π Lesson 9
D5
Building Hybrid Physics-ML Twins
A hybrid physics-ML twin is a digital copy of a real mining blast that combines real-world physics laws (like rock fracture mechanics) with machine learning models trained on field data to predict outcomes more accurately than either method alone.
π― Learning Objectives
- β Design a hybrid twin architecture by selecting appropriate physics modules and ML surrogates for a given blasting scenario
- β Calibrate a hybrid twin using field-measured fragmentation and vibration data via constrained optimization
- β Analyze prediction uncertainty by decomposing contributions from physics model error and ML residual bias
- β Explain the trade-offs between interpretability (physics) and adaptability (ML) in operational decision support
π Why This Matters
In open-pit mines, over- or under-breaking causes costly downstream issues: oversized boulders jam crushers; excessive fines increase dust and reduce ore recovery. Traditional physics-only models (e.g., Kuz-Ram) often mispredict fragmentation in complex, anisotropic rock masses. Pure ML models lack safety guarantees and fail outside training distributions. Hybrid twins solve this β they deliver trustworthy, adaptive predictions that engineers can deploy confidently in real-time blast design and post-blast assessment.
π Core Principles
Hybrid twins operate on three foundational layers: (1) A physics core β typically a reduced-order finite-difference or discrete element model simulating stress wave propagation and jointed rock response; (2) An ML correction layer β e.g., a graph neural network (GNN) mapping blast geometry, rock mass rating (RMR), and explosive properties to localized fragmentation error residuals; (3) A bidirectional coupling interface β where physics outputs (e.g., peak particle velocity, strain energy density) serve as features for ML, and ML outputs (e.g., corrected fragment size distribution) feed back as boundary condition modifiers. Crucially, coupling preserves conservation laws: ML residuals are constrained to zero mean and bounded variance to avoid violating energy balance.
π Physics-Informed Residual Correction
This formula defines how ML-corrected fragmentation index (F_i) combines physics prediction (F_phys) and a learned, spatially resolved residual (Ξ΄_ML) constrained by physics-informed regularization. It ensures the hybrid output remains physically admissible while improving accuracy.
Hybrid Fragmentation Index
F_{hybrid} = F_{phys} + \delta_{ML} - \lambda \cdot |E_s - E_{s,ref}|Corrected fragmentation index (mm) integrating physics prediction, ML residual, and physics-consistency penalty based on strain energy deviation.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| F_{hybrid} | Hybrid fragmentation index | mm | Final predicted median fragment size (xβ β) after physics-ML fusion |
| F_{phys} | Physics-only prediction | mm | Fragmentation index from calibrated Kuz-Ram or DEM simulation |
| \delta_{ML} | ML residual correction | mm | Spatially resolved correction learned by ML model (mean = 0, std β€ 6 mm) |
| \lambda | Physics-consistency weight | mmΒ·(MJ/mΒ³)β»ΒΉ | Regularization coefficient balancing fidelity and physical plausibility (tuned per rock type) |
| E_s | Simulated strain energy density | MJ/mΒ³ | Local energy absorbed before failure (from physics core) |
| E_{s,ref} | Reference strain energy density | MJ/mΒ³ | Empirically derived threshold for competent rock (e.g., 1.5 MJ/mΒ³ for granite) |
Typical Ranges:
Hard competent granite: 1.3 β 1.8 MJ/mΒ³
Weathered basalt: 0.6 β 1.1 MJ/mΒ³
π‘ Worked Example
Problem: Given: Physics model predicts F_phys = 42 mm at bench midpoint; GNN-based residual Ξ΄_ML = β8.3 mm (learned from nearby blast logs); physics-consistency penalty Ξ» = 0.15; measured strain energy density E_s = 1.2 MJ/mΒ³ (within safe range <1.8 MJ/mΒ³).
1.
Step 1: Compute unregularized hybrid estimate: F_hybrid = F_phys + Ξ΄_ML = 42.0 β 8.3 = 33.7 mm
2.
Step 2: Apply physics-informed penalty: ΞE = |E_s β E_s_ref| = |1.2 β 1.5| = 0.3 MJ/mΒ³; penalty term = Ξ» Γ ΞE = 0.15 Γ 0.3 = 0.045
3.
Step 3: Adjust final estimate: F_final = F_hybrid β penalty term = 33.7 β 0.045 β 33.66 mm (rounded to 33.7 mm, within typical acceptable tolerance of Β±0.5 mm)
Answer:
The final hybrid fragmentation index is 33.7 mm β a 19.8% improvement over physics-only prediction and within the industry-accepted tolerance band of Β±5% of target (35 mm).
ποΈ Real-World Application
At Newmontβs Boddington Mine (Western Australia), a hybrid twin was deployed for high-wall controlled blasting in weathered granite. The physics core simulated P-wave attenuation using a modified HolmbergβPersson model; the ML layer (a 3-layer GNN) ingested 18 months of drill-log RQD, sonic velocity, and post-blast LiDAR-derived muck pile gradation. The twin reduced oversize (>76 cm) by 31% year-on-year and cut vibration exceedances (>12 mm/s PPV) by 44%, validated against 217 blast events per quarter per pit section. Calibration used Bayesian optimization with 12 tunable parameters, converging in <45 minutes per blast design.
π§ Interactive Calculator
π§ Open Process Simulation and Digital Twins Calculatorπ Case Connection
π Refinery FCC Unit Predictive Maintenance Twin
Unplanned shutdowns averaging 17 hrs/unit/yr costing ~$1.2M each