🎓 Lesson 8
D5
Calibrating Kinetic Models with Bayesian Inference
Bayesian inference is a way to update our best guess about how a blasting model works—like rock fragmentation—by combining prior engineering knowledge with new field data.
🎯 Learning Objectives
- ✓ Calculate posterior parameter distributions for a fragmentation kinetic model using Bayes’ theorem
- ✓ Design a prior distribution for blasthole burden based on geomechanical rock mass classification (e.g., RMR)
- ✓ Analyze posterior predictive performance of a calibrated Kuz-Ram model against image-based fragment size data
- ✓ Apply Markov Chain Monte Carlo (MCMC) sampling to estimate credible intervals for powder factor sensitivity
📖 Why This Matters
In open-pit mines, over- or under-fragmentation wastes energy, damages equipment, and delays loading. Traditional 'tuning' of kinetic models (e.g., Kuz-Ram) relies on trial-and-error or least-squares fitting—ignoring uncertainty in rock properties and sensor noise. Bayesian calibration turns subjective expert judgment (e.g., 'this granite usually needs 0.28–0.32 kg/m³') into mathematically defensible priors—and lets each blast refine the model like a learning digital twin. That’s how BHP’s Olympic Dam reduced oversize by 22% after integrating Bayesian-calibrated fragmentation models into their blast design loop.
📘 Core Principles
Bayesian inference rests on three pillars: (1) Prior knowledge—expressed as probability distributions encoding geotechnical experience (e.g., burden ~ Normal(3.1 m, 0.4 m) for competent granite); (2) Likelihood—how probable observed fragment sizes are given a set of model parameters; and (3) Posterior—the updated belief about parameters after seeing data. Unlike deterministic calibration, Bayesian methods yield full uncertainty distributions—not just point estimates—enabling risk-aware decisions (e.g., '95% chance fragmentation index P80 < 120 mm'). For kinetic models like the Kuz-Ram equation, parameters such as rock factor K and exponent n become random variables, and their joint posterior reveals correlations (e.g., high K often pairs with low n), guiding targeted field testing.
📐 Bayes’ Theorem for Model Calibration
Bayes’ theorem links prior belief, data likelihood, and updated posterior: p(θ|D) ∝ p(D|θ) × p(θ), where θ represents kinetic model parameters (e.g., K, n in Kuz-Ram) and D is observed fragment size data. When analytical solutions are intractable—as with nonlinear fragmentation models—Markov Chain Monte Carlo (MCMC) approximates the posterior numerically.
💡 Worked Example
Problem: A blast in porphyritic granite yields P80 = 92 mm (measured by photogrammetry). Prior for rock factor K is Normal(17.5, 2.1) (unitless, per Kuz-Ram). Likelihood assumes Gaussian error: p(P80_obs|K) ~ N(P80_pred(K), σ=8 mm), where P80_pred(K) = K × (burden × spacing)^0.8 × (powder_factor)^−0.2, with burden = 3.4 m, spacing = 4.2 m, PF = 0.31 kg/m³.
1.
Step 1: Compute P80_pred = 17.5 × (3.4 × 4.2)^0.8 × (0.31)^−0.2 ≈ 17.5 × 6.57^0.8 × 1.24 ≈ 17.5 × 5.12 × 1.24 ≈ 111.2 mm
2.
Step 2: Evaluate likelihood: p(92|K=17.5) ∝ exp[−(92 − 111.2)²/(2×8²)] = exp[−462.4/128] ≈ exp[−3.61] ≈ 0.027
3.
Step 3: Weight prior density (N(17.5,2.1) at K=17.5 ≈ 0.191) by likelihood → unnormalized posterior density ≈ 0.191 × 0.027 ≈ 0.00516. MCMC sampling across K ∈ [12,24] shows posterior mode shifts to K ≈ 15.3, reflecting data-informed downward revision.
Answer:
The posterior mode for K is 15.3 (95% CI: 14.1–16.6), indicating the prior overestimated rock resistance. This revised K improves next-blast P80 prediction error from ±24 mm to ±11 mm.
🏗️ Real-World Application
At Newmont’s Boddington Mine (WA), engineers calibrated a modified Kuz-Ram kinetic model for saprolite–granite transition zones using Bayesian inference. Prior distributions for rock factor K were derived from 120 historical RMR–K correlations; fragment size data came from drone-based LiDAR scans of 18 production blasts. Using PyMC3 MCMC, they estimated posterior joint distributions for K and explosive energy factor E. The calibrated model reduced median P80 prediction error by 37% and flagged two outlier blasts—later traced to undetected water saturation—demonstrating Bayesian diagnostics' value in model validation and anomaly detection.