AI & Ore Sorting

Artificial intelligence and sensors are changing how ore is found, mined and processed. Learn sensor-based ore sorting (XRT, XRF, NIR, laser), bulk sorting, pre-concentration mass balances, machine-learning applications (grade prediction, image-based fragmentation, predictive maintenance, prospectivity), classification metrics, and risks of AI — with a Python sorting simulation.

Rejecting waste before grinding saves energy, water and tailings. Sensor-based sorters scan every rock on a belt and fire air jets to remove waste; bulk sorters divert whole truckloads or belt segments. Machine learning interprets the sensor signals — and increasingly supports decisions across the mine.

An analogy

An ore sorter is like a fast fruit-packing line: cameras and sensors inspect each piece and air jets flick away the bad ones.

Ore sorting

Sensors: XRT (atomic density), XRF (elemental), NIR (mineralogy), colour cameras, laser/LIBS, magnetic and electromagnetic. Particle sorters handle ≈ 10–100 mm rocks; bulk sorters use belt-mounted PGNAA/XRF or truck-level analysis to divert ore/waste.

Machine learning uses

Grade and hardness prediction from MWD and core images, fragmentation from shovel cameras, predictive maintenance, prospectivity mapping, ventilation and plant process control, and computer vision for safety (PPE, proximity).

Measuring models

For ore/waste classification: precision (how much 'ore' is really ore), recall (how much real ore is caught), confusion matrices and economic cost of each error type. Always validate on new, independent data.

Risks and governance

Models trained on one ore zone may fail in another (data drift); black-box decisions affecting safety need explainability and human oversight; data governance and cybersecurity are essential. AI supports engineers — it does not replace understanding of geology and physics.

Formulas

Worked example: Sorting mass balance

Given: Feed 1.0 g/t; Reject 40 % of mass at 0.2 g/t

Find: Product grade and recovery.

Answer: 1.53 g/t with 92 % recovery; mill treats 40 % less rock

Worked example: Classifier metrics

Given: TP 90, FP 10, FN 15, TN 85

Find: Precision, recall, accuracy.

Answer: P 0.90, R 0.86, accuracy 0.875

Simulate a sensor-based ore sorter

import numpy as np

# Sensor-based ore sorting simulation (e.g. XRT sensor on a belt)
rng = np.random.default_rng(1)
n = 50_000
grade = rng.lognormal(np.log(0.6), 1.0, n)          # g/t Au per rock
sensor = grade * rng.lognormal(0, 0.35, n)          # noisy sensor estimate
feed = grade.mean()

for threshold in (0.2, 0.4, 0.6):
    keep = sensor >= threshold                      # air jets reject the rest
    mass = keep.mean()
    rec = grade[keep].sum() / grade.sum()
    print(f"thr {threshold:.1f}: keep {mass:5.1%} mass, "
          f"recovery {rec:5.1%}, grade {grade[keep].mean():.2f} g/t "
          f"(feed {feed:.2f})")
# thr 0.2: keep 84.9% mass, recovery 97.6%, grade 1.13 g/t (feed 0.98)
# thr 0.4: keep 64.6% mass, recovery 90.6%, grade 1.38 g/t (feed 0.98)
# thr 0.6: keep 49.6% mass, recovery 82.5%, grade 1.63 g/t (feed 0.98)

Raising the sensor threshold rejects more waste mass (saving grinding cost) but loses more metal. Sorting pays when the ore is heterogeneous at rock scale and the sensor correlates well with grade; the optimum threshold comes from an NPV trade-off.

Common mistakes

Applications

Key takeaways

Going further