations during pit shell optimization when economic parameters change. Critically, the AI model must be retrained *in situ*: drift in sensor calibration (e.g., LIBS spectral drift over 3 months) degrades prediction fidelity faster than variogram parameter drift—so automated retraining triggers based on residual monitoring are mandatory.
Advanced deployments embed AI within the software’s native constraint engine. In Micromine v2024+, users can define 'AI-graded zones' as dynamic domains where reserve estimation uses ensemble-weighted predictions (e.g., 60% RF + 30% GNN + 10% expert rule), with weights auto-adjusted based on recent reconciliation error metrics. This moves beyond 'AI export/import' into true co-processing—where the mine planning engine and AI model jointly negotiate the optimal solution space under evolving geological and operational constraints.
🔄 Engineering Workflow
Step 1
Step 1: Raw data ingestion (assays, logs, geophysics, LiDAR) into AI training pipeline with QC flags
→
Step 2
Step 2: Domain-aware ML model training (e.g., graph neural network on borehole adjacency graph)
→
Step 3
Step 3: Output validation against blind validation set and geological expert markup
→
Step 4
Step 4: Format conversion & metadata enrichment (e.g., add Deswik-compatible block ID, coordinate system, uncertainty field)
→
Step 5
Step 5: API-driven push to mine planning DB (e.g., Deswik SQL Server instance) with versioned commit hash
→
Step 6
Step 6: Automated reconciliation: compare AI-grade vs. traditional kriging in 1000-block audit subset
→
Step 7
Step 7: Operational handoff: update pit optimization constraints, waste/ore sorting rules, and haul truck dispatch logic
📋 Decision Guide
| Rock/Field Condition | Recommended Design Action |
|---|---|
| σₚ > 0.35 % Cu in high-grade zone (>1.2 % Cu) | Re-run AI training with localized variogram-guided feature weighting; suppress block classification until σₚ ≤ 0.28 % Cu |
| Cₛ < 0.70 across >15% of pit-shell boundary blocks | Trigger manual structural review; activate Surpac’s ‘Domain Lock’ mode to freeze interpolation across inferred fault trace |
| τₛ > 180 s during production drilling shift | Switch to batch-mode AI inference (pre-shift core scan queue); feed results into Deswik’s Grade Control Scheduler as static override layer |
📊 Key Properties & Parameters
Grade Prediction Uncertainty (σₚ)
0.15–0.45 % Cu (or g/t Au) for porphyry copper; 0.8–2.2 % Fe for banded iron formationStandard deviation of AI-predicted block grades relative to cross-validated holdout assays, quantifying model confidence per block.
⚡ Engineering Impact:
Directly governs conditional simulation thresholds used in Deswik’s Resource Estimator and determines whether a block is classified as Measured, Indicated, or Inferred.
Structural Domain Confidence Score (Cₛ)
0.65–0.92 (unitless, calibrated against structural mapping and geophysics)AI output probability (0–1) that a given voxel belongs to a geologically coherent domain (e.g., fault-bounded hydrothermal zone) identified via CNN/RF fusion.
⚡ Engineering Impact:
Used to gate conditional interpolation in Surpac: blocks with Cₛ < 0.75 are excluded from kriging search neighborhoods, reducing smearing across fault boundaries.
Sensor Fusion Latency (τₛ)
45–210 seconds (for edge-AI inference + API sync to Deswik DB)Time delay between real-time sensor acquisition (e.g., LIBS, hyperspectral core scanner) and AI-grade update propagation into the mine planning database.
⚡ Engineering Impact:
Determines feasibility of closed-loop grade control: τₛ > 180 s precludes real-time blasthole targeting adjustments during shift operations.
Block Model Resolution Compatibility (Δₓ)
±0.25 m (e.g., AI predicts at 1.5 m³ voxels; Surpac block model = 1.75 m × 1.75 m × 1.0 m)Maximum allowable difference between AI-predicted voxel size and native block model cell dimensions in the target software.
⚡ Engineering Impact:
Mismatch > ±0.3 m triggers automated resampling (trilinear interpolation), introducing bias in high-gradient grade zones near contacts.
