🎓 Lesson 12
D5
XRF and NIR Calibration Best Practices
XRF and NIR calibration is like teaching a sensor to correctly 'read' the chemistry of ore by comparing its measurements to known lab results.
🎯 Learning Objectives
- ✓ Explain the physical principles underlying XRF and NIR spectral response differences in mineral matrices
- ✓ Design a calibration validation protocol compliant with ISO 12099:2017 for feed-forward control applications
- ✓ Analyze calibration model residuals to diagnose matrix effects or outliers
- ✓ Apply spectral preprocessing techniques (e.g., SNV, first derivative, detrending) to improve PLS model R² > 0.95 for Fe grade prediction
- ✓ Calculate and interpret key validation metrics (RMSEP, bias, RPD) for an NIR calibration model using cross-validated test sets
📖 Why This Matters
In sensor-based feed-forward control, XRF and NIR analyzers mounted on conveyors or in-line slurry streams provide real-time grade data—enabling immediate adjustments to crushing, grinding, or flotation circuits *before* suboptimal material enters downstream units. But if calibration drifts or misrepresents complex ores (e.g., mixed hematite/goethite, clay-coated particles), feed-forward decisions become dangerously misleading—causing grade loss, reagent overuse, or circuit instability. This lesson bridges lab-grade accuracy with plant-floor reliability.
📘 Core Principles
XRF excites core-shell electrons, emitting element-specific X-rays; intensity correlates linearly with atomic concentration—but attenuated by matrix effects (e.g., Fe absorption suppressing Mn signal). NIR measures vibrational overtones (e.g., O–H, C–O bonds); responses are non-linear, highly sensitive to particle size, moisture, and surface heterogeneity. Calibration thus requires not just regression, but *physics-informed preprocessing*: XRF demands matrix correction (e.g., Fundamental Parameters or empirical alpha coefficients); NIR requires scatter correction and spectral filtering. Both demand representative calibration libraries spanning geological variability, mining sequence, and operational conditions—not just lab homogenates.
📐 PLS Cross-Validated RMSEP
Root Mean Square Error of Prediction (RMSEP) quantifies average prediction error during cross-validation—a critical metric for assessing calibration robustness before field deployment.
RMSEP (Root Mean Square Error of Prediction)
RMSEP = √[Σ(y_i − ŷ_i)² / n]Quantifies average magnitude of prediction error for a calibration model during cross-validation or external validation.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| y_i | Reference value | % | Laboratory-assayed concentration of analyte i (e.g., %Fe) |
| ŷ_i | Predicted value | % | Instrument-predicted concentration of analyte i |
| n | Number of validation samples | unitless | Size of independent validation set |
Typical Ranges:
NIR Fe grade prediction (hematite ore): 0.15 – 0.30 %
XRF SiO₂ in bauxite: 0.20 – 0.45 %
💡 Worked Example
Problem: A NIR calibration model for iron grade (%Fe) was developed using 120 samples. During 10-fold cross-validation, predicted vs. reference values yielded squared residuals: [0.04, 0.09, 0.01, ..., 0.16] (120 values total). Sum of squared residuals = 8.4.
1.
Step 1: Compute mean squared residual = total sum / number of samples = 8.4 / 120 = 0.07
2.
Step 2: Take square root: √0.07 = 0.2646
3.
Step 3: Round to two decimals and interpret: RMSEP = 0.26 %Fe
Answer:
The result is 0.26 %Fe, which falls within the safe range of <0.30 %Fe for high-grade iron ore feed-forward control per Rio Tinto’s Sensor Integration Standard v3.2.
🏗️ Real-World Application
At BHP’s Newman Operations (Pilbara), an inline XRF analyzer on the ROM conveyor feeds forward to adjust primary crusher settings and blending ratios. Initial calibration used 200 lab assays from 2022–2023 drill-core composites. Model drift was detected after 4 months due to seasonal moisture variation (>8% RH shift) affecting XRF vacuum chamber pressure and detector gain. Recalibration included humidity-controlled reference standards and a dynamic gain-correction algorithm—reducing Fe prediction error from RMSEP 0.41% to 0.23% and enabling 12% reduction in downstream grinding energy via optimized feed hardness targeting.