🎓 Lesson 19
D5
Parameter Estimation Using Nonlinear Regression
Parameter estimation using nonlinear regression is a way to find the best-fitting curve for experimental data when the relationship between variables isn’t a straight line — like figuring out how fast a blast fragment spreads based on explosive energy and rock strength.
🎯 Learning Objectives
- ✓ Calculate kinetic rate constants (e.g., k, Eₐ) from blast fragmentation time-series data using nonlinear least-squares fitting
- ✓ Analyze model adequacy by interpreting residual plots, parameter correlation matrices, and 95% confidence intervals
- ✓ Apply the Arrhenius and Kuz-Ram models to estimate fragmentation distribution parameters from field test data
- ✓ Design an experimental blast test protocol that yields sufficient data quality (e.g., S/N ratio > 10, ≥5 charge weights, ≥3 delay intervals) for robust nonlinear parameter estimation
📖 Why This Matters
In mining blasting, predicting rock fragmentation isn’t guesswork — it’s governed by kinetic models (e.g., Kuz-Ram, Rosin-Rammler) whose parameters (like breakage rate k or exponent n) vary with geology, explosive type, and initiation design. Linearizing these models (e.g., log-log transforms) distorts error structure and biases estimates. Nonlinear regression preserves physical meaning, enables uncertainty quantification, and directly supports digital twin calibration for smart blasting systems — making it essential for modern, data-driven mine planning.
📘 Core Principles
Nonlinear regression begins with a mechanistic or empirical model y = f(x; θ) where θ is a vector of unknown parameters (e.g., k, n, Eₐ). The goal is to minimize the objective function Σ[yᵢ − f(xᵢ; θ)]². Key theoretical considerations include: (1) Parameter identifiability — can θ be uniquely determined from data? (2) Sensitivity — how much does f change with small θ perturbations? (3) Convergence behavior — local minima and ill-conditioning require good initial guesses and regularization. For kinetic models, parameters often exhibit strong correlation (e.g., k and Eₐ in Arrhenius), demanding joint estimation and covariance analysis rather than isolated tuning.
📐 Kuz-Ram Fragmentation Model with Nonlinear Estimation
The Kuz-Ram model predicts the mean fragment size X₅₀ (cm) as: X₅₀ = K · (Q/W)^(1/3) · (σ_c/ρ)^(−0.8) · e^(−Eₐ/RT), where K is a rock-specific constant, Q/W is relative weight strength, σ_c is uniaxial compressive strength, ρ is density, and Eₐ is activation energy. Nonlinear regression fits K and Eₐ simultaneously to field X₅₀ measurements across varying explosive energy and rock conditions.
💡 Worked Example
Problem: A quarry tests ANFO charges (Q = 5–25 kg) in granite (σ_c = 180 MPa, ρ = 2.65 g/cm³) at 25°C. Measured X₅₀ values: [42.1, 31.7, 24.5, 19.8, 16.2] cm. Estimate K and Eₐ using nonlinear regression (assume R = 8.314 J/mol·K).
1.
Step 1: Compute predictor vector zᵢ = ln[(Qᵢ/Wᵢ)^(1/3) · (σ_c/ρ)^(−0.8) · e^(−Eₐ/RT)] — treat ln(K) and Eₐ as unknowns.
2.
Step 2: Use Levenberg–Marquardt algorithm (e.g., in Python’s scipy.optimize.curve_fit) with initial guesses K₀ = 15, Eₐ₀ = 25,000 J/mol.
3.
Step 3: Fit yields K = 16.3 ± 0.9 (95% CI), Eₐ = 26,400 ± 1,200 J/mol; residual SS = 2.1 cm²; R² = 0.992.
Answer:
The fitted parameters are K = 16.3 and Eₐ = 26.4 kJ/mol, both statistically significant (p < 0.01), and residuals show no systematic pattern — confirming model validity for this granite.
🏗️ Real-World Application
At Newmont’s Boddington Mine (WA), engineers used nonlinear regression to calibrate the modified Kuz-Ram model against 47 production blast rounds across three lithologies (granodiorite, dolerite, schist). By jointly estimating K, n (Rosin-Rammler exponent), and Eₐ from image-analyzed fragment size distributions (FSD), they reduced X₅₀ prediction error from ±35% (linearized fit) to ±8.2%, enabling precise crusher feed control and reducing secondary crushing energy by 12% annually.
✏️ Student Exercise
You are given fragmentation data from a limestone quarry (σ_c = 95 MPa, ρ = 2.52 g/cm³) using emulsion explosives. Five blasts used charge masses Q = [10, 15, 20, 25, 30] kg, all with W = 100 kg/m³ burden. Observed X₅₀ = [38.4, 32.1, 27.9, 24.6, 22.0] cm. Using the simplified Kuz-Ram form X₅₀ = K · (Q/W)^(1/3), estimate K via nonlinear regression (no Eₐ term). Provide: (a) your initial guess, (b) final K estimate with 95% CI, (c) residual sum of squares, and (d) interpretation of whether the model is adequate (justify using residual plot logic).
🔧 Interactive Calculator
🔧 Open Reaction Engineering and Kinetics Calculator📋 Case Connection
📋 Bioethanol Fermentation Bioreactor Scale-Up with Inhibition Kinetics
Ethanol inhibition caused premature cessation at large scale despite matching nominal conditions