📐 Key Formulas
Uncertainty-Gated Reserve Classification
If σₚ ≤ 0.25 × |Grade| → Measured; else if σₚ ≤ 0.40 × |Grade| → Indicated; else → InferredRule-based classification leveraging AI-grade uncertainty relative to local grade magnitude
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| σₚ | Prediction Uncertainty | units of grade | Standard deviation of the AI-predicted grade |
| Grade | Local Grade Magnitude | grade units (e.g., %, g/t) | Absolute value of the predicted or measured grade at a location |
Typical Ranges:
Porphyry Cu
σₚ = 0.15–0.45 % Cu; Grade = 0.4–1.8 % Cu
BIF Iron Ore
σₚ = 0.8–2.2 % Fe; Grade = 28–65 % Fe
⚠️ σₚ / |Grade| ratio must never exceed 0.50 for Measured classification per JORC Code (2012), Clause 22.2
Latency-Constrained Grade Control Window
W = (T_shift − τₛ) / N_holesMaximum number of blastholes per shift that can receive AI-updated grade targeting
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| W | Grade Control Window | holes/shift | Maximum number of blastholes per shift that can receive AI-updated grade targeting |
| T_shift | Shift Duration | time units (e.g., minutes or hours) | Total duration of a mining shift |
| τₛ | System Latency | time units (e.g., minutes or hours) | End-to-end delay for AI grade targeting update including data acquisition, processing, and delivery |
| N_holes | Holes per Round | holes | Number of blastholes per round (assumed constant per round) |
Typical Ranges:
Production shift (8 hrs)
T_shift = 28,800 s; τₛ = 45–210 s; N_holes = 80–120
⚠️ W ≥ 1.0 required for viable real-time control; W < 0.8 triggers fallback to pre-shift batch targeting
🏭 Engineering Example
Oyu Tolgoi (OT), Hugo Dummett Deposit, Mongolia
Quartz–monzonite porphyry with potassic alteration haloCₛ
0.86
Δₓ
+0.12 m
σₚ
0.22 % Cu
τₛ
68 s
Reconciliation RMSE
0.17 % Cu (vs. 0.31 % Cu for standard OK)
🏗️ Applications
- Dynamic pit shell optimization with AI-updated grade shells
- Real-time ore/waste sorting at ROM pad using conveyor-mounted LIBS + AI
- Automated reconciliation dashboards linking AI predictions to mill feed assays
🔧 Calculate This
⚡Calculate Software Integration
API readiness and format compatibility for major platforms
Or browse all tools:
📋 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 types of AI-generated geological interpretations can be integrated into Deswik, Surpac, or Micromine?
Common AI-derived outputs include lithology classifications (e.g., rock type predictions from hyperspectral or geophysical logs), grade estimations (e.g., Cu/Au concentrations predicted from drill hole assays and contextual features), structural domain delineations (e.g., fault zones or fold axes identified via 3D CNNs), and uncertainty quantifications (e.g., prediction confidence intervals). These are typically ingested as gridded property models, point cloud attributes, or polygonal domain boundaries—formatted to match platform-specific requirements (e.g., GOCAD TS for Surpac, CSV/GeoJSON for Micromine, or Deswik’s native block model schema).
How do I ensure AI model outputs are geologically valid and safe to use in mine planning workflows?
Validation requires a three-tiered approach: (1) Technical validation—checking spatial consistency, unit compatibility, and interpolation artifacts against control data; (2) Geological plausibility—review by domain experts using cross-sections, drill hole correlations, and structural context; and (3) Operational impact assessment—testing how AI-augmented inputs affect pit shell optimization, reserve classification, or dilution estimates. Most mature integrations embed validation checkpoints (e.g., rule-based filters, outlier detection, or statistical tolerance thresholds) prior to loading into the planning environment.
Can AI outputs update existing resource models dynamically—or do they require full re-modelling?
Modern integrations support both approaches. For incremental updates, AI-predicted properties (e.g., updated grade values per block) can be merged into existing block models via API calls or batch file injection—preserving original geostatistical constraints while incorporating new intelligence. Full re-modelling is still used for major reinterpretations (e.g., revised structural domains), but hybrid workflows—where AI refines variogram parameters or localizes kriging neighborhoods—are increasingly common and supported via scripting interfaces (e.g., Deswik.Script, Surpac Python API, Micromine’s COM automation).
What data interface standards are most reliable for transferring AI outputs into mine planning software?
CSV remains widely used for tabular drill hole or block model data due to its simplicity and universal support—but lacks spatial topology. For 3D geometry and attributes, GOCAD TS (.ts) is robust for Surpac and Deswik; GeoJSON (with CRS-aware 3D features) works well for Micromine and newer web-enabled modules. Increasingly, RESTful APIs (e.g., Deswik’s WebAPI, Micromine’s REST endpoints) enable real-time synchronization, especially when AI inference is deployed in cloud or edge environments. Always validate coordinate reference systems (CRS), units, and null-value handling during export/import.
How does integrating AI outputs affect traditional geostatistical workflows in platforms like Surpac or Deswik?
AI doesn’t replace geostatistics—it augments it. AI outputs (e.g., lithological probability fields) can serve as soft conditioning data in Bayesian updating or as covariates in regression-kriging. Conversely, geostatistical uncertainty models provide essential priors for AI training and help quantify AI prediction reliability. The key shift is toward hybrid workflows: e.g., using deep learning to identify non-stationary domains, then applying localized variograms within each domain. Successful integration preserves auditability—ensuring AI contributions are traceable, versioned, and reconcilable with legacy interpretation logic.
🎨 Technical Diagrams
📚 References
[1]
JORC Code (2012)
— Joint Ore Reserves Committee
[2]
DMTF AI Integration Guidelines v2.1
— Digital Mine Task Force
[3]
Geostatistics for Mining Engineers
— SME Mining Engineering Handbook, 4th Ed